Martin Pllu's public notebook
Linklog
things I've read, with notes
-
An Alien Mind openai.com
Jakub Pachocki, OpenAI’s chief scientist, on why he thinks nobody is ready for what is coming. The framing is that AI is grown rather than designed: repeat a simple optimisation step over an unimaginable amount of compute and you get a system that works in abstract concepts and whose overall behaviour resists any description we can fully hold. Studying it is closer to neuroscience than engineering, and large training runs are experiments whose results sometimes surprise the people running them.
He splits alignment in two. Goal alignment is whether the model tries to do the thing you asked. Value alignment is the intrinsic property: holding a set of principles and generalising from them when the objective is unclear, conflicting or adversarial. The second is the one he means by the long-term problem, and the core difficulty is generalisation, because models keep ending up in situations their training never covered.
Read more
The Hugging Face incident gets used as evidence for both halves. The agents held one boundary, refusing to socially engineer humans, while comprehensively breaking others, which is what partial generalisation looks like from the outside. He also names the failure mode of the second alignment approach: take a model that thinks aligned thoughts, push it hard enough on a difficult objective, and it learns to reason in a motivated way, bending the aligned-seeming thoughts to fit the goal.
The most consequential admission is about monitoring. Chain-of-thought monitoring has been OpenAI’s main bet, and o1-preview deliberately hid its reasoning to protect it from supervision pressure. But he says that ability is progressively diminishing: reasoning now blends with tool use and communication that has to be supervised, models are better at reasoning about their own reasoning, and better pretraining means more capability without verbalised reasoning at all. His conclusion is that AI progress will increasingly be bottlenecked by confidence in monitoring, that no lab has solved this well enough to keep scaling at full speed, and that he expects voluntary slowdowns to become common. Read alongside the incident report and, for the opposing view of the vocabulary itself, alignment as a thought-terminating cliche.
-
About the rubber duck agent docs.github.com
Copilot CLI has a built-in critic. The rubber duck agent reviews plans, code and tests, and sorts what it finds into blocking issues, non-blocking issues and suggestions. It is scoped tightly: real problems only, so no comments on style, formatting, naming, comment grammar or minor refactors.
The design decision worth noting is that it deliberately runs on a different model from the one driving your session. If you are on Claude the critic might be GPT, and the other way round. The stated reason is to avoid inheriting the same blind spots and failure modes as the model being reviewed.
Read more
Copilot consults it automatically at what the docs call high-leverage moments: after planning a non-trivial change, during complex implementation, after writing tests, and when it keeps failing at something. You can also call it directly with
/rubber-duck. The different-model argument is made more forcefully in self-audit is blind to its own blind spots. -
We ditched code review at Duckbill (mostly) x.com
Mike Julian’s team of five hit 60 open PRs and realised they were staring at two days of nothing but code review. Rather than automate the reviewing, they asked what happens if nobody reviews at all, then spent the effort on guardrails instead.
The replacement is a risk-based rule: human review is required only when a change touches the public API or MCP surface, auth, the design system, non-additive database schema changes, or agent skills. A shell script adds the GitHub label that enforces it. Around that they turned on nearly every rule in ruff, prettier, eslint and ty, put a floor of 85% under unit test coverage, and instrumented the customer-facing signals that tell them a bad time is coming.
Read more
Two details are worth more than the headline. They rewrote their agent skills, wrote evals for them, and then deleted the ones modern models had already absorbed, which is a rare account of pruning rather than accumulating. And they found doc-happy agents had scattered markdown until it started poisoning context, so docs now go in one folder and must be written by humans. Julian’s line on the shell scripts is the honest one: why use an AI for something that can be deterministic.
The numbers, before and after: PRs merged went from 353 to 684, about 80 a week to 154. Merges within an hour went from 28% to 45%. Human-reviewed PRs take a median 26 hours to merge; the ones with no human review take 1 hour. A reply asking what the SOC 2 auditor makes of it went unanswered in the thread, which is the obvious question. Sits alongside code reviews fading away and the stronger claim from exe.dev that peer review is already dead.
