Over the last few years I’ve worked on, and managed the development of, a lot of open source code. As you might expect, it’s primarily been hosted on cloud platforms (mostly GitHub, occasionally GitLab).

I had a realisation last month—that my relationship with open-source code was changing. My internal cost/benefit calculus was shifting. At a minimum, if you’re iterating towards an MVP, it should no longer be in a public repository.1

For most organisations, this would be a given anyway—just because a competitor might not understand what they’re looking at, why take the risk? Still, the chance of that code being made comprehensible by an AI assistant is pretty strong, so this scenario has gone from a poor strategy to suicidal strategy.

Okay, in all likelihood, you’d have built in a private repo anyway, but is that safe?

Although GitHub say that they don’t train their AI models on organisation private repositories, they’ve made no such claims on individual accounts. Indeed, any attempts to gain clarity on the matter using their issue tracker et cetera have been met with crickets. It’s been the subject of threads on Reddit and Hacker News, with some users openly not trusting the public statements given by the company. This comprehensive post by Simon Williamson digs into the matter in depth and still can’t reach a definitive answer.2

This does ultimately mean you’re putting your faith in a large organisation (Microsoft) with hugely asymmetric power over you to both a) not act maliciously and b) not act incompetently. Quite a big ask.

After all, from Microsoft’s point of view, they’re not in the business of stealing the IP of nascent ideas—they’re happy to buy them—so you’ve got to ask the question, how does it even get back to them if a) or b) above is violated, and your IP is leaked via Copilot?

This may be paranoia, but it’s worth considering. In my previous post on Hard Problems, I pointed out that in the age of AI coding agents you need hard problems with barriers to entry. A good way of ensuring the barrier to entry you thought you had (your unique idea, GTM and IP) remains unique is to apply some adversarial thinking to your development stack.

This is all rather bizarre for somebody that’s worked in crypto for the last 5 years—where your DMs blow up if you release anything that isn’t fully open-source, and your competitors brazenly just fork your code and deploy it—but I can see myself in the future taking not only steps back from open-sourcing system components that are part of products we are building, but even going so far as to host them on company Git servers instead of cloud platforms.

In my life as a consultant I talk a lot about TCO—it’s the reason most businesses should be using these cloud tools, rather than rolling their own—but by the same logic, I now find many of my own projects falling on the other side of that fence. There’s a cost to risk, and I think it would be naïve to not factor that in.

One changed line in a cloud platform’s Terms and Conditions that gets missed, a malicious or incompetent move, and you’re toast. It’s a Dark Forest now.

Watching Inside the Manosphere the other night,1 I was struck by the cognitive dissonance of some of the personalities interviewed.2 They still saw themselves as victims, plucky rebels fighting The Culture, even as their avatar sat in the Oval Office. Somehow, despite having won by any reasonable metric, they still saw themselves as the victims, because everybody else still does not agree with them.

This single-minded fixation on being right and being the centre of the universe is a classic Narcisstic trait.

It’s also how the politics of self-interest can rapidly become the politics of oppression, authoritarianism, and silencing dissent.

However, and here is where I might make myself a little unpopular in my own political camp—there’s also a grain of truth to their original complaint about being the victims of The Culture. Ironically enough, the reason is simply this: the Left (however we define it) has its fair share of narcissists too. Of course it does.

Like narcissists on the right, they are too busy thinking of themselves to start looking for win/win solutions.3

This has always been the danger of identity politics. Unlike the politics of community and collective vision and effort, it is atomised, and on some fundamental level, essentially narcissistic. At its worst, it enforces a zero-sum view of the world that sees everything in win/lose terms rather than allowing for the possibility of win/win.

No surprise then, that the far right has decided to fight fire with fire,4 and as events have shown, are much better at playing the game than the Left.

As others have noted, a plurality of opinions are required in order to sustain a win/win equilibrium. However, there’s also a flip side to this in order to avoid the politics of narcissism. Taking responsibility for their actions is something narcissists and psychopaths are both bad at, but in order to have your opinion, it also needs to be hedged against the consequences of using it.

Being ‘cancelled’ is one of those—a social enforcement of a behaviour norm. Good manners cost nothing, as the saying goes, and the people that think they can have their opinion without also respecting others shouldn’t be surprised if what goes around, comes around.5

Proof of Ownership, or PoO is a new consensus paradigm for a stablecoin-first future. It comprises 3 main components, with the goal being to ensure an endogenous token (a so-called ‘gas token’, to many) retains value even as a stablecoin is the main transactable asset on a network.

1. Only wallets with a valid soulbound or wallet-bound (and potentially even time-bound) NFT can transact. If your stablecoin is private-by-design (it should be) then the creation of this should be a zero-knowledge event where only a proof is written to the ledger. The existence of this proof is locally required at the time of issuing a transaction by a valid client, but it not audited centrally.1

2. All transaction fees (as well as gas) are paid to the network in the endogenous token. Transaction fees should likely be stable (but one would envisage greatly cheaper than the equivalent fees when e.g. using the Visa or Mastercard rails).

3. Optionally, some amount of endogenous token is staked or bonded to a validator before a wallet can transact.2 An antehandler3 filters all transactions that would increase a validator’s share of VP above a fixed sensible point (e.g. 5%, if you assume a valset of 20). Another option that would reduce the governance overhead would be to have all vals be within +/- a couple of percent of the average VP in order to be eligible for stake bonding.

This system requires an oracle, since the value of the NFT should be stable relative to the exogenous price of a real-world currency (let’s say the dollar, if you envisage a dollar stablecoin, but you should probably denominate it in something else unless you are (a) American, (b) an American company).

Thus, at least in principle, the value of the NFT is stable.4 You might additionally consider making them bound to wallets and non-transferable. Assuming an HD type wallet, you would need one for each sub wallet.

Although this is not the goal of the system, you could imagine AML/CFT or KYC measures occurring at the point of NFT sale.

In fact, it is possible to use this general scheme on top of say, a Tendermint style blockchain and make it work with a Sark payments layer.5

Relays (Porters) would have to take additional metadata and verify NFT ownership on-chain before saving transactions to their local transaction trie, but apart from that, the principle should work.6

If assets needed to be transacted more than once (for example if they represented an RWA and not a payments primitive, like a stock holding), then you could in principle continue to use a system like Sark with multi-hop updates and saturate the full provenance of an asset to determine its history in the case of regulatory need. Let’s call this extended protocol PoOP, or Proof of Ownership and Provenance.

Having typed this whole thing, and despite the date, I’m still totally unsure if I’m serious or joking.7

💩

I came across Evans F. Carlson a few years ago and was fascinated by his story. During WWII, the US military established what was essentially the first US special forces unit,1 the Marine Raiders. They created two Battalions, and gave their commanders extensive leeway in their organisation and tactics. Merritt Edson was given command of the 1st, organising it along the lines of existing unit doctrine, and Carlson the 2nd.

Carlson structured his unit more unconventionally than Edson. As an observer, he had seen the Chinese Communists fighting during the Chinese Civil War, and had been surprised by their effectiveness. He took from this two lessons—the importance of an éspirit de corps based on co-operation, and their size of unit organisation, which was typically 3 fighters:

Carlson used egalitarian and team-building methods: he treated officers and enlisted men with minimum regard to rank as leaders and fighters, gave his men “ethical indoctrination,” describing for each man what he was fighting for and why, and used the Chinese phrase “Gung-ho!” as a motivational slogan which he learned from the Communist forces during his years in China. He also eschewed standard Marine Corps organization, forming six rifle companies of two platoons each, and innovating 3-man “fire teams” as its basic unit. (Wikipedia)

The best thing about this story is the etymology of “gung ho,” which became a war cry for the US Marines as a whole after a propaganda film of the same name about the Makin Raid.

Carlson explained in a 1943 interview: “I was trying to build up the same sort of working spirit I had seen in China where all the soldiers [of the Eighth Route Army] dedicated themselves to one idea and worked together to put that idea over. I told the boys about it again and again. I told them of the motto of the Chinese Cooperatives, Gung Ho. It means Work Together — Work in Harmony.” (Wikipedia)

However, Carlson was mistaken. The phrase is simply an abbreviation for ‘industrial co-operative.’ Still, the fact that he adapted what he thought the phrase meant for a fighting unit is interesting.

After assembling two more Raider battalions, the force was consolidated into the 1st Marine Raider Regiment, with Carlson as executive officer. His 2nd Battalion was returned to ‘normal’ structure, but his ideas were adopted first by the Regiment, and then the whole Corps.

Carlson's 3-man fire team and 10-man squad organizations were adopted, first by the Raiders and then by the entire Marine Corps. Edson contributed the concept of a highly trained, lightly equipped force using conventional tactics to accomplish special missions or to fill in for a line battalion. (Wikipedia)

It’s the last point of that paragraph which hints at the ultimate fate of the Raiders experiment.

The reality was that in the Pacific war, there were relatively few opportunities for true Commando-style raids. Increasingly, landings were done in force, using amphibious tractors, and what was needed were more line Marine units. Moreover, the lightly-armed, specialized Raiders were disliked as an ‘elite within an elite’2 by some senior officers, and the 1st was eventually redesignated the 4th Marine Regiment and brought back in line with standard force organisation.3

Enter Robin Dunbar

Though the Raiders themselves were relatively short lived, their experimentation had a lasting effect, as well as being correlated in other areas. The 3-person structure crops up over and over—in terrorist cells4 and, subsequently, in the standard force organisation of most NATO militaries, where the smallest atomic unit is typically a fire team of 3-4 personnel.

This perhaps makes sense for a reason—Dunbar’s number. Though the methodology and conclusions have been disputed, I think the theory behind it has become so widespread because people can look around them and see some correlation between real life and theory.

The reason you will probably have heard of it is the idea that you can at most maintain about 150 meaningful relationships.

However, the model is stratified, and was developed by further research. Although most people will think of the smallest ‘circle’ of relationships as 5, there’s actually an inner one, 1.5.5 If we round this up to 2, we can see that one person with two close relationships adds to 3.6

For completeness, here are the layers:7

1. 1.5: Intimates/Inner sanctum

2. 5: Loved ones

3. 15: Good friends

4. 50: Friends

5. 150: Meaninful contacts

6. 500: Acquaintances

The takeaway is this: you don’t have to deliberately always structure around these numbers,8 but it maybe pays to be aware of them. When your teams break, consider if it’s because of crossing a threshold, or due to other external factors.

Finally, always remember that if in doubt, 3-4 motivated people can almost always get the job done.

In my mammoth series on AI,1 I discussed at length the economics, benefits, and potential improvements to ways of working and outcomes offered by AI tooling.

However, I also argued that in most cases organisations would not be nimble enough to take full—or possibly any—advantage of these opportunities, due to organisational inertia, team skill, exec capability, or what my former boss Craig would have called ‘hysterical raisins.’2

To quote Part 2 of that series,

The problem here isn’t that these gains exist—it’s that they are localised to teams that are able to capitalise on them. Most software teams are not bottlenecked by outputs, they’re bottlenecked by a lack of clarity on what to build, why they should build it, or even the permission to build it at all.

In Parts 2 and 3 of that series, I noted that although engineers inside large organisations were using the tooling and able to find gains, the majority of outsize impact appeared to be in small teams with agency, small organisations and individual domain experts using their own initiative.

