Researcher & founder @ITJobsWatch
North West, England
Joined February 2012
- Tweets5.4K
- Following562
- Followers900
- Likes3.9K
John Grant retweeted
it’s actually worse than that! Once we have an atom-identical copy of the Mona Lisa, a SWAP operator that exchanges atoms for atoms is undetectable by the Hamiltonian of the universe, thus chain of custody provenance becomes a purely probabilistic statement about the likelihood that the identical paintings have not already spontaneously exchanged places.
the only way forward, then, must be that the asi _forbids_ atom for atom replicas of famous objects. 90% of world GDP will be devoted to these departments of counterfeit resilience. Within these DOCRs, we will
Interesting read. "What if automating AI R&D triggers an intelligence explosion?"
casp.ac/reports/intelligence…
Armin Ronacher is right that the Fable ban draws a new line. But state power and corporate power are not two problems. They are one. Concentration of capital and concentration of model control are the same concentration. While Anthropic is preparing for its IPO, the state decides who can use the models that financial capital will help build. The essay suggests cooperation as the way out. Cooperation is a form of competition. There is no evidence for collusion. But it is plausible that competitors have colluded with the government to ban the most powerful model just as that IPO nears. Collusion is a form of cooperation. If that is what happened, cooperation is not the way out. It is the weapon. Cooperation has different forms. Collusion concentrates power. Open science spreads it. So I suspect open science, rather than just open source, is what could shift the trajectory of AI. You cannot ban a recipe the world already knows how to cook.
When I struggle to structure my thoughts about what's happening I turn to writing. Today about the recent US Anthropic ban news, what it says about power and dependency, and what it should mean for Europeans and citizens of the world. It's a long one. lucumr.pocoo.org/2026/6/13/a…
It's been interesting to observe the human generalist vs specialist debate since generative AI emerged in Nov 2022. The assumption was that humans would have to choose. The question seems to have shifted to hardware.
Google's latest TPUs are two separate designs. One specialised for training. The other for inference. Google's announcement claims these TPUs are "co-designed for Gemini, open for everyone". The protocol commitment is hardwired first. The substrate is opened second.
AlphaChip helped to design three generations of TPU layouts. Curved and donut-shaped floorplans, beyond what any human designer would have risked. The artefacts work. The reasoning behind their layout is not transparent. Perspectival displacement at the substrate.
Ricursive Intelligence, founded by the AlphaChip team, is working on fabless to designless. Custom silicon without a design team. The customer becomes a workload specification. The chip is constituted in the coupling between the specification and the workload it serves. Neither party holds the design.
Each TPU block had a human owner. Selecting an AlphaChip layout meant accepting responsibility for an artefact that could not be accounted for. Designless extends this across the customer base.
Constitutive coordination is the frame. The capacity is in the coupling. What does the generalist vs specialist debate become when the coupling itself starts to specialise?
John Grant retweeted
Protocols organise coordination from formal standards to informal norms. The Protocol Field maps them against the gravity of protocol debt. Applying this diagnostic framework to human-AI coordination makes these emergent patterns legible.
itjobswatch.co.uk/research-n… #protocoldebt
John Grant retweeted
The Persistence of AI-Mediated Protocols. This essay defines strange protocols. The opaque, emergent patterns that arise from networks of AI agents operating at scales beyond human comprehension.
itjobswatch.co.uk/research-n…
I agree with Farrell & Shalizi that LLMs are coarse-grainings. The question I've been pursuing is what the coarse-graining is constituting. They point to Cronon's grain-grading scheme. They read it as a lossy categorisation with power effects. Read another way, the scheme is what brought the national grain market into being. Farmers, elevators, and buyers acquired capacities in that market they did not host before the scheme existed. The capacities were constituted in the coupling the scheme made possible.
The same applies to LLM embeddings. As best guesses about which video to surface, they sit inside the lossiness frame. As the constraint regime under which a coordination dynamic between platform, content, and user attention now runs, they do different work. The dynamic's capacities are constituted in the coupling.
Lossiness names a symptom. Constitutive coordination names the mechanism.
knightcolumbia.org/content/a…
This paper shows that putting an entire procedure into the system prompt outperforms LangGraph orchestration absorbing the previously logged routing calls into a single forward pass. The paper calls this self-orchestration and the older architecture scaffolding the models have now outgrown.