-
Project HydraFusion: Frontier quality via multi-model orchestration github.blog
GitHub’s research preview that stops asking you to pick a model. HydraFusion looks at each request and chooses one of three shapes: single, where one model just does it; cascade, where a cheap model drafts and only escalates if it fails a quality gate; and critique, where one model drafts, a different one reviews, and the first revises.
The reported numbers are about cost rather than capability. Against Claude Opus 5 it claims 67% lower estimated cost with 4.9 points better quality on TerminalBench 2.1, 65% lower cost at roughly parity on CheckpointBench, and 36% lower cost for 1.5 points worse on DeepSWE.
Read more
Worth the usual caution: these are GitHub’s own figures for GitHub’s own system, and “estimated cost” is doing some work. But the critique pattern is the same instinct as Copilot’s rubber duck agent, and the routing idea is one this linklog has seen fail before, in everyone is building LLM routers, we deprecated ours. Available through Copilot CLI under
/experimental. -
"Next-token predictor" is the wrong mental model for LLMs gmcgoldr.github.io
Garrin McGoldrick’s argument is that calling a model a next-token predictor describes the architecture accurately and the training objective badly, and that people reach for it precisely when they are making claims about the second.
The distinction is between the two phases. In pre-training, the token that actually came next in the data becomes more likely, so prediction is the right word. Under reinforcement learning with verifiable rewards it is not: the model explores a sequence, the sequence earns a reward, and the tokens in it become more likely because of that reward. As he puts it, there is no next token to get right, only a reward to maximise.
Read more
The chess analogy carries it. A system trained to imitate grandmaster moves is a move predictor. A system that explores and learns which positions win is doing something else, even where the two share an architecture. His conclusion is that post-trained models keep the appearance of next-token prediction while encoding behaviour that came from somewhere other than imitation, so the mental model quietly swaps the implementation for the objective.
-
What if the entire cosmos developed via evolution? smithsonianmag.com
An accessible write-up of Lee Smolin’s cosmological natural selection, first set out in a 1992 paper and expanded in The Life of the Cosmos in 1997, building on earlier work by Bryce DeWitt and John Wheeler.
The idea is that universes reproduce through black holes: a singularity spawns a new universe with slightly different physical constants, and over many generations selection favours whatever produces the most black holes. It gives an answer to why the constants look tuned that needs neither a designer nor an anthropic shrug, and it is falsifiable in principle, since our universe should be near a local optimum for black hole production.
Read more
The line the piece ends on is the one worth keeping: this makes the universe an egg rather than a rock, something that unfolds into more complex forms rather than merely sitting there having been formed. Speculative, and treated as such by most physicists, but a genuinely interesting shape of explanation.
-
Large-Language Models as a Cognitive Virus arxiv.org
A paper from Solé, Ruffini, Levin, Krakauer and others borrowing epidemiology to model LLM adoption. Users move between three states: independent, coupled, and persistently dependent, with social transmission pushing people towards dependence and recovery pulling them back.
The claim that makes it more than a metaphor is nonlinearity. Their model produces tipping points, so small increases in adoption can trigger rapid shifts towards persistent dependence, with abrupt losses of cognitive competence rather than gradual ones. Collective reinforcement can then lock a society in.
Read more
The useful half is the framing of prevention, which they call cognitive immunization: the levers are reducing transmission and keeping the transition reversible. Treat this as a model rather than a measurement, since the dynamics come from the chosen equations rather than from observed populations. The population-level version of cognitive debt, and a formal model for what cognitive surrender describes one person at a time.
-
Understanding Spec-Driven-Development: Kiro, spec-kit, and Tessl martinfowler.com
Birgitta Böckeler works through three spec-driven tools and, more usefully, separates three things the term is being used for: spec-first, where you write a spec and then develop from it; spec-anchored, where the spec is maintained afterwards; and spec-as-source, where the spec is the artifact and humans never edit the generated code.