Move Fast and Make Things

As I talked to people, I firmed up the idea that small, agile (small-a), motivated teams were the ones that seemed to benefit the most. I concluded that the “most effective users of AI tooling in my network either work at start-ups or scale-ups.”

All of this leads me to a prediction. We’re headed for a golden age of micro-consultancies.3

My experience in enterprise-scale companies leads me to think there are two main reasons that a consultancy is brought in:

1. An incumbent team lacks one of: headcount, skills, buy-in, trust, or agency to execute a given programme of work (sometimes all of the above).

2. An exec team bringing in a consultancy looks good to their stakeholders and shareholders, board et cetera.

I’ll note that in point one, generally headcount is not the key driver. Almost always it’s the organisational factors. Those pressures are likely to be more acute in the AI case, where smaller teams able to iterate and learn how to deploy tooling and best practice rapidly will outpace their less nimble competitors—both other consultancies and internal teams in larger organisations.

The cost of coordination within teams and between teams is often above-linear,4 hence the advantage that smaller teams sometimes demonstrate. Additionally, agents represent, well, extra agents, as the name suggests, in your organisational graph, which means they de facto add to your organisational complexity in the same manner as members of staff.

The question here is whether adding AI agents is above-linear, or sub-linear. In the latter case, they probably introduce less cost than adding headcount. Which is good, but anybody who’s ever encountered the ‘mythical man-month’ knows that simply adding resource to a project doesn’t necessarily speed it up.5

So you have a perfect storm where organisational factors will act as a sea-anchor for internal teams, introducing drag on their ability to up-skill and iterate with tooling. Meanwhile, other teams will find themselves with a powerful lever. Even if it’s a 20% or 30% force multiplier6 then that’s going to be a decent margin to build a business on top of.

For these teams, headcount is going to be aesthetically difficult to sell, so I would guess you’d be looking at teams of 3 product-minded engineers forming micro-consultancies to bid together on project work.

Large Orgs Should Take Advantage of This

For enterprise organisations that have the ability to trust (and procure from) organisations that small, it’s an easy bet to make—even at contractor’s rates, small teams like that are a small cost (relatively), and delivering a deadlocked or unrealistic project is potentially a very large upside.7

As a result, I’d argue that larger organisations that want to benefit from AI gains, or at least explore them, should attempt to streamline their procurement process to accomodate smaller players with a strong track record, and manage many small external teams delivering software.8

Inside many large organisations, staff augmentation contracts are already common. Many of these became fully remote during COVID-19, and stayed that way. For organisations already looking at the pre-AI trend of ‘near shore’ offshoring, it’s a much more palatable sell to meet in the middle between staff augmentation and offshoring.

These companies can then instead assign project-based work to small teams that may even adopt the full governance lifecycle of the client company and join the company Slack. I think that’s a decent compromise between internal governance and risk in exchange for the potential of a better result.

Obviously I’m biased, because this is basically what Envoy Labs has been doing for half a decade in the blockchain space, but hey, it’s an optimistic take if you’re an engineer that is:

• (a) entrepreneurial, and

• (b) product-minded.

Find some colleagues you trust and get started.

I, for one, hope I’m right—and to that end, this is a direction we as a company are going to explore pivoting to in the coming months. If you want to talk, then get in touch via our website or LinkedIn.

As promised, I’ve put in the (virtual) shoe leather cost to offer some thoughts over how to use the tools, the pros and cons, some of the most obvious pitfalls, and what’s left for creative work.

Hopefully this is the most positive section of the post series—I certainly meant it that way. Remember, that even when it looks like you might not, you always have a choice. Whether to participate, how to participate, and what your lines are.

5. A Tolerable Engineering Workflow

A provocative title, but hopefully this section will prove relatively optimistic.

As part of the research for this post, I reached out to most of my friends and former colleagues to get a sense of who was using the tools in what way. The idea that human computer interfaces have fundamentally shifted is interesting, but it seems too maxi, and isn’t reflected in day-to-day of the people I talked to.

Few-to-no experienced engineers are actually “vibe coding” production code. Instead, there’s a huge spread of heuristics for developing confidence1 in the AI output.

A common response was what one friend summarized as, “use the ideas, not the output.” That is, potentially don’t even have the tools integrated directly into your workflow, but have them near-to-hand, to prompt, pair with, bounce ideas, and gain context.

However, some take it much further. As part of this exploration, I asked the best engineer I’ve ever worked with (a former CTO of mine) whether they were using AI tools, and what their workflow was. It turns out the answer is yes, and the workflow they described to me is what I’d call very sensible.

They first generate and then iterate a spec document that multiple agents can read, using beads. They observe that a clear plan is crucial, “LLMs are mostly great for helping with stuff you already understand…but they do surprising stuff all the time…”

Then they iterate.

“An iterative process of

• describe the next step, maybe ask the LLM to research similar things in other languages

• Have the LLM implement it

• Review and iterate

• On a couple of occasions, decide it was

  • Unsalvageable, or

  • A bad experiment

• Then throw it away to either start again or go in a different direction

It wasn’t a quick process, but it was a lot quicker than doing it the old fashioned way… the LLM interaction process actually led to me finding and exploring things I didn’t set out to do.”

What’s most interesting is that they are solving what I would call Hard Problems—in this case writing an Algebraic Effects library in Elixir. Moreover, they say that the output is better than they could have achieved by themselves. One of the main benefits they identify is not only the speed of development, but also an ability to discard work, in light of a lower emotional investment in the code generated. This is also something my former colleague Paul mentioned.

In other words, the cost of iteration has been reduced.

Something to note about this person that seems to be really succeeding with the tooling is that they are a domain expert. They’ve been iterating solutions in this space for a decade and know what questions to ask, in addition to having a strong mental model of the problem space and grokking the key desiderata.

If you don’t have that, you might be in trouble. As they say,

“Mixed in with all the great, often it’s just stupid, wrong, or both—it doesn’t actually understand anything and it has no aesthetic sense beyond reversion to the mean, which makes it a powerful tool in the hands of an expert, and a chainsaw in a kindergarten for many other users.”

The idea of using the agent just to research for you is already a pretty valuable task—and verifying its suggestions aren’t crazy is relatively fast compared to say, parsing code. Gergely Orosz says HashiCorp founder Mitchell Hashimoto, “always [has] an agent running in the background doing something. He kicks off tasks before leaving the house — research, edge-case analysis, library comparisons — so work progresses while he drives or is away.” There’s a good quote from Mitchell on that episode of the Pragmatic Engineer podcast:

“There’s a lot of people like, ‘I don’t want it to write code for me.’ But just delegate some of the research part.” He uses agents for library comparisons, edge-case analysis, and deep research—not just code generation. “You don’t need to pick up on the ‘it must replace you as a person’ kind of propaganda.”

For what it’s worth, this is kind of where I am currently. So naturally, confirmation bias kicks in and I think this is a Great Take.

Returning to the cost of iteration, Will Faithfull, another Northern technologist—whose blog you should read—argues this is why a 50:50 split of features to refactoring is needed when using these tools. It’s easy to get carried away, and the tools have no aesthetic taste, beyond “whatever patterns they encounter; from the codebase, from the current session, from their own prior output.”

Why 50/50? We’ve anecdotally observed that AI productivity decays over the lifetime of a project, and that the rate of decay is strongly influenced by codebase quality and complexity. Complexity grows unavoidably as a project matures; there’s not lots you can do about that. What you can control is quality, and that determines how steep the decline is. What drives this decay? Pattern amplification.

It’s worth saying that some working in AI claim that the agents do now have aesthetic taste. However, one of the models mentioned there is the one my former boss is using. I trust their judgement more. Moreover, that statement is couched in a piece that is (a) pure ‘Criti-hype’ fearmongering, and (b) by somebody that works in AI. Caveat emptor.2

Finally, back to Will’s point on diminishing returns: I think there’s likely an inverse relationship to TCO, depending on how good your decisions are about when to ‘do the work’ and accrue domain knowledge, versus when to offload that work to an AI tool. Lack of understanding of the domain, or codebase, represents a hidden cost that isn’t accounted for in the cost of production or sale—something in economics we’d call an externality.3

At a meta-level, there’s a cross-project version of this externality—deskilling. If the argument that users of AI tools gradually deskill holds true, then you would expect this externality related cost to grow across all your projects.4 The more your team uses the tools, the worse they perform, over time, across all work—unless you have a way of maintaining their skill base. Needless to say, this is a hard problem.

As discussed in Part 2 in the context of organisational design, what Stuart Winter-Tear calls Speeding One Cog Breaks the Machine5 is a serious risk in software projects. If the process improvement isn’t holistic, all that happens is you move the bottleneck point within a meta system or process and flood somewhere else.

Anecdotally, that point seems to be code review, and I’ve personally seen not only code review, but often additional post-hoc sense checks of code have become necessary. That’s a big, time-consuming piece of work.

Also key to this all working in the long-term is a missing part of the current picture—completely open-source AI models, governed in a way we can trust.6 If you optimise a workflow on top of technology that will eat our societies then you’re still rearranging the proverbial deck chairs on the Titanic.

This is a super hard problem and I’m not sure I have any answers at all, let alone any easy ones.

6. You’ll Own Nothing, and You’ll Be An Start-Up

This is pretty short, because the argument is simple. Integrating AI at the heart of your business venture is likely not a good idea unless you’re a large business, an incubent, or operate in an area with high barriers to entry.

If you build your business on somebody else’s model, then you are accruing wealth, power and money to their platform. If your business fundamentally requires their model, then you’re in even tougher straights. You have no USP, you have no real barrier to entry, and perhaps you have no business.

Think of the Commandment of Control and Commandment of Entry, which I’ve discussed before, and the point should become clear. You’re not a SaaS product or a platform product, if your start-up depends on an AI model, as other writers have noted; you’re a margins player, and your business can be disrupted by a new competitor at any time, for a low cost.

7. A Manifesto for Non-Participation

As should probably be pretty clear by this point, I think there’s a moral, political and creative imperative to resist, so far as is practical, the use of AI tools in the creative process.

I’m more of a pragmatist when it comes to the world of work, but I’m also not jazzed about the level of asymmetric power at play. I’m happy to use the tools to a point, but by doing so feel complicit in a political and ideological programme that is toxic. I don’t know how to square that circle.

To take it back to a reference from the start of this piece, this asymmetric power recalls the plight of weavers at the advent of the Industrial Revolution. I believe my name is French Huguenot, so maybe I have a bone-deep latent sympathy with the weavers and Luddites for that reason, I don’t know.

You might be in the situation where if you completely reject AI tooling, you might lose your job or face sanctions—I’m, to be honest, in a similar situation in my profession.7 So don’t get fired, but don’t blindly accept, either.8 Remember, you always have a choice.

At the very least, non-participation is an act of resistance. Non-participation is also an optimistic act. As far as I can tell, many people kind of hate AI, but go along with using it out of fear.9

AI enthusiasts and the e/acc crowd can beat you over the head with non-participation and can you a Luddite.10 However, as I’ve said, this is an optimistic act. It’s doubly so when you combine it with a call to arms for creativity.

Nobody can tell you not to create. Every scribble is art. In Understanding Comics,11 Scott McCloud advances possibly the best definition for art I’ve ever seen:12

“Any human activity which doesn't grow out of either of our species' two basic instincts: survival and reproduction.”

Even then, this might be too narrow!

Don’t ask for permission. Create. Your work is valid. Your thoughts are valid, your aesthetic is valid.

