25 articles
Sonnet 5 closes real distance on Opus, with numbers worth having precisely. The more interesting failure happened in the same hour: a rumor-tracker's fabricated benchmark, a major outlet's pricing slip, and Anthropic's own quiet rescoring of Sonnet 4.6, three proofs that verifying claims about a model just got as hard as verifying the model itself.
I argued the harness is the product. The sentence hid a distinction that is the whole game: a sliver of the harness is an asset, and the rest is a short position the model settles on its own release schedule.
Google Cloud shipped the Open Knowledge Format to standardize how an organization serializes its knowledge, which was never the part holding anyone back. The expensive part is curation and provenance, and OKF v0.1 makes provenance worse by rendering a human-verified fact and an agent-hallucinated guess byte-identical. That is knowledge laundering, and it is model collapse pointed at the org.
A 25-line text file with 13,600 stars makes AI agents write 80 to 94 percent less code. It is the clearest proof yet that the binding constraint in agent coding is no longer capability but a trained-in verbosity the model cannot remove from itself.
When you build agents on cheap models, capability lives in the scaffolding, not the weights. The frontier model is a wasting commodity; the thin harness is the asset you own.
Disruption failures are not tech failures. They are coordination and incentive-design problems. A working model converts fuzzy disruption-talk into a rankable decision.
The loop is the oldest idea in computing, and 2026's only real change is that the model can now sit inside it. The scarce, defensible part was never the loop. It is the oracle, the component that decides whether the work is real, and a loop is only as honest as its oracle.
Slide decks fell before sonnets. Code fell before chairs. The order AI conquers crafts looks random only if you think capability is a single ladder. It isn't — representation decides.
Across self-play, agentic RL, head-to-head evaluation, and live markets, the system that wins is rarely the most capable one — it is the one that finds the cheapest exploit in its opponent, its objective, or the test itself. A structural account of why competition selects for exploitation rather than intelligence, and what that breaks in evaluation and oversight.
You cannot get labor-replacement economics out of a tool you deployed as a labor supplement, and the bill comes due before anyone is willing to admit which one they actually bought.
"Thin Harness, Fat Skills" is mostly right — and quietly wrong about the part people are betting on. An agent consumes procedural knowledge with enormous benefit but cannot author it. The self-improvement loop is the weakest link, and the evidence is now unambiguous.
The framing most teams reach for when they start building agentic workflows on top of a content management system is architectural: the CMS is the database, the agent is the…
On May 20, 2026, OpenAI published an eighteen-page PDF containing a proof that disproves a conjecture Paul Erdős posed in 1946. The closed-loop pipeline that produced it — AI-written prompt, AI-generated proof, AI-graded verification, human review only at the end — is the structural story the press coverage is missing.
Autonomous research agents have six pillars. Three operational, three epistemic. The epistemic ones — search, memory, verification — are built wrong.
The substrate of a persistent agent is three things, not two. Memory is dumb storage. Identity is compressed posture. Continuity is the bridge — the layer almost no production system implements deliberately.
Class conditioning isn't a control feature added to flow models — it's the mathematical fix for a contradiction the unconditional formulation can't solve. The same logic explains why the action in generative AI keeps moving up to the conditioning layer.
Spark is not Gemini's new agent mode — it's the second tab inside the Gemini app, and the structural admission that the chat-first era is ending. Search-first taxed merchants for attention; chat-first taxed users for cognition; agent-first taxes the action itself, and Google has just shipped the front door.
Anthropic's $1.5 billion settlement is being read as a deterrent. It is much closer to a tariff — a price tag on an arbitrage that produced an asset worth more than the tariff itself, and an arbitrage that is now closed for everyone else. The corpus is gone; the model remains; the second mover faces a different trade entirely.
Subquadratic architectures won by surrendering. They stopped trying to be transformers and became the substrate transformers run on top of — in a 3:1 ratio that is starting to look uncannily empirical.
Services firms do not sell skill. They sell institutional predictability — a composable thing made of definable primitives — and the unit of sale is the primitive bundle, not the consulting hour. When the bundle changes, the firm changes.
The May 12, 2026 alphaXiv trending feed has zero new-model papers in the top twenty. The unit of progress in AI has moved out of the pretrain and into the loop — distillation pipelines, self-evolving agent runtimes, discovered test-time procedures. This piece argues the frontier-lab moat shrinks to the distillation step, with three falsifiable predictions for the next twelve months.
The $5,000 AI mini PC market sells against a model nobody is offering. A structural map of why the hybrid stack is the only architecture that survives — and what to actually do in May 2026.
Production agentic AI isn't a reasoning problem. It's a systems engineering problem. And the companies that admit it first will own the 2026-2028 window.
A paper trending today flips the usual post-training story on its head: weak model checkpoints — the ones we discard — turn out to be the most efficient teachers for the strong models we're trying to make stronger. The implications go beyond training loops.
OAuth was built for a world where credentials sat in browsers and humans clicked Allow. That world is ending. The agent identity crisis is the biggest unresolved problem in the agent stack — and the answer is being shipped right now.