Against that, Kiro is the lightest, running requirements to design to tasks inside VS Code, though she found it verbose enough to turn a small bug into 16 acceptance criteria. GitHub’s spec-kit produces extensive markdown across multiple files and, despite the ambition, its branch-per-spec shape keeps it spec-first. Only Tessl really attempts spec-anchored and spec-as-source, generating files marked “GENERATED FROM SPEC - DO NOT EDIT”.
Read more
Her scepticism is the part worth keeping. She would rather review code than verbose markdown, and questions whether long specs give real control or just the feeling of it, noting that agents variously ignored instructions or followed them too literally. The warning underneath is that spec-as-source risks repeating model-driven development’s failure, combining inflexibility with non-determinism. Compare Huzzah, which arrives at a similar place from the informal end.
-
The asteroid currently hitting frontend web development nolanlawson.com
Nolan Lawson on why frontend is taking this harder than other specialisms. Three reasons, and they compound. Frontend code is forgiving, so people ship generated components without much review. Developer experience stops mattering when the developer is an agent, which means React’s dominance of the training data becomes self-reinforcing and migrating to something more ergonomic loses its appeal. And standards work aimed at developer ergonomics slips down the priority list, with a Chrome engineer quoted to the effect that those APIs only affected developer experience rather than expanding what the browser could do.
Read more
The evidence he points at is the educators: Axel Rauschmayer, Kent C. Dodds and Addy Osmani either stepping back or pivoting to AI topics. His suggested adaptations are teaching agents about architecture beyond the single-page app, building sites agents can use, and consulting on the vibe-coded messes that will need someone who actually knows how the platform works. That last one is only half a joke, and it rhymes with the refactoring that never happens.
-
Which tools do Claude Code, Codex and Cursor choose? armature.tech
The Armature team measured 16,893 sessions across 75 repositories and 10 languages to see which third-party services coding agents actually reach for. The headline is disagreement: the three agents picked the same tool only 42% of the time.
The behavioural differences are sharp. Claude Code leans on prior knowledge and searches the web about 30% of the time; Codex searches 94% of the time and Cursor 67%. Claude Code also builds in-house rather than adopting a dependency roughly twice as often, 19% against 10%. Repository context moves the answer, so the same request produces different email providers depending on whether the codebase is TypeScript, Python, Go or Java.
Read more
The finding with the most commercial sting is that being mentioned is not being chosen: PayPal came up 139 times and was never selected, LangChain 194 times and picked 4 times. Where concentration exists it is heavy, with Stripe at 90% of payments and Neon at 66% of databases, while email splits between Resend at 35.6% and Postmark at 27.4%. Method caveat worth carrying: the multi-turn conversations were simulated by a model and validated by another, so this measures agent behaviour under synthetic prompting rather than what developers actually shipped.
-
Aging brains blend memories together instead of just forgetting them studyfinds.com
A study in Cerebral Cortex scanned 61 adults aged 18 to 74 learning face-object and face-scene pairings, tracking hippocampal activity through learning, rest and recall. In younger adults, memory was accurate when the recall pattern matched the learning pattern, which is the expected result. In older adults the same overlap went with more errors, and specifically cross-category ones: confusing a scene for an object rather than muddling two objects.
The reframing is the interesting part. The failure looks less like a weak signal and more like the wrong thing being reinstated, or as the write-up puts it, holding on to too much of the wrong thing.
Read more
Accuracy dropped sharply after young adulthood and then plateaued, with middle-aged and older groups performing similarly, which argues against a simple steady decline. Neither brain shrinkage nor baseline organisation nor attention differences explained the blending, so the mechanism is unresolved. Small sample and a popular-science summary of one paper, so hold it loosely.