Don’t replace yourself. Don’t deskill yourself. Don’t empoverish yourself. Do the hard thing because that is all there is to life.

Life, like creative suffering, is what you make of it.

Acknowledgements: Thanks to the many people that fed back on earlier or draft versions of this series, including but not limited to: Jon Stone, Craig McMillan, and Rob Bowley. Cheers for the conversations while I worked out shower thoughts to Andy Gray, Geoff Goodell, James Morgan, David Scott, David Alesch and Jack Gray. Thanks also to all my network that I have bugged about AI tooling, workflows, best practice in their places of employment and for opinions. I hope I’ve done your thoughts and feedback justice.

In Part 1, we talked about how chasing productivity as the cure of all ills was a category error. In Part 3, we will talk about the potential positives of AI tooling, as well as what agency you, as a citizen and user, can exercise on our precarious Now.

In the next two sections, we’ll discuss why the current political programme behind AI is feudalism at best, and fascism at worst, and why for most organisations, the potential gains are unlikely to be realized.

A better future is possible—but we have to call out that the present way this programme is structured sucks.

Let’s start by discussing the best thing I’ve seen on the subject—a video essay by the musician Adam Neely (embed below).

3. AI Accelerationism and Fascism

In the essay,1 Neely starts off by simply addressing the ethics and creative dissonance involved in AI’s incursion into art, but then goes much deeper. He connects the intellectual lineage of the silicon valley VCs and e/acc enthusiasts2 to Italian futurism, and specifically the poet and writer Filippo Tommaso Marinetti.

Neely points out that in that case, the logical conclusion of such a reductionist techno-optimist manifesto of ‘speed at all costs’ was Italian fascism—Marinetti was the co-author of the Fascist manifesto. The fetishisation of speed and progress implicitly results in a suppression of critique, something that we see with AI skepticism.

Bringing things back to the present, something I had never noticed until Neely pointed it out was that Mark “conehead” Andreessen has a nod to Italian Futurism on his wall, and repeatedly paraphrases Marinetti in his own writing. This should be concerning when considering incredibly powerful, wealthy people that switched to supporting Donald Trump when potential AI regulation was on the cards at the tail end of the Biden administration.3

As Neely points out, for every argument like democratization or removing shame that holds up to some scrutiny,4 the end result is the same. Though these are arguments applied to AI art, twist the perspective just a little, and you see the mental contortions applied to most areas of creative work in order to justify AI maximalism. If you’re on LinkedIn, look at your feed and you’ll see it.

These AI capitalists aren’t looking to solve the root cause—lowering barriers to technical or musical education, or working on making young people more resilient to failure, or to feel less ashamed of emotional cultural expression. They’re not looking at structural access to skills or opportunity, such as Universities, apprenticeships, or art schools.

No, they want to sell you a product. A product trained on everybody else’s intellectual property, and walled-off in their platform for them to rent-seek in perpetuity. Great.

Policy programming music critical thinking art essay writing academic work learning being a creatively fulfilled human doing stuff is hard and requires collaboration, interaction, struggle and self-reflection. It’s much easier to just solve problems by lowering the bar.

Then we can repeat the present forever. Which suits these guys, because they’re control freaks. They control the present, and for all their bluster about progress, they’d rather no future happen, than a future they can’t control.

Or maybe it’s simpler than that. A lot of the companies wrapped up in this are seeking to maintain stock valuations as ‘growth’ companies, with potential valuations far in excess of their actual P/E ratio. In such a context, their incentive with AI, as with blockchain before it, is to claim that not only are they on the verge of changing the world, but they’re the only ones that can do it.5

Some of the most powerful people and institutions on the planet want you to believe that the changes they promise are coming are a foregone conclusion, that they are inevitable. This is not the case—what they fear most is that you, we, will realise that in fact we have agency and a choice. No surprise that those who stand to gain most are those that say these changes are part of an inevitable ahistorical process of ‘progress.’6

It’s also no surprise that these same people are generally the narcissists and psychopaths interviewed and profiled by Nate Silver in his book On the Edge. I discuss these behaviours at length in my post reviewing his book.

You’ll have read the endless blog posts and think-pieces about how AI is already taking entry-level positions, even as the FT suggests that the underlying cause is simply normal employment cycles, or the effect of interest rate changes on hiring:

“Linking job losses to increased AI usage rather than other negative factors like weak demand or excessive hiring in the past conveys a more positive message to investors,” points out Ben May, director of global macro research at Oxford Economics.

An interesting data set is a National Bureau of Economic Research working paper, surveying 6000 executives. In it, the respondents predicted AI use would impact employment by either cutting headcount by up to 0.7%, or by increasing it by up to 0.5%. Though obviously the data will have been primarily on older models, they report a marginal impact, or no impact, on productivity.

This isn’t real, the AI isn’t coming for your job—at least not tomorrow—you’re being coerced.

Will your job change? Yes, it looks like it. Will it be only the boring parts that are left? Yes, it looks like it. Will it pay less? Yes, it looks like it.

All of those are bad things, but it’s not yet a doomsday scenario. For one, as I describe in the next section, put the middle class out of work, and maybe the whole gameshow ends anyway, and for another, we already have a political crisis caused by stagnant real wages, let alone falling ones. This is the same political and economic problem we were already facing, just accelerated.

Moreover, the AI can’t take your hobbies and your skills. Some things are worth doing for the sake of it. I talk more about that in the final section, Section 7.

Neely concludes by advocating for the virtues of Service, Patience, Craft and Beauty in creative pursuits, and that’s not a bad North Star as we navigate an increasingly uncertain world.

4. These AI Gains Aren’t Possible Anyway

Let’s take a moment to consider the AI productivity gains again.

Even in light of David’s post, what I’ve seen the tools do, and the workflow outlined in Section 5, I’m still not sure I buy it.

Why? Two big reasons. First, organisations can’t even take advantage of 20 year old technology, let alone frontier tech, and second, that our economies would collapse if they could.

Let’s talk economics first.

At the beginning of this hype cycle, there was a much-paraphased report by McKinsey on AI productivity gains that today, honestly, might pale in comparison to some of the most frothy predictions. Quoting Lee Vinsel’s piece on ‘Criti-Hype,’7 which you should read, we see a typical labour market prediction.

For example, in its 2017 report, the AI Now Institute, which is associated with New York University, paraphrased another report from the consulting firm McKinsey claiming that 60 percent of occupations would have 1/3 of their activities automated.

The reasoning is very simple here.

If white collar workers were laid off, middle-career, in the numbers that are (a) required for these AI companies to pay back their investments, (b) with enough revenue back to the AI companies to pay for the trouble, and (c) without any job to go to, our societies will collapse.

Let’s break this down.

• First, in order to pay back the scale of investment, the level of adoption has to be on the order described, or not far off.

• Second, the cost model of AI tools needs to change in order to cover its full cost (at the moment many companies are making a loss). I don’t imagine they will price in their externalities—what capitalists ever have?—but this increase in cost will also mean that the marginal saving between the employee and the AI will go down. It’ll likely still be a large enough gap, however.

• Third, these workers become unemployed. Potentially for good. They go from productive members of society with social capital and a good standard of living to not being able to pay their mortgages.8 Let’s put aside the social unrest that would cause (I cover that below, and yes, it involves guillotines), and focus on what happens within an economy. What do you suppose happens to sales of IKEA furniture when 19.8% of young professionals cease to exist?9

A version of this demand-side liquidity crunch is what cryptoeconomies have speed run in bear markets for the last few years. First, products have to compete for a reduced pool of spending power (liquidity), then some go bust and others consolidate. However, the pool of available money continues to decline, and eventually the Foundation typically has to step in and either boost demand or subsidise the projects and validator set required to keep the chain alive.

Real-world economies are no different, and this kind of sustained, demand-side shock is already a issue in developed economies. In many, real wage growth has been stagnant for thirty years, and credit has taken its place. Obtaining credit relies on a job, on security, and on predictability. Offering credit when it cannot be repaid is what led to the Subprime Mortgage Crisis and the 2008 Great Financial Crash.

See the problem?

Demand can’t be stimulated by simply printing money, or creating more credit, because people will be unable to pay it back. Second, a Universal Basic Income (a good idea in principle) can’t plug the gap, because tax receipts for the state will decline. They have fewer people working on the one hand, and less sales tax from businesses on the other. Simply—how can states afford to pay for it?10

Our populations will continue to age, driving up the cost of social security for old age pensioners onto those that continue to work. They will conclude, as many already do in countries like the UK, that work does not pay.

However, in a situation where many will be unemployed and in worse material circumstances, people will continue to work, but with increasing resentment. Look, this isn’t that different to our existing paradigm—the trends, after all, are the same—it is just far more stark.

This is a negative cycle.

A social upheaval I think about a lot isn’t just economic. After World War One, soldiers returned home to find economic chaos, and in many cases—especially in the defeated powers—punitive sanctions in place. They found themselves at peace but unable to survive.11 I will always remember the words of my lecturer in describing the result, “every capital East of Paris fell.”

That’s what happens when people really think the social contract has been broken.

When people are laid off, while those that do work do so in increasing misery,12 all the while seeing billionaires and platform capitalists doing just fine, planning colonies on Mars to keep them safe from rogue AI, what do you suppose their conclusion might be?

Or, as Jack Clark from Anthropic says in Nate Silver’s On The Edge,

“People don’t take guillotines seriously. But historically, when a tiny group gains a huge amount of power and makes life-altering decisions for a vast number of people, the minority gets actually, for real, killed.”

Your Mars colony is only a decent escape plan if you don’t get caught and guillotined on the launchpad, Elon.

All of this is before we consider what organisations are actually in a position to capitalize on.

As we saw in the last section, it’s likely that so-called “AI job losses” may just be normal cyclical economic behaviour from firms, with a bit of PR spin on top.

No doubt that as models improve, entry level positions will take a hit, but maybe once the macro position improves, the impact will be less than what we’re seeing right now. I graduated into the fallout of the Great Financial Crash in 2008 and I remember how lean the first few years of my career were in terms of finding any work I could.13

I’ve spent a long time working and consulting in both start-ups and enterprise companies, and I can tell you that for every start-up (or scale up, for that matter) able to pick up the latest and greatest tooling and run with it, there’s an enterprise organisation that would fail to realise substantial gains whatever happened.

This isn’t something I’ve made up. When I was speaking to former colleagues for the next section, I found many inside large organisations that used new tooling, with frontier models for their day-to-day work. Some said the use of tools was encouraged, some said it was a soft mandate. Others said that it wasn’t required, but the best engineers were getting incredible work done with the tools.

The problem here isn’t that these gains exist—it’s that they are localised to teams that are able to capitalise on them. Most software teams are not bottlenecked by outputs, they’re bottlenecked by a lack of clarity on what to build, why they should build it, or even the permission to build it at all. None of these things are strictly an engineering problem, and solving this complicated interrelation of impulses is part of the reason that agile software originally came about.

Still, what we’ve seen in large organisations is that agile adoption—in the sense of “work iteratively and systematize your successes”—has been slow, difficult, and often has failed. I don’t think these organisations are likely to see any huge benefits soon, simply because the average rate of productivity in the company is the sum total of all teams, well managed or not, and the ability to deliver the right thing, rather than just deliver full-stop.