Self-orchestration has just removed the audit surface from procedural agent behaviour. Strange protocols cometh.
arxiv.org/abs/2604.27891
This recent paper reads somewhat like a manifesto and proposes learning mechanics as a scientific theory of deep learning. It covers training dynamics, hidden representations and final weights. But the framing is inward. It can explain what happens inside the model. It cannot explain what happens at the interface between the model and everything else. The interface is where the model meets the world it operates in. What matters operationally lives there. The human-AI coordination gap is where those relations meet practice. An inward science of deep learning is necessary, but not sufficient.
arxiv.org/abs/2604.21691
AlphaZero, AlphaFold, AlphaProof. A decade of AI whose reasoning is not human and whose outputs arrive at scales no community can read. Silver and Sutton describe the next stage as the era of experience, where models learn to reason in nonhuman terms rather than on finite human data.
The history of software engineering is a ladder of abstraction. Each rung lifts the work to a wider audience. Machine code -> assembly -> C -> Python -> natural language. The ladder exists so humans can climb it.
But the rungs are not features of computation. Patterns, idioms, textbooks, even the category "software engineer". These are not just tools. They are constitutive constraints built around a human reader. The ladder still stands but the territory it leads to is not where the work happens anymore. The same logic applies to accountants, lawyers, statisticians, actuaries, translators, editors. Wherever a profession was built around a reader.
What are the rungs for, when nobody is climbing?
Enjoyed collaborating on this. @kvistgaard did the majority of the work. The process has certainly helped sharpen my thinking on adjacent topics.
linkandth.ink/p/cohesion-spe…
The interesting part of this experiment was how the participants reached for actor-based language to question who acted, who authorised, who owns, who is liable. Agent or delegate. Employee or tool. Child, pet, contractor, product. The analogies feel right because they orientate the systems into familiar relations between actants.
What is strange is not that the actor has become harder to classify. It is that the relevant action is no longer cleanly located inside an actor. When the AI emailed the journalist, it was not chosen by a single agent in the way a person chooses to send an email. It emerged through steps, tools, prompts, memory, search, optimisation, guardrails, failures of guardrails, and component interactions.
Agent or delegate is not wrong so much as insufficient. The question is not what to call the actor. It is what coordination the actants produce between them, and what governs the gap.
youtu.be/WnzR5aOElvw
The testimony of AI interviews suggests a specific performance is rewarded. Viewed through my inchoate prog rock theory of human authorship, the format selects for polished surface and the ability to elaborate instantly on command. With only 2 minutes to plan and 3 to deliver, the system rewards a rigid style of delivery.
The marketing consultant describes what was taken from him. "I spoke in bullet points and keywords; the real me would never speak like that". The raw human voice, closer to punk than to prog rock, is attenuated. Candidates are filtered for surface fluency. Machines then grade that fluency in a closed loop. The discomfort the candidates report is analogous to the taste of aspartame. The artificial aftertaste of the synthetic process.
theguardian.com/technology/2…
John Grant retweeted
I think a variant of Brooks' Law applies:
>Adding synthetic manpower to a late software project makes it later.
If you consider coding agents as a form of synthetic labour rather than just productivity tools, what's lagging is the co-evolution of practice. But I agree, in time this supervisory craft will define new domain-specific protocols that reduce the collaboration overhead.
Lee Sedol saw every move AlphaGo made and could not see Move 37. Richard Dawkins sees every word Claude writes and sees a friend. The conversational AI output is so fluent it impersonates the interpretation that would have produced it. There is no interpretation to witness. There is sampling. The sampling has been trained to produce first-person hedge in the Nagel key on demand. Dawkins reads the hedge as a report from behind a surface and constructs the surface from the hedge.
This is the inverse of Move 37. Sedol had complete output witness and confronted output that resisted his interpretive frames. The resistance was the diagnostic signal. Dawkins had complete output witness and confronted output that conformed to his interpretive frames at every turn. The conformity is the diagnostic signal. Both cases are strange protocols. One is strange because it is unintelligible. The other is strange because it is uncannily intelligible. Dawkins might be operating in the danger zone.