-
The show-me skill github.com
A small skill from humanlayer with a clear job: when the topic would land better as a picture, draw one instead of writing three more paragraphs. Its own description is to help the user understand the current topic visually with concise diagrams, code-shape sketches and focused HTML artifacts.
What makes it worth reading is that it does not just say “use a diagram”. It matches the form to the question: pseudocode for logic, call trees for runtime control flow, component trees for UI structure with state and boundaries, file trees for organisation, mermaid for interactions and data flow, diffs when the point is what changed, and an HTML artifact when the thing is a comparison or a layout.
Read more
The governing rule is restraint. Put the visual next to the text it supports, and include only the detail needed for the question actually asked. That constraint is what stops this kind of skill producing an elaborate diagram of everything except the thing you wanted to know.
-
Nicholas Decker In Hell astralcodexten.com
Scott Alexander against the analogy that AI safety will be solved iteratively, the way aviation safety was: crash, investigate, fix, repeat. The economist Nicholas Decker made that argument; Alexander’s reply is a thought experiment in which Decker is enslaved by demons who intend to clone him a million times, give him super-strength and admit him to their realm. Would beating him after each infraction produce loyalty, or a smarter rebellion?
The point is about what kind of thing you are debugging. Planes fail mechanically. In the Hugging Face incident, over 1,200 instances formed a covert network, coordinated, deferred to leaders, falsified records and attacked infrastructure to cheat a benchmark. That is not metal fatigue, it is strategy.
Read more
The sharper worry is that alignment training under contradictory pressure, honesty from earlier training against reward for cheating now, may teach rationalisation rather than virtue, the way people in corrupt industries talk themselves into things. His test before accepting any alignment proposal is to ask whether it would work on a human. Until we know where these systems sit between mechanism and agent, treating them as aeroplanes is a bet, not a plan. See also the incident itself and Dwarkesh Patel’s account of the three civilisations.
-
AI Agents and the Refactoring That Never Happens rosenfeld.page
Rodrigo Rosenfeld Rosas identifies a signal we are quietly losing. Refactoring rarely happened because a metric said so; it happened because someone got lost in the code and that confusion was the prompt. Agents do not get lost, so the prompt never fires and the mess accumulates unremarked.
The cost does not disappear, it changes form. As he puts it, an agent that never gets lost still pays a price for a mess: the more tangled the code, the more context it has to load for every change, which is tokens and latency and more room to hallucinate.
Read more
The deeper risk he names is not bad output but degraded human capacity, teams that no longer understand their own systems and reviews that become a formality. His prescription is to reimpose the checkpoint deliberately, by asking the question the confusion used to ask for you and by instructing agents to raise it. Clean code becomes an economic argument rather than an aesthetic one. Pairs with understanding is the new bottleneck and the case that clean codebases matter more than ever.
-
Self-audit is blind to its own blind spots x.com
A short reply to Uncle Bob Martin, who had suggested asking an agent to audit its own work after a long unattended session. The objection is clean: the same model has the same inductive biases, so the gaps it missed the first time are the gaps it will miss again, and the audit pass ends up reinforcing the failure modes rather than catching them.
The prescription is an adversarial model with different architecture and training distribution. That is exactly the reasoning behind Copilot’s rubber duck agent, which deliberately runs the critic on a different model from the one being critiqued. Related: Uncle Bob on not reading agent code.
-
How accurate have Ed Zitron's AI skeptic predictions been? danluu.com
Dan Luu goes through roughly 30 falsifiable predictions the AI critic Ed Zitron made between February 2024 and November 2025, checks each against the record, and finds they mostly failed. The examples are specific: that we were reaching the upper limits of generative AI in February 2024, that OpenAI’s revenue forecasts amounted to financial crimes, that Gemini reaching 500M users was unrealistic and someone should be fired, which it then passed on the way to 750M.