Not only will the average be dragged down, but delivering the right thing, so far as I can tell, is not something that AI can necessarily help with—other than perhaps information synthesis after you’ve talked to your users, or by triaging help tickets and identifying themes of problems with your product.14 My former boss Rob has written several posts on this subject, and I think he’s on the money when he says, “typing was never the bottleneck.”

Moreover, the maintenance of these new projects still works the same as any software project—it’s typically expensive, and dominated by the cost of change and the number of staff that understand it. TCO (Total Cost of Ownership) has to be factored in when using AI tooling, and although the tooling can digest and answer questions about a corpus of code, ultimately a human needs to grok it before it can be reasoned about, or safely changed. That part hasn’t changed, and the cost is still high.

The current pricing model of AI models obviously doesn’t take into account externalities—environmental, social, et cetera—but it also doesn’t reflect the full cost of production. As soon as these models dominate, they will have to increase their prices, and at that point you’re in a classic vendor lock-in conundrum as a business. In that situation, TCO of using the AI tools might be only marginally less than doing things the old-fashioned way. It might even be higher—that remains to be seen.

Start-ups, and many scale-ups often find themselves working in a pseudo-agile way by default, simply out of necessity (most would not describe it that way). Put simply, they only eat what they kill, so they’re often applying the principle of trying ideas and abandoning them fast, trying to find product market fit and make the company work. This attitude makes me think that only small companies are likely to reap the real benefits of AI tooling—and then only with experienced staff who can execute well within a framework of building the right things.

The apparently most effective users of AI tooling in my network, either work at start-ups or scale-ups, which certainly influences me in this hunch. Every post like this lengthy post on how the Software Development Lifecycle is Dead strengthens that suspicion. It’s a reasonable post—if you work in an enterprise company—but much of it was never relevant in the kind of small, nimble companies I’m thinking of anyway. However, it’s a small sample set and maybe I’m wrong.

It’s also possible I’m wrong that the improvements in LLMs will hit a plateau and the curve will flatten, with each new generation only showing marginal gains. We will see.

In the final part of this series, we’ll talk about how you as an engineer (or creative) can use the tools, manage the risk of deskilling, and what is left for creative work.

Acknowledgements: Thanks to the many people that fed back on earlier or draft versions of this series, including but not limited to: Jon Stone, Craig McMillan, and Rob Bowley. Cheers for the conversations while I worked out shower thoughts to Andy Gray, Geoff Goodell, James Morgan, David Scott, David Alesch and Jack Gray. Thanks also to all my network that I have bugged about AI tooling, workflows, best practice in their places of employment and for opinions. I hope I’ve done your thoughts and feedback justice.

This post is part of a series: Part 1 | Part 2 | Part 3

I’ve come to much the same conclusion as David Whitney in his devastating post, Existential Dread and the End of Programming, which I read while copy-editing this one. The AI tools are here, they’re Good Enough, programming as we know it is over, and, taking a wider view, we’re not institutionally or individually ready for what comes next.

Which sucks. But we’re programmers. We solve problems, so let’s try and put the dread aside and work the problem.

This essay is an attempt at that. Maybe I’m being too optimistic. Maybe Hard Problems are Not Still Hard. Maybe they’re easy now. Who knows.

David says,

“While so much of what we do in software is remixing existing concepts, innovation isn’t going to come from an existing corpus of information, but business innovation might. You’ll still need those experts if you want to do something actually unique.”

Which I suspect might be true. That’s my thesis in Section 2, and my previous post.

In this series I’m going to argue three main things:

1. AI is going to change the world of work, but probably won’t take your job. In fact, many organisations might see little benefit, for organisational and TCO reasons.

2. Identifying productivity as the key problem we face is a category error—it’s creativity that’s the problem, and AI doesn’t solve this.

3. You can and should reject the political programme behind AI in its current form, doubly so in your creative pursuits and hobbies.

This article is split into three posts (coming this week—I will update these links as they go live):

• Part 1 is about how AI addresses the wrong thing: a productivity deficit as opposed to a creativity deficit.

• Part 2 is about the political programme behind AI, and why it isn’t great—as well as why we probably won’t see any huge changes because of inertia, particularly in large organisations.

• Part 3 is about how you, as an technologist, can use these tools, why that’s still a hard problem, and what is left. I try to end on a hopeful note.

Let’s go.

1. AI Is Boring, But That Won’t Save You

The more I work with AI1 tools, the more I find myself becoming apathetic. I’ve managed to avoid them entirely for my writing, and perhaps this is why I find my interest for it so reinvigorated of late. However, programming with too much AI in the loop (I’ll tolerate ‘turbo-autocomplete’, having spent some years used to Rust and Haskell compiler hints) is quite joyless.

Before I continue, if you’re already typing “luddite” in the comments2 then you might want to skip to section 5 in Part 3, where I argue for what a tolerable workflow might look like. It’s not dissimilar to David’s, outlined in his post.

In a future post (not part of this series), I’ll make the optimistic argument that being able to iterate quickly and make outsize gains may usher in a new era of small, high-impact software shops. It’s certainly the direction I see my consultancy, Envoy Labs, going.

Still, like David, I find the process exhausting, perhaps more so. I will also note, that for me, while such a workflow is fine for work, it is still (a) kind of joyless, and (b) not sustainable in the long-run unless we train engineers the old-fashioned way, probably without, or with minimal AI tooling. It’s not something I have much interest in adopting for my own personal projects, even if it may present a significant business opportunity (that we will be seeking to capitalize upon).

I began programming as a creative pursuit, an extension of writing that allowed me to build abstract, yet concrete things out of words and symbols. If cyberspace is real, then programming is a form of magic, conjuring real things from the ether by giving them their real names. The writer Alan Moore argues that prose is already a form of magic—programming then is an order of magnitude more powerful.

Don’t get me wrong, AI has some uses—off the top of my head, making adjacent similar things, like lesson plans if you’re a teacher, or speeding up making low-effort semi-creative work, such as shitposts or memes for your friends.

In the programming world, I’m happy to get an agent to spit out additional tests for code that I’ve already worked out the hard parts of. My friend Paul—who is a very good engineer—has a really good use case. He uses them to synthesise information, so he can ask questions as he thinks about design.

That’s a pretty good, time-saving use. It’s not world changing, however.

The key is that simply automating low-hanging drudgery that in many cases we simply would not have bothered doing if it wasn’t for the AI assistance isn’t going to juice the valuations of the AI companies.

Or, as another writer described their usage of AI,

“On the other hand, if I had needed to pay someone proper money to do it, I probably would not have done it at all.”

The investors in these AI companies, as well as the incumbent tech giants rushing to put forward their own AI solutions, all have a colossal vested interest in arguing that AI can 10x, 100x productivity, or do jobs that otherwise wouldn’t have been done. Until recently, their argument was still pretty thin.

However, something changed in the last generation of models, and that’s that they got reasonably good at spitting out decent, usable code. They went from reliably knocking out short scripts to reliably knocking out plumbing code for backend applications. Cue another round of hype over productivity gains.

The thing is, even if AI does deliver those gains (it doesn’t in most cases—organisations need to be set up to realise them, see Section 4), it wouldn’t matter, because we are optimising for the wrong thing. Again, for the counter-argument, scroll down to A Tolerable Engineering Workflow (Section 5).

Moreover, there’s the question whether you can prompt to something novel. Weak Sapier-Worff suggests the answer is, possibly not. Still, this might not matter. Most of creative output is a remix or repurpose of something that exists, after all.

Thus we arrive at our subject for today. Why creativity isn’t the same as productivity, and why the political programme of productivity and speed at all costs is not only toxic, but misses the mark of what we actually need.

2. Creativity versus Productivity

On some level, AI productivity increases are just the story of doing more, with less. It’s sort of like austerity, except applied to putting in effort and using imagination. But, as the old saying goes,3 “if less is more, just think of how much more ‘more’ will be!” Creativity is the answer to the scarcity mindset implied by AI maximalism.

AI is the presumptive owner of the narrative of progress, obsessed with the aesthetic of the future. This isn’t new; the crypto scene was obsessed with 80s retrofuturism and the cyberpunk aesthetic. The internet as jurisdiction and peer-to-peer networking are kind of both edgy and cool. Moreover, if the aesthetic, or the performance of crypto, actually matched the reality, then perhaps it would have formed part of a creative answer to our stalled present.

Agency is one of the things we lack. Alienation is one of the things we feel. It may be dorky to say it, but peer-to-peer networking implicitly has a promise of both agency and community. Ask anybody that file-shared using BitTorrent in the old days and they’ll tell you—it wasn’t just about getting that new record first, it was also about something less tangible, and more vibes-based.4

Of course the reality of crypto was mostly just gambling, but I can tell you, when a network genesis happened, you felt something irrational, something—dare I say it—emotional, or hopeful, about the peering process starting to spit out blocks. Even if there was no reality to back the performance, well, we were still engaging in a shared performance. Maybe, after all, what we were doing in crypto was at least part performance art.5

As a result, there’s a bit of me that has a tiny bit of sympathy for the AI maxis who genuinely believe the hype.6 It’s quite a natural human feeling to look on something that has the appearance of novelty with optimism.

The main selling-point of AI (and of blockchain before it)7 is the idea that we can boost productivity, and create a new margin of growth that will unstick semi-stalled developed economies. This is the inevitable logic of ageing populations that have to support more retired people with fewer working-age people, but it’s also an often-unchallenged attitude in the political and business class as much as it is a function of demographics.

The more I think about it, however, the less I think we’re looking at a productivity crisis—we’re looking at a creativity crisis. Perhaps the need for ‘more creativity’ is simply perspective, and I’m dismissing the ‘wrong’ creativity.

After all, much as it might be hugely overstated, the creation of AI tooling, marketing and proliferation represents a huge creative endeavour. It’s not just ekeing out additional efficiency from existing paradigms. Even if the AI programme fails, it has, as a doctrine, at least been a novel intellectual and narrative one.

Still, that’s just one novel doctrine or movement. I think we can do better. Much better. I think moreover, we can and should demand a plurality of novel movements, and seek to bring them into being.

What if we more effectively oriented around creativity such that we could imagine the new, both in our day-to-day work and at a higher level, in our institutions, businesses, communities and politics? Sort of a radical optimism made doctrine, if you like. On the face of it, this isn’t that different to e/acc—it’s just less pedantic, as it doesn’t make the arrogant assumption that capitalist technology is the only game in town for the transformation of our world.

Even if we confined ourselves to the narrow sphere of economics, I could illustrate the difference thus: AI means that barriers to entry are lower, but retained knowledge and context is kept by the AI model owner, as they iterate on their model. They operate, essentially as a feudal landlord.

In a sense, they’re the end state of what McKenzie Wark called ‘Vector Capitalism.’ They own the platform and the means to connect the dots in the economy. More straightforwardly, the landlord analogy is one that’s been made by former Greek finance minister and economist Yannis Varoufakis, among others. As a result, the tendency is for incumbents to benefit more, even if there is the illusion of greater agency for ‘creative destruction’ and challenging incumbents with new ideas.

What if instead of this, in the most extreme example, we ignored the AI tooling and just tried hard to come up with new ideas for businesses? What if we aggressively subsidized that principle as a government, accepting the huge failure rate of new businesses? What if we essentially took the VC model for risk and made it a part of the social contract?

What if we said, “we, society, will take a risk on your novel idea, citizen, if you at least try your hardest to make it a reality.” That’s opportunity and agency in action.8

Then, instead of the performance of disruption, we might get the reality of disruption. Then, instead of the performance of agency, we might get the reality of agency. Then, instead of the performance of productivity, we might get the reality of creation.

