10 assumptions I make daily

Last revised 12 October 2026.

You can’t argue for all of your beliefs and get anything else done, so some things are simply assumed, for better or for worse. You’re entitled to know where I sit before weighing what I say.

1. TEI is preferable to AGI, which is preferable to ASI.

Task-emulated intelligence is just enough artificial intelligence to power automation and limited autonomy, and not enough of either to be much more threatening than the software we already run. It’s the sweet spot of current capability, and we’ve already reached it, more or less. It’s enough for the work, and the steps after it create risk without creating much of anything the necessary work needs.

Everything Pentad Labs builds assumes an agent whose intelligence is emulated for the task at hand rather than general or “super”, whatever that means.

2. WunderOS assumes an infinite scratch disk for storage

S3 is one of the ten greatest feats of software engineering, and most new databases, perhaps nearly all of them, should be built on it or on something like it. WunderOS’s database, PlatypusDB, uses node-local NVMe as a hot cache and S3 Express as an effectively infinite scratch disk.

3. The UX era is over; the era of AX and OX is here.

UX isn’t over, but it’s going to rest for a while. What takes its place at the front of the queue is agent experience, the interface an agent meets, and operator experience, the particular subset of UX that people need in order to manage agents and fleets of agents.

The second is Lisanne Bainbridge’s revenge, that is, the return of the irony of automation she described in 1983. Automating a task doesn’t remove the human; it moves them to supervising a system they no longer practice operating, and the supervisor needs tools no one has built.

The key distinction to internalize for AX and OX is the difference between automation and autonomy.

4. The AI era is very early, and no one knows what will happen.

Karpathy observed recently that two groups are talking past each other. Most people formed their view of AI from a free tier and a model since deprecated, and the quirks they laugh at are real. This is Team Free Tier.

A much smaller group pays for access to frontier models and uses them professionally in programming, mathematics and research, where the gains this year have been staggering, and that group assigns the capability a gravity the first can’t see. This is Team SWE.

Both groups are right, because the capability is peaky: it has climbed fastest where the reward is verifiable and the B2B value is highest, and least in search, writing and advice.

So it’s early, earlier than the discourse allows, and one thing that follows is that VC-led kingmaking is premature. The companies being anointed now are being anointed on the strength of a capability almost no one has used.

The kingmakers don’t have much choice about this, since capital has to be placed somewhere and on a schedule. So this is advice less to them than to those crowned early, who do have a choice about what to do with a crown awarded before anyone has seen the kingdom.

5. Alignment requires semantic judgment & mechanical authority. AI can supply only the first.

Alignment needs two things.

• The first is semantic judgment, the capacity to tell whether an action is the right one in context, and this will come from AI regulating AI or it won’t come at all.
• The second is mechanical authority, the capacity to decide whether an action is permitted and to stop it when it isn’t, and this has to be not-AI: good old-fashioned software, deterministic and inspectable, for reasons I argue at length elsewhere.

Hence, alignment and agentic engineering are both neurosymbolic programs.

6. Over time the dominance of LLMs in the neural zoo will diminish.

The neural zoo is the population of model types in production use, and the LLM is its dominant species for now. I don’t know which other types will surge. LeCunesque world models and fast System 1 classifiers are the obvious candidates.

The details matter less than the posture, which is that an operating system for agents can’t assume the only model type it hosts is, and will remain, an LLM.

7. Model Minimalism is correct and inevitable.

Use the cheapest mechanism that preserves the answer. I won’t say more here, since I’ve written about it at length.

8. The whole computing stack is being redesigned for agents, and the OS isn’t immune.

Agents will live on Linux, Windows and macOS for all sorts of reasons, and none of those is an operating system for agents, for reasons I’ve argued elsewhere. Both things are true at once.

I expect every WunderOS deployment to run on Linux, until or unless we work out how to operate BEAM/OTP at scale as a unikernel.

You will understand this assumption when you understand that it has nothing to do with Linux, Windows, or macOS.

9. Agents are the weirdest thing we’ve ever made.

I owe the world an elaboration of this claim, and I mean it literally. The phenomenology of agents is strange.

An agent is a program and a user at once, which no operating system before now has had to assume, and from one moment to the next it’s capable of insight and of stupidity in ways its average competence does nothing to predict.

The implications of this weirdness will take years to unravel.

10. AI has the potential to move society to a post-capitalist political economy.

I don’t assume this will happen, only that it might.

It’s unclear to me that wide-scale unemployment is consistent with a stable social order in the G7 countries, even with a welfare state and a basic income at Nordic levels. If AI produces that unemployment, then either the political economy changes or the social order doesn’t hold. Stability is everyone’s concern here, not one side’s: even those who would hasten the transition to a post-capitalist economy should want it to be a stable process. This sits beside the argument about where value concentrates without contradicting it.

That page describes the economy we have. This assumption is about what it might turn into if the social order is to hold.