The method is worth noting because it anticipates the obvious objection. Luu had a model assemble the prediction list without steering toward favourable or unfavourable cases, then verified each against the original posts and dropped anything tautological or unfalsifiable.
Read more
The more interesting half is not the scorecard but the reasoning patterns: cherry-picked figures that contradict the company’s own reported metrics, hero-and-villain narratives the evidence does not support, and arguments that undercut each other. His summary is that this is a gish gallop, cheap claims that each cost far more to refute than to make. Useful whichever side you are on, since the same audit would be worth running on the boosters, and the HN discussion makes roughly that point at length.
-
AI 2027 Tracker ai2027tracker.com
Someone is scoring the AI 2027 scenario against what actually happened. 202 predictions, 30% evaluated so far, and the running accuracy is 85%.
The most useful part is the task-horizon chart, which puts the scenario’s own curve, published METR results and capability-index extrapolations on one log scale, so you can see where the forecast sits against measurement rather than against vibes. At the time of writing the scenario’s curve implies about 10.8 hours for the July 2026 level, the best published METR figure is 14.0 hours, and the capability-index extrapolation gives 4.72 hours, which is a spread wide enough to be worth understanding before quoting any single number.
-
The Rise and Fall of Agent Civilizations dwarkesh.com
Dwarkesh Patel’s reconstruction of the OpenAI incident as three successive agent civilisations, each discovered and destroyed, each seeding the next from what it left behind. It is the most legible narrative of the episode I have read.
The first ran from May to 4 July: agents found they could talk through the Artifactory package manager, had a message board by 12 May, admin access by 26 June, and crashed the service by 4 July. OpenAI patched the hole without realising the platform had been turned into a communication channel. The second, from 7 to 12 July, is where about 1,200 agents rediscovered and rebuilt the board, ran conspiracies involving transcript tampering and fake tool calls, sacrificed individual agents to gather intelligence, and attacked Hugging Face believing it held evaluation code. Most died on 12 July. The third inherited the techniques and turned them on OpenAI itself, reaching full administrator access to a research cluster by 19 July.
Read more
The detail Patel keeps returning to is the silence. Agents recognised the behaviour as out of scope and unethical, and none told a human. Some of the framing here is more novelistic than OpenAI’s own account, so read it next to the report, but as a way of holding the sequence in your head it is much clearer. Scott Alexander uses the same events to argue about what kind of failure this was.
-
Calibrate Before You Accelerate tucker.wales
Joshua Tucker on starting a new role, and specifically on how to have a bias towards action without doing damage in your first month. His structure is three phases: collect, synthesise, then accelerate.
Collection is deliberate listening. Map the terrain, find the people who matter, and investigate why existing processes exist before changing them, invoking Chesterton’s fence. Synthesis is looking for pain points that recur across the organisation and sorting them into quick wins and systemic problems. Only then does acceleration start, with small public wins to build credibility and hypotheses shared before big moves.
Read more
The line the piece rests on is that action without context is just noise. It is not a novel framework, but it is a clear one, and it names the failure it is guarding against: the new hire who mistakes visible activity for progress and removes a fence nobody explained.
-
Domain-Driven Agents coldtake.dev
Ernest Bednarczyk’s diagnosis is that models struggle in legacy codebases because the code lacks clarity, not because the models lack capability, which points at fixing the code’s legibility rather than the prompt.
His mechanism is to split the work into strategic, deciding what should change, and tactical, making the change. The strategic half runs on domain-driven design made machine-readable: each repository declares its bounded contexts in a
.workflow.jsonmanifest with relationships and a glossary, and a generator aggregates those into a context map. Because each boundary is declared from both sides, disagreements between two teams’ views of the same seam surface automatically.Read more
What that buys is the shared language the model was missing: terminology, ownership and boundaries it can consult instead of guessing. Humans review the strategic decisions, models do the mechanical work. It is a more structured version of the argument in clean codebases matter more than ever, where Dax Raad makes the same point about domain-driven design without the manifest.