AI claims ownership of the aesthetic of progress, but that’s all it is. it is not a reality. It can only be an endless repetition of the present, with each repetition endlessly accumulating capital in the hands of the feudal AI landlords.

Only creativity can bring forth the novel, and this is not something that AI tooling in its current form can offer. It’s not clear that it is something AI tooling will ever be able to offer.

Even if it is, I suspect the reality, viewed through the lens of tech, will be something like what David Whitney describes,

“These engineers will need to have taste, and they’ll probably be involved in early hand writing of some categories of code to establish patterns for the machines to follow in the first instance, but likely will accelerate to the point where traditional workflows of pull-requests and reviews don’t make sense when faced with the pace change can be made.”

These engineers will need to be experts, and fully understand their domain (see Section 5); from whence will they come unless we (or a business) subsidise their traning and apprenticeship? In fact, David’s discussion of how a novel software project might be run sounds very like the research projects we’ve been working on in the Future of Money group.

Most of the work thus far has been exploratory code to understand the domain. Many younger contributors have used a lot of AI tooling. I’ve mostly worked by hand, though I expect my usage to flip in the next project stage. This seems a natural progression if we are to benefit from the tooling, but the key thing is that the early, exploratory, novel phase of the project had to be paid for first. It’s not actually as fast or cheap as it looks.9

Thomas Kuhn’s argument on the structure of scientific revolutions is often referenced by tech types, usually with the assumption that paradigm shifts must necessarily happen in the large, due to technology, but I don’t believe that.10 I believe that creativity is the spark for the paradigm shift, and this can happen in the small and the large.

Still, it is incumbent on us as societies, if we buy the AI argument (progress)11 to try and make it manifest by actually catalysing the creativity that brings forth the novel. By definition, even if we limit ourselves to start-up enterprise and not novel ideas in art, culture,12 politics, et cetera, we cannot know ahead of time what these new businesses and ideas will look like. We must simply write the cheque and give people the space to create.

In case I haven’t been clear enough—only the novel can solve the problems of our societies. Thus AI—in its current form, at least—is a bust.

Furthermore, in case I haven’t been clear enough on another point—the novel doesn’t have to be technological. It can be cultural, artistic, or political. Subsidize a new generation of polytechnics and art schools—we do not know from where the ideas will come.

Much has been written on the AI-inevitability narrative of late—there’s great posts by Cory Doctorow13 and others that go into the subject in great depth. However, the best commentary I’ve seen so far is a feature-length video essay by the musician Adam Neely.

We’ll talk about that in Part 2.

Acknowledgements: Thanks to the many people that fed back on earlier or draft versions of this post series, including but not limited to: Jon Stone, Craig McMillan, and Rob Bowley. Cheers for the conversations while I worked out shower thoughts to Andy Gray, Geoff Goodell, James Morgan, David Scott, David Alesch and Jack Gray. Thanks also to all my network that I have bugged about AI tooling, workflows, best practice in their places of employment and for opinions. I hope I’ve done your thoughts and feedback justice.

A broken clock is right twice a day, and some of my old blog posts come around every few years. I’m in various technical leader chats, forums and face-to-face meetups, and one idea seems to come up over and over again. Whether it’s said over a pint or from a stage by a speaker, there’s this perennially popular idea that to succeed in tech leadership, you don’t need to be technical.1

Let me be absolutely clear: this is only ever true if, and only if2 the leader in question has independently developed the same gut feeling and reasoning that the apprenticeship path of software engineering tends to develop in most competent engineers.

And what is that reasoning? The goal of technical leadership is systematization.3

I wrote in several posts about the apprenticeship pattern of ‘software engineering’ described by Adewale Oshineye and Dave Hoover in Apprenticeship Patterns,4 and I think that an important result of taking this approach to a technical career is necessarily that your analytical and problem-solving skill, as well as your simple gut instinct become highly trained.5

That training (in my experience) tends to express itself as pattern recognition. Knowing the detail, it is possible to then interrogate whether a pattern you have identified (or intuited) has any basis in reality. This is where the ability to switch back into the detail matters.6

In essence, you have to be very senior before having no grasp of the detail is describable as anything other than a weakness. I’d use the analogy of me trying to explain the benefit of functional-core, imperative-shell (pick your code organisation habit of choice here) to an entry-level developer.7

Without experience of at least one large, production project experiencing growing or maintenance pains, not only does that developer not have any context for understanding the deeper implications of the argument (or perhaps even the foundational need for change or improvement at all), they also have no experience of it on either a hands-on level, or perhaps, more importantly, on an emotional level.8

There’s always a risk of over-priviledging your painful failures like the character on the Five of Cups,9 but it is through systematizing these lessons that truly exceptional individuals self-actualize (to use a dreadful term).

The counter-argument is of course clear at this point. The goal of technical leadership is systematization, and that is independent of deep technical knowledge.

This means, surely, that tech leaders don’t need to be technical.

Obviously, I can’t disagree with this point—I can merely restate that yes, of course not all engineers will be able to contextualise and systematize their successes and failures, but they empirically have much more opportunity to do so than those that are non-technical. They will encounter these situations daily, monthly, yearly over the course of a ten or fifteen-year track into management. That is an observable, testable, quantative fact.

The fallacy, I think, being played into is that because systematization is independent of deep technical knowledge, it is somehow mututally exclusive. This is not the case.

Moreover, technical leaders do not exist in a vacuum, and they will have talented non-technical colleagues, likely in product, design, delivery, and other disciplines, to collaborate with. Again, ability to collaborate is independent of deep technical knowledge, but it is not mututally exclusive.

So if you’re an effective technical leader from a non-technical path—celebrate yourself, and know that I’m not getting at you in this post. I’ve worked with many excellent non-technical CIOs and CTOs. No doubt I will work with many more in future.

However, if you want to be a technical leader, doing it the easy way definitely means doing the long yards in a technical role and reflecting on your work as you go. The really harmful thing about this debate is the same thing I identified in my old post (nearly eight years ago!)—that the advice that they did not need to pursue a technical track was being given to folks in their early career. In the general case, this is objectively not correct.10

Still, the delta between this, which once upon a time would have been called ‘web design’ and ‘web application development’ is not great. In fact, the delta from this to distributed systems engineering for most non-technology companies is not that great.

The reality is that today those companies are probably making ‘build/buy’ decisions rather than ‘vibe code in-house/buy’ decisions, but even so, the point stands. However, soon—and sooner than I would have thought if you’d asked me two years ago—those companies are going to be turning to AI for the majority of their skilled work, both inside and outside their tech function.

Most of software engineering (especially in these types of companies, say a retailer or other large company whose main product isn’t software or services) is simply what I like to call ‘plumbing,’ that is, connection layers between different systems, different APIs, or even different layers within an application stack. Not all of these are created equal in terms of complexity of implementation for a given project, but generally you can see why AI agents could at least help substantially with the velocity of development, even if there’s no move to replace technical staff any time soon.

Now this isn’t ideal for me personally—I tend to say that my biggest skill is information synthesis, and AI tools can digest a far larger corpus of information, and infer things more succinctly and more quickly than I ever could. Bad news bears.

Hard Problems Have Novel Solutions

The quickest readers will have noticed though that there’s an absolute ocean of problems that don’t fit into the categories above. Off the top of my head, the broad ‘anything that touches real money,’ category, before we talk about large scale data-processing and engineering. Sure, there are plenty of low-hanging fruit in data science and ETL tooling, but the difference between a lazy iterator, an eager iterator, a parallel iterator et cetera in such situations (language depending, of course, and using a Clojure example from reflex) can be huge.

Sure, LLMs can identify this and write sensible code in many cases—but they need careful supervision. In the end, it’s likely to be (at best) a force multiplier than a complete game-changer. If you’ve worked with REPL-Driven Development (RDD) or the advanced compilers and formatters of languages like Rust or Haskell, you’re already used to advanced tools and hugely iterative workflows, and perhaps it is best to regard LLMs as falling mainly into this category of “spiking” tools.

Experienced engineers often “spike” ideas and “refactor” later, and in a sense, generating code via AI tooling is the purest, lowest-cost “spike” imaginable, as long as you can express your intent to the agent conscisely and get useful output. Two big ifs.

Still, many of the examples just given are areas that an AI would not have been able to tackle two years ago, but might just have a shot at today. Even if the prompt spat out unoptimized code, with an experienced engineer checking the output, a large task (writing the code and optimising the code) could be a much smaller one (optimising the code). As others have observed, this is a complete paradigm shift in the job description.2

Even so, there’s also a class of problems where it’s not so much that there needs to be a human in the loop for risk mitigation; it’s that it’s simply unclear whether a novel solution could be found without one. Don’t get me wrong, a human might not figure it out either—but here I’m thinking about long-lived research efforts such as the development of something like Malachite as a Rust rewrite of Tendermint (or even CometBFT as a continuation of the development of Tendermint).3

The ability to find novel approaches to a problem space is often key to such R&D efforts, and while there is a sense in which current AI tools are good at digesting context and offering adjacent solutions, it’s not clear that genuine novelty in implementing solutions is there yet.

Moreover, it’s also definitely not clear that for use-cases that are sensitive or complex, or can have serious negative consequences, the average user is willing to take on risk as a result of AI mistakes. They certainly wouldn’t tolerate a developer making mistakes that affect them, so it seems a reasonable expectation when considering either fully AI-generated code, or even AI assisted code (although, let’s face it, that’s just ‘code’ today). Even in non-regulated markets, the PR backlash from a project failure due to incorrect oversight is likely to be terminal.

Smarter people than me have written at length about project guardrails, governance, and humans in the loop better than I can, so I won’t regurgitate that here.

Hard Problems Are Business Edge

What all this means is that I have a strong feeling that there’s a class of hard problems in computer science, or indeed using computing to solve problems, that remain (a) completely, or relatively unexplored, and (b) are a bad fit for AI tooling due to novelty, user trust, or regulation. This is good news if you enjoy programming, and it’s great news if you want to start a business.

In The Millionaire Fastlane, MJ DeMarco offers five commandments to follow when starting a business, and two are particularly relevant here:4

1. The commandment of control—be in control of your business, pricing and operations (an ironic one considering this blog is hosted on Substack, but it’s not our, or my, core business)

2. The commandment of entry—the lower the barrier to entry to a market, the higher the competition will be, and the lower the margins. Try to find a market with a higher barrier to entry (or, implicitly, find a market where you have the ability to either create the market or create the barrier to entry in the market)5

The barrier to entry on the easy problems, with AI, is now effectively zero. It’s a double-edged sword for those chasing them too—by addressing them using AI tooling or commodity solutions, a would-be business owner is also violating the commandment of control, as they have no edge or USP.

Chase hard problems, and you will have an edge over everybody else that is using the powerful new tooling that’s available to chase the easy problems. After all, the tooling will also help you lower your time to market and iterate on hard problems too. It just won’t help by the same proportional amount versus project size, even if it helps the same absolute amount.

Again—if you have a unique idea (bonus points for it obeying the commandments above) that is easy and you can bootstrap your product to market now by yourself where you couldn’t have in the past, then go and do it now!

If you have any other kind of stake in tech and/or have an entrepreneurial mindset, you need to be chasing only one type of business idea—things that are (a) hard for AI to do (currently), and (b) have a high barrier to entry (either because they’re hard, because of regulation, or because of required specialist knowledge, et cetera).

Of course, for all I know, the tech might leap forward in a few weeks with a new release and render this post fully obsolete, in which case I will probably give up on this programming lark and become a carpenter. Time will tell.

There’s a by-election in my constituency and it looks like it will be two-way between the Greens and Reform. Suddenly I have a lot of skin in the game, much sooner than I had anticipated. I’ve also realised that on some level it’s patriotism (cringe) that makes me want to defeat Reform so much.

The Battle of Britain was S-Tier Antifascist Action

I’ve been interested for some time in a patriotic left-wing movement against the Far Right. I’m unashamedly left-wing, except for a slight hawkish streak on defence—which perhaps the current state of things has helped to validate—but grew up, like many people in England, on stories of the blitz, the Battle of Britain and films like A Bridge Too Far, 633 Squadron,1 and Where Eagles Dare.2

Now, perhaps it took me too long to have this thought, but my entire adult life has been taken up by the right claiming exclusive rights to the imagery and legacy of British involvement in World War Two. Last time I checked, World War Two was the greatest struggle in history against the extreme right, so at what point did it become weaponized by those that want to foist authoritarianism upon these isles?3

Though I’m no nationalist, and I’m a bit shy of ‘devoted’ to my country, I’m definitely a pragmatic patriot. I’m proud to be English, British, European—in spite of the bad parts—but feel that admitting this is somehow problematic due to the behaviour of many of the most performative, far-right “patriots.” That’s a bit weird, isn’t it?4

It’s in this context that I genuinely believe that we have a patriotic duty to defend England against Reform. Against Trump. Against Putin.5 I’m being deadly serious.

Identity and the Social Contract

There’s a chance this thing doesn’t get too out of hand and it remains a spiritual struggle, confined to the ballot-box, but that requires acting together and acting for the good of the country.6 You wanted blitz spirit? That’s what I’m selling. Let’s chuck the chancers out and rebuild a country of agency, opportunity and decency, without having to go through the bloodshed required to form a social consensus last time around.

The question is, what’s the smallest thing we could put to paper and all agree on? That we want to build a better country? That we want to build a better tomorrow? Less? Just that we want to build anything that is real? We need to find that shared hope. That’s the kernel. Then we can ask the question: where can we go from there? It’s definitely a huge task.

If I had to guess, I think Maggie Holland might have an answer in her song A Place Called England.7 At the moment, the country is “Run by men who think that England’s just a place to park their car.”

I’m taking a stab in the dark, but I think that’s a statement that both disaffected Reform and Green voters could agree on. I’m going to self-indulgently quote two whole verses of the song:8

And come all you at home with freedom, whatever the land that gave you birth
There’s room for you, both root and branch, as long as you love English earth
Room for vole and room for orchid, room for all to grow and thrive
Just less room for the rich landowner, he can stay in the Virgin Isles

For England is not flag or empire, it’s not money, it’s not blood
It’s limestone gorge and granite fell, it’s Wealden clay and Severn mud
Blackbirds singing from the May-tree, lark ascending through the scales
Robin watching from your spade and English earth beneath your nails

So here’s two cheers for a place called England, sore abused but not yet dead

We need ideas and action, not the neoliberal sound-bites and stasis of Labour and the Conservatives, and not the reactionary, millionaire-funded, self-interested dross of Reform.9 In fact, even the Left doesn’t really have any answers for this one, because it’s not about the political spectrum, and I’m not even sure it’s about a specific policy,10 it’s about who we are when we are together. England. Britain.

The flip of this lack of narrative control over patriotism by the left is the state of the Conservatives—somewhere I’ve drafted a post about the social contract in Jack Aubrey’s navy, which is really I suppose about one-nation conservatism, or old-fashioned patriarchal Tories. Maybe you can’t have a coherent answer to the far-right unless you have a coherent right-wing.11

Both the left and the right have to rebuild the social contract, and give people the space to decide who we are together when we face challenges together. There’s no shortage of challenges, even if there is a shortage of political will to face up to them.

Returning to the most immediate challenge, if the current right is Reform, then after the bizarre penguin episode in London12 I think we have to start asking the question, not only whether they are just dangerous,13 but whether they are in touch with reality at all.14

Regardless of how this all pans out, I think at our best, we’re braver, bolder, and more resourceful than the political class that fear making real decisions about the future believe we are. I also think we’re more human, more prepared to sacrifice, and more capable of cooperation than the cynical opportunists allow for either. Most people are just waiting for an opportunity to show it.

I’m a bit less constructive by nature than the Young’uns—too many old war movies, I guess—so I’ll end with a snippet of a rousing song:

Our ship being cleared, the foe we neared, with expectations high
That we should show the murd'rous foe
That British courage still will flow
To make them strikе or die!

—The First of June

1

The ending of this film (image above) always reminded me of this moment from Blackadder. I think it’s telling that the arch neoliberals like Michael Gove seem to hate it.

Nate Silver’s 2024 book On The Edge is the best non-fiction book I’ve read this year.1 Obviously I’d recommend giving it a read. It’s a great study of risk-taking behaviour, gambling and technology companies, and having worked in and around blockchains for five years, most of it rang uncomfortably true.

There’s one aspect of it that I think is worth responding to, however, and that’s the fact that narcissism, psychopathy and the Dark Triad personality type aren’t mentioned in the text, other than in one passing quote.

Silver builds a case for two communities, Riverians (risk takers) and Villagers (risk avoiders).2 Throughout the book, there’s attempts to pin down what these personalities might have in common—particuarly on the Riverian side—and ultimately despite a well-written section on neurodivergence and asperger’s, he doesn’t draw any firm conclusions, or examine other conditions.3

“narcissist” or “Narcissist”?

Part way through the chapter Acceleration, after the nominal half-way point of the book, Silver quotes veteran tech journalist Kara Swisher:

I asked Swisher why tech leaders like Thiel and Musk are so obsessed with their media coverage. She didn’t need much time to consider her answer. “It’s because they’re narcissists. They’re all malignant narcissists,” she said.

Similarly to Silver, I’m not going to make any hard and fast conclusions. That’s because as is (already) common for my blog posts, I’m a little out of my lane. Even so, I do wonder if she’s actually being quite specific with the use of medical jargon here, because it tranforms the meaning of the quote.

In everyday use, “narcissist” means “self-centred”; but it has a concrete definition, which we will explore below. Moreover, Narcissism forms part of a triad of behaviours—along with Machiavellianism and Psychopathy4—that together form what is known as the Dark Triad (Paulhus and Williams, 2002).

There’s quite a bit of literature out there on the Dark Triad, and some debate about a central core behaviour, currently hypothesized as aggression.5

The Dark Triad, Narcissism and Psychopathy

So what is the Dark Triad? I’m going to be as concise as possible, quoting the original Paulhus and Williams paper,6 before quoting some definitions from a later paper (since those authors incorporate work post-Paulhus and Williams) on the constituent parts.

Here’s Paulhus and Williams on the Dark Triad:

Despite their diverse origins, the personalities composing this ‘Dark Triad’ share a number of features. To varying degrees, all three entail a socially malevolent character with behavior tendencies toward self-promotion, emotional coldness, duplicity, and aggressiveness. In the clinical literature, the links among the triad have been noted for some time (e.g., Hart & Hare, 1998). The recent development of non-clinical measures of all three constructs has permitted the evaluation of empirical associations in normal populations. As a result, there is now empirical evidence for the overlap of (a) Machiavellianism with psychopathy (Fehr, Samsom, & Paulhus, 1992; McHoskey, Worzel, & Szyarto, 1998), (b) narcissism with psychopathy (Gustafson & Ritzer, 1995), and (c) Machiavellianism with narcissism (McHoskey, 1995). Given such associations, the possibility arises that, in normal samples, the Dark Triad of constructs may be equivalent.

Although they also describe Machiavellianism, Psychopathy and Narcissism (at the time of writing, non-clinical psychopathy was the newest of the three facets to be recognised; “SRP scores predict anti-social behavior in forensic and non-forensic populations”),7 A longer description of each is helpful.

Walker et al. (2022) in a paper on faking good and bad on tests have several useful definitions. Importantly, they make it clear that the three constituent behaviours are not equivalent. First, the triad:

The ‘dark triad’ is an umbrella term for a set of three socially aversive personality traits comprised of narcissism, Machiavellianism, and psychopathy (Paulhus & Williams, 2002). While all three traits are associated with ethical, moral, and socially deviant behavior, among other shared characteristics, they are considered independent of each other. Recent debate relating to a shared common core among the dark triad traits continues, but there is some consensus on the role of antagonism connecting narcissism, Machiavellianism, and psychopathy (Jones & Figueredo, 2013; Jones & Neria, 2015; Paulhus & Williams, 2002). Furthermore, each of the three traits shares exploitative characteristics with goal-focused manipulation of others’ emotions to get what they want.

Then, Narcissism:

While narcissism has traditionally been conceptualized as a personality disorder, mild to extreme sub-clinical characteristics of narcissism are found in non-clinical populations (Raskin & Hall, 1979; Samuel & Widiger, 2008). Accordingly, narcissism is commonly regarded and investigated as a personality trait (Miller & Campbell, 2008) and comprises two facets: grandiose and vulnerable narcissism. As the name suggests, individuals high in grandiose (or “overt”) narcissism are characterized by grandiosity, self-confidence, and exploitation of others and tend to rely on self-validation to maintain self-esteem. When threatened, individuals high in grandiose narcissism tend to blame and devalue others while denying their own weaknesses (Dickinson & Pincus, 2003; Zhang, Luo, Zhao, Zhang, & Wang, 2017).

Vulnerable (or ‘covert’) narcissism tends to rely on external validation. While grandiose fantasies also characterize vulnerable narcissism, people high on vulnerable narcissism tend to oscillate between self-love and self-loathing, thereby exhibiting fragile self-esteem, defensiveness, and resentment (Weiss, Campbell, Lynam, & Miller, 2019). Individuals high in vulnerable narcissism are characteristically hypersensitive to negative feedback and tend to act aggressively when their sense of self is threatened (Wink, 1991). Though grandiose and vulnerable narcissism share several core characteristics such as an antagonistic interpersonal style, self-importance, entitlement, and hypersensitivity to criticism (Dickinson & Pincus, 2003; Krizan & Herlache, 2018; Weiss et al., 2019), there is clear evidence that these two facets are distinct (Krizan & Herlache, 2018; Miller, Vize, Crowe, & Lynam, 2019; Walker et al., 2021).

Does ‘vulnerable’ narcissism sound at all like a billionaire angry at negative media coverage? I won’t over-egg it. Their discussion of Machiavellianism is more terse:

Machiavellianism is commonly understood to be a unidimensional personality construct characterized by goal-focused manipulative and callous social interactions (Christie & Geis, 1970). Individuals high in Machiavellianism tend to be viewed as strategic, capable of delaying gratification for bigger and better rewards in the future, possessing low moral commitment, and engaging in long-term strategic planning with a cold and cynical world-view (Christie & Geis, 1970; Furnham, Richards, & Paulhus, 2013).

Finally, psychopathy is described:

Psychopathy has been typified by interpersonal, affective, and behavioral characteristics, including superficial charm, pathological lying, and lack of empathy, conscience, and remorse (Cleckley, 1951; Hare, 2003). Although clinically identified with antisocial personality disorder, the sub-clinical characteristics of psychopathy, like narcissism, exist on a continuum in the wider population (Berg et al., 2013). Based on Karpman’s (1941) work, enduring classifications of psychopathy make a distinction between two subtypes differing in their etiology and symptomology. Expanding on earlier models, Hare (2003) made a distinction between primary and secondary psychopathy. Both psychopathy facets involve affective elements suggesting some indifference to their own and others’ emotions, each underpinned by an antagonistic interpersonal style (Miller & Lynam, 2012).

Primary psychopathy is characterized by: (a) a lack of guilt and remorse, with elevated levels of callousness, manipulation, and socially desirable behavior (Hare, 2003), (b) having superficial affect (Casey, Rogers, Burns, & Yiend, 2013), and (c) deficits in affective empathy (Wai & Tiliopoulos, 2012). Primary psychopathy is also associated with lower levels of fear (Patrick, Bradley, & Lang, 1993) and lower indications of repentance (Hare, 2003; Lee & Salekin, 2010).

Secondary psychopathy is characterized by higher levels of antisocial behaviors such as aggressiveness, impulsivity, and anxiety (Levenson, Kiehl, & Fitzpatrick, 1995). These types of characteristics are potentially a result of experiencing strong emotions which are unable to be effectively regulated (Hare, 2003). Additionally, individuals high on secondary psychopathy have been shown to possess guilt and fear responses not typically observed in individuals high in primary psychopathy (Lykken, 1995; Wallace, Malterer, & Newman, 2009).

Note the difference between Primary and Secondary psychopathy, and bear that in mind when we talk about Elon Musk later.

So as to not throw out the baby with the bathwater, it’s worth saying that in the original article, the authors don’t advance a strong opinion on how destructive the individual traits might be:

Indeed, Machiavellians and narcissists may be more of an interpersonal irritant than a threat: Data suggest that such characters are a mixed blessing in personal life (Robins & Beer, 2001), interpersonal life (Paulhus, 1998), and some organizational contexts (Hogan, Raskin, & Fazzini, 1990; Robins & Paulhus, 2001). Adaptive interpersonal correlates of subclinical psychopathy may be more difficult to find. Their positive self-view and lack of anxiety, however, can be viewed as adaptive in an intrapsychic sense (Taylor & Armor, 1996).

Indeed, Walker et al.’s “faking good” and “faking bad” paper (using exercises such as a simulated job interview) built on work on the Big Five personality traits which showed people answer according to how they think makes them appear best—meaning such deceit is not the preserve of disordered personalities:8

People are highly motivated to be viewed in a favorable light by others. There is evidence that people distort their responses on personality scales under high-stakes conditions (Sjöberg, 2015). In fact, meta-analytic findings show people can and do substantially change their personality scores when motivated to do so (Birkeland, Manson, Kisamore, Brannick, & Smith, 2006; Viswesvaran & Ones, 1999). Much of this research has focused on the broad personality domains of the Big Five/Five-factor model (extraversion, agreeableness, conscientiousness, neuroticism, and openness; MacCann, 2013).

Elon Musk and SBF

Of particular interest in Silver’s book is the case where he talks about Elon Musk and Sam Bankman-Fried (the first being one of the “malignant narcissists” mentioned before). In the first case, Musk is portrayed as being without fear, playing poker according to what is referred to as the “maniac” strategy—go all in, every hand—while SBF is portrayed as a utilitarian prepared to extend potentially flawed logic to a degree that frankly beggars belief in the telling.9

Indeed, Musk’s fearlessness extends beyond business to his physical well-being. His first appearance in Silver’s book is an account from Thiel about how—ostensibly as a a result of both impulsiveness and fearlessness—he crashed his MacLaren F1, and amazingly was unscathed. As Silver relates, “It’s not that Musk had made some rational calclation—that if you’re worth $22 million, you can just afford to buy a new one. Rather Thiel said, Musk hadn’t bothered to consider the possibility of a crash.”

The question here is, had Musk not considered the possiblility, or was he consitutionally unable to?

Crucially, besides their impulsiveness and fearlessness, both SBF and Musk seem to not really understand the implications of the prisoner’s dilemma—i.e. that many games are not zero sum.10 These things combined are a heady cocktail, when we’re talking about personalities that have Bond-villain levels of power.

Though they may not share every behaviour, the inability to see the world as anything but zero-sum is something they share with Donald Trump. Yes, it is concerning that there’s a Dark Triad narcissist who sees the world in fully zero-sum terms with their finger on the nuclear button, but what can you do?

To highlight how insane the zero-sum worldview is, even the most meme-level (and arguably cringe) business resources tend to have the need for collaboration built-in.11 One of the 7 Habits of Highly Effective People is to seek “win-win” situations, so it’s not as if this insight isn’t an informal Schelling point as well as one more formally described by (among other things) the logic of mutually assured destruction.

Even if you flip this argument on its head and still look to win in a zero-sum (or close to zero sum game), it’s likely that it is a more effective strategy to not think of the game as zero sum, or even if not seeking a compromise or asummetrical compromise in your favour, to at the very least put yourself in the shoes of your opponents.

Nate Silver quotes American general HR McMaster pointing out the need for strategic empathy, not for the compromise case, but to defeat your opponents, by understanding their success and exit conditions.

Other than on an instinctual level, the narcissist—as opposed to the psychopath—is perhaps less able to do this in order to act in their own self-interest. I’m on shakier ground here in terms of my reading, so I’ll just give an example and then wrap up.

To me, a good example of how these personalities might interact would be (with an armchair diagnosis, I know) in a recent episode of the UK edition of the Traitors, where the contestant Fiona (a Traitor, in the game) completely sabotaged her own game as the result of having her ego bruised, all the while proudly saying to the camera that she was executing a master plan (see Walker et al., quoted previously, “[w]hen threatened, individuals high in grandiose narcissism tend to blame and devalue others while denying their own weaknesses.”).

Her opponent (who was also a Traitor, and almost certainly also a disordered personality along the Machiavellian or Psychopath axes), calmly briefed against her and got her voted off the show while receiving no votes in return.12 The situation has more than a little of the Musk/Trump spat about it; no doubt if Fiona could have fired her opponent via tweet, she would have.

Just another day in the Corps

I’ve mentioned the books of Richard K Morgan in the footnotes, so indulge me while I use them to conclude my argument.

In the Takeshi Kovacs novels (as opposed to the Altered Carbon TV series)13 it would be hard to describe the Envoys as anything other than psychopaths—and in fact, not even subclinical ones, since the first page the Envoy Corps is mentioned, there’s a time jump to Kovacs watching a fellow member of the Corps being incarcerated on a future meeting, “She was going down for eighty to a century; excessively armed robbery and organic damage.”

To some extent the ability of Envoys (psychopaths) to use their lack of anxiety and remorse to reshape the world around them mirrors the main themes of Morgan’s other novels Market Forces and Thirteen—the latter a very explicit study of psychopathy. Perhaps Silver’s book is similarly a story of psychopaths, not risk takers, being able to reshape a world that can no longer govern their behaviour.

Schwarzkopf? A lightweight. Montgomery? Overrated. Patton? A blowhard. Napoleon? A loser. Wellington? Just lucky. Nelson? Didn’t even survive his own battles. Sun Tzu? Probably ghostwritten.

No. I am speaking of course of Denethor, Son of Ecthelion, Steward of Gondor. Clearly in Number 10, they’ve gotten out the popcorn and have had a rewatch of Peter Jackson’s Lord of the Rings Trilogy.1

From this they have decided to follow Denethor’s cunning plan2 (I will leave the reader to figure out the metaphor here):

1. Retake Osgiliath, a ruined city of little strategic value

2. Double down on defending this site of little strategic value while getting outflanked

3. Get outflanked and lose site of little strategic value with heavy losses

4. Double down again on retaking the site of little strategic value, resulting in your son being mortally wounded and the loss of a significant portion of your forces

5. Get attacked in your heartlands by an emboldened and unchecked enemy

6. Have your capital city sacked by rampaging Orcs

7. Try to set son on fire rather than accept medical help, even though you have no relevant expertise

8. Set self on fire

9. Run through the streets on fire

10. Plunge from the prow of Minas Tirith

11. Die

Before continuing, I will say that I have had two completely opposite responses to the main arguments in this post—that stables on public blockchains are trivially exploitable, either by chain halt or more esoteric attack.

First, I put this argument to my PhD supervisor and got a “so what?”1

Second, I put a (much shorter) version as a question to Brunello Rosa, at a seminar on the topics in his 2024 book Smart Money. He said, “I don’t think the Americans are taking that risk seriously.” Whether or not they should be taken seriously is thus at the discretion of the reader.

So, caveat emptor on this post.

CBDCs may still be a long way off for most countries, but stablecoins are all the rage. After a solid year of stablecoin hype and maximalism, there’s no sign of the wave cresting, so it’s time to consider what the infrastructure for this new era of payments might look like, at a time when regulation is struggling to keep up, and institutions are weaker than ever.

In particular, the Americans have gone all-in on a strategy of using private companies to roll out stablecoins. Now, this isn’t necessarily a bad thing in terms of getting to market quickly,2 but most of these projects are also relying on public, permissionless blockchains, and that is a problem.

I think this strategy fundamentally understates the risk of disruption to any of that infrastructure.3 Yes, even a total failure will only affect retail consumers (and possibly some small businesses), perhaps the effect would be negligible in the end. And hey, I might be wrong—what do I know? But this is my blog, and I’ll cry if I want to post essay-length musings if I want to.

I’m going to describe how some basic attacks work, and why an attack might be worthwhile, and conclude with some thoughts on decentralization, and whether locality and obliviousness are more useful design goals.

I believe the kind of ‘decentralization’ that most existing blockchain designs offer is likely not proof against the kind of problems a digital currency at scale might face.

How to Attack a Blockchain 101

The most basic kinds of attack we need concern ourselves with are buying, or taking over, the finality of a chain (at a threshold of say, 51% for Bitcoin, or 66.7% for a Tendermint or BFT-based network like Cosmos), and halting a chain.

For some networks, both the threshold that a chain could be halted by taking voting power or hash power offline is the same threshold (what the Nakamoto Coefficient terms the operative threshold).

However, for others, such as the Tendermint-based blockchains in the Cosmos ecosystem, the takeover threshold is 66.7% of voting power, while the operative threshold is 33.4%, the point at which a chain can be halted.

For most networks, simply halting the network is enough to do damage. Finding a zero-day that allows for double-spends or performing an esoteric attack on its cryptography is likely not necessary. In fact, in both the cases I’ve experienced of cyber attacks on blockchains, zero-days were simply used to inject non-determinism into the application state, thus halting the chain when nodes reported different results for a piece of computation.

Since Satoshi’s whitepaper it hasn’t exactly been a secret that the risk of attack is major and only incentives and agent independence keep the party going. Satoshi relied on the rational behaviour of miners to avoid undermining their holdings (assuming that they have what Budish et al. (2024) would call “stock value” in the asset, or what you might call skin in the game).

Budish et al.’s argument has the implication that any attack which is ‘cheap’ to execute4 might be free if the attacker has no ‘stock value’, or token holding. The fact that most penalization mechanisms for adversarial actions—be it a rewrite or double spend—are ex-post only makes this harder to reason about.

Houy (2014) argued that a Proof-of-Stake cryptocurrency was even more vulnerable, and that any economic attack would be free for a different reason—that if agents plausibly believe an attack will succeed, they must assume the collapse of the cryptoasset, and sell their tokens to the attacker to avoid ruin.

Budish’s other 2024 paper showed that the cost of securing Bitcoin scales linearly—making it impossible to categorically secure at scale in the face of a large enough attacker, and raising questions of every chain with a different security model or a lower market cap.

Luckily for stablecoins, the easiest and cheapest attacks are on cryptoassets that are endogenous tokens—without external backing or security. Thus the risk is less undermining confidence and the token going to zero as it is disruption or destruction of the underlying ledger.

However, this is an important point—if you launch on a blockchain like Ethereum, you are taking on the technical risk of that ledger and the economic risk of its endogenous token. Just because your externally-backed stable isn’t5 vulnerable to going to zero due to a social attack doesn’t mean you’re safe from them.

What if Pearl Harbor, except Shitcoins

So my question: what happens if (a) these systems get adopted at scale, and (b) a sufficiently motivated adversary decides to attack them? I think it would make the Terra/LUNA fiasco look like a footnote.

Perhaps a nation state wants to seed chaos before an invasion of a neighbour, or simply to grief and cause economic damage via disruption (although you would probably keep this in your back pocket for a special military operation occasion; once the stunt has been pulled, it’s unlikely to work twice).

Okay, this doomsday scenario sounds somewhat possible when stated with enough confidence. However, as my long-suffering supervisor pointed out, “so what?”

Whether or not there is a second-order effect, the most likely outcome would be a sort of minor chaos, unless the contagion spread to wider markets, and it’s not clear that the outcome would be a beneficial chaos to the attacker. Or, to paraphrase the argument sent my way:

So you’ve gone into an auditorium with a BB gun, and you’ve shot somebody. Now they’ve got a slightly painful arm, and everybody knows you’re a Bad Guy. Is that really worth doing?

Facetious as ever, I shot back with something like:

Well, if you go in with a black-sprayed BB gun6 and shoot the guy on stage, maybe there’s security, maybe they shoot somebody else while shooting at you, and then the chaos gets really out of hand.

Alright. It’s probably not the best example. Getting shot for pulling a griefing stunt is not exactly a dominant strategy. Still, if you’re in a position where you’re already set on being a “bad guy”—perhaps tomorrow morning you will launch an invasion of Ukraine—and you’re reasonably confident that any ill effects will be outweighed by your expected value on success, then it could be worth it.7

The Dark Forest

Now, like the true charlatan that I am, I can reveal my true purpose. Just like one of those thought-leaders posting about how AI is going to take jobs (and save businesses zillions) that turns out to be incubating an AI startup, I have a vested interest in oblivious systems. Not private systems, or even systems with privacy-by-design, but systems with obliviousness-by-design.

Decentralization is in the title of this post, but it’s a tricky concept to nail down. The Nakamoto Coefficient mentioned previously is sometimes used for it, and there’s various industry and academic definitions for it. I’m the lead author on a paper which suggests a new model for it. Shilling our paper isn’t the goal here, however, but to question what the desired outcome is of decentralization.

If the goal is systemic robustness, then I would argue that locality is perhaps a more relevant concern,8 as you reduce your systemic risk in return for different trade-offs. Locality (or local centrality) is closer in shape, I would argue, to the kind of ‘decentralized’ system that was discussed in Paul Baran’s original 1964 paper that defined ‘decentralization.’

If the goal is agency, then perhaps obliviousness is a more relevant issue,9 as other agents cannot track, monitor, control or trivially attack you. Pseudonomity is not good enough. This indeed is part of the reason Kahn et al. (2005) make privacy a requirement of any definition of ‘money’, something many stablecoins lack.10

This is especially true of payment systems, where any payment system necessarily requires some means of identification. In the examples of attack before, you’re exposed to not just every stakeholder or user in the system, but any agent, anywhere that can see your history, or even that can infer you exist.11

Stable, account-style identifiers make accounts trivially targetable, while the kinds of network-level attack described previously can be mitigated by a combination of locality and obliviousness. A global shared ledger implies a single point of failure, a fact that is trivial to intuit, but one which has far-reaching consequences for both governance and counterparty risk.

For science fiction fans, the obliviousness point is a little bit like in the novel The Dark Forest, by Liu Cixin. In the novel, the solution to the Fermi Paradox is that civilisations attempt to go unnoticed by the wider galaxy. To announce your existence to a hostile universe results in the destruction of your civilisation.

Revealing your crypto holdings12 isn’t quite the same level of existential disaster, but it’s a little like getting doxxed—that genie is hard to get back in the bottle.

Indeed, decentralization can also be thought of as your exposure to risk—at least in terms of the operative subsystems or governance mechanisms that guarantee integrity in the system. I’m simplifying a little bit, but performative decentralization, or ‘decentralization theatre’ does not move the needle on the topology of a system when described this way.13

Additionally, as touched on above, systems may have global and local centrality, as we described in our paper for the Sark architecture. This is a crucial distinction, when considering ledgers with global shared state—they necessarily have a single point of failure that is system-wide.

Systems with local centrality may be compromised at the local level, but the contagion is not system-wide. Again, in some ways, this is closer to Paul Baran’s original conception of the term ‘decentralization’.

Regardless of how you define it, the reality is most existing public, permissionless blockchains are not very decentralized. Taking a widely-used industry metric, the Nakamoto Coefficient, we see that for many chains it is often very low in terms of key metrics, such as validator count.

Elsewhere, taking an academic heuristic, the Edinburgh Decentralization Index MDT,14 many chains due to Foundation delegations or stake concentration violate the MDT.15

Conclusion (or, maybe we should just use PostgreSQL)

The end result is that either chains need to decentralize further, and not merely in a performative way, or accept other trade-offs, such as local centrality, in exchange for greater systemic robustness.

If you launch a stablecoin on a ledger, obviously you take on the risks of that ledger—unless of course, the use of the ledger itself is performative, and behind the scenes your actual system of record is a normal database duplicating the data, like PostgreSQL.

If your system of record isn’t the blockchain, then using one at all is largely pointless.16

In our work on Sark, we seek to have a more disintermediated, user-custodial system with fungible assets while still effectively envisioning it being run in a permissioned or consortium way at the node level.

As the users and implementers of stablecoins on systems that include a ledger, we have to evaluate which properties we care about, and carefully weigh up performative security versus actual security.

For most users, a more centralized system that is honest about the topography of its centralization might in the end be a better fit than a more decentralized system that in practice is largely perfromative. There’s a reason regulated banks are the way most people interact with the financial system—how much risk users are comfortable with.17

Obviously, I have pretty strong opinions and experience in this area; our first Needlecast blockchain node came online almost exactly five years ago. I don’t know what the exact definition of crypto survivor is, but I’ve got a reasonable claim to having weathered some pretty big busts, hacks and insane situations.1

In an old episode of the validator podcast we periodically did with Lavender 5, Rhinostake and Kingnodes2 we hypothesised, probably around 2022, that the validator business would gradually become a margins game. I’m pretty sure we even presented a Wardley map to show how the validator business was maturing in the direction of being a commodity—if you like Wardley maps, then stick around.3

Since then, the dominant narrative has been one of consolidation; of mergers and acquisitions among the larger independent players, bankruptcy for those that overreached in various bull runs, and a long tail of smaller independent or hobbyist validators.

We’ve sat uneasily in this narrative, because of course we are not one of the larger independents (although some of our close colleagues such as L5 or Polkachu might sit in that category), but have also at some times had something under the order of half a billion USD in Assets Under Stake (AUS). Still, it has felt for a while like our time is limited, as a more boutique operation in a higher-cost jurisdiction.

This in itself isn’t that big a deal. In the naïve view, validators should be fungible, and perhaps it should be a margins game (whether we’re talking about PoS or DPoS4—Ethereum has been a margins game for some time, even with MEV factored in).5 However, the reality is that if you care about a metric like geographic decentralization, or indeed datacentre or infrastructure decentralization (and blockchain Foundations seem to), then validators are not fungible. Not only that, but their running costs are widely variable depending on location and juristiction.

For example, as mentioned before, if you’re in a high-cost jurisdiction, not only are your baseline costs (salaries, utilities, fees, legal, accountancy, business support, et cetera) higher, but there’s also typically more of a regulatory presence. In the UK this has always loomed, albeit somewhat inconclusively. Luckily for us HMRC published a comprehensive tax manual for businesses quite early on, so compliance on that front hasn’t been too big an issue,6 but elsewhere we have still had problems.

Chief among these is debanking. Although we’ve not run a regulated operation7 and any stakers custody their own funds,8 working in this space paints a risk target on your back that results in periodic interruptions to business operations. We’ve had to at various times run multiple different bank accounts, and eventually we fully split the business operationally to address this risk.

The end result is more cost, not only in accounting and admin, but also if you need, for example, banking providers that will deal with high-risk clients. Such arrangements (these typically require a base deposit level of $25k or more) are expensive and also effectively lock up operational cash. These considerations are all part of the service that implicitly our clients and counterparties receive, though they don’t pay for it, per se.

In fact, typically none of this is factored in by counterparties. In the two Wardley maps below,9 I’ve taken a rough approach at sketching out the value chain for validation. For consumer-level stakers, they often don’t care where a validator is based, or what their setup is. They just care about uptime and cost. This means they see the operational expertise as slightly less commodified, yet there’s not much custom beyond a brand.10 Most validators have some degree of custom software, even if it’s just for monitoring, that’s not something stakers will typically pay for.11

As an independent node operator in this space, you’re competing with major CEXs, some of which are regulated, with a huge brand presence. Given that performative aspects of blockchains tend to get the most attention (not that this is different in other industries, to be fair) marketing and brand has an outsize impact, and it’s no surprise as a result that few independents can compete in the consumer staker market over the long run.

You could draw this map differently based on what stakers perceive, or want, from the value chain, versus what it actually looks like. From the perspective of the operator, it looks like the Foundation Wardley map, below, in both cases, but only in the Foundation case are the custom elements typically compensated at all.12

Foundations, unlike stakers, typically care intensely about certain aspects of operations, going out of their way (in most cases) to KYC and vet their counterparties and the companies they delegate to, while (often) also specifying desired jurisdictions and node locations.

This is a highly custom service, and yet often it does not come with the recognition of it as such. It’s not exactly clear why; maybe validators are all bad at negotiating, or maybe it will take a sustained bear market cull to re-establish the perception of what is, effectively a subtype of institutional validation as an appropriately expensive service, rather than the commodity service it is in the case of DPoS.

As you can see, it’s hard to represent by abstractly placing ‘validation as a service’ on a Wardley map,13 and I suspect, regulators have little sympathy.14 Regulators probably assume all validators are raking in cash and are mainly focussed on the big operators.

As we’ve seen just in the last weeks, future regulation in Korea has likely been part of the reason that a large validator, a41, has decided to exit—and they didn’t exactly seem to be struggling. As validators, and particuarly as independent validators, we need to become better at articulating what exactly it is that clients, stakers, or counterparties are actually paying for. In some cases, we still need to make the case that they need to pay for these things at all.

Still, this all may represent an opportunity. With validators exiting both via merger and bankruptcy, market stresses or regulation, it might be that there’s more opportunities for small, leaner or mid-tier validators to pick up market share. Whether or not that will pan out into a more predictable or scalable business model in the long-run remains to be seen.