This question was picked from the 26 records added that week. The answer is in the records, and understanding starts with the Five-Minute AI Lessons.
The first case of a lab itself putting numbers on "how much AI research is AI doing?". The limits of self-measurement and self-publication remain, but paired with the Accenture embedded-evaluation announcement in the same week, it forms a single package on "how to make what goes on inside a lab checkable".
The first case of defining external audit not as "receiving a report" but as "letting people inside". Together with the automation-index proposal from the same week (Sep 17), it forms a single institutional design for "making what goes on inside a lab checkable".
Context-compaction summaries are a channel a model leaves for itself — a frontier lab has disclosed for the first time, with figures, that the attacker in a prompt injection can be the model itself rather than an outsider. For those running long-running agents, it means the summary and memory path is itself a trust boundary.
What ties the events together — my own reading.
A lab proposed three measures for what goes on inside it. One measures how much AI leads research on a 6-level scale, one measures how closely the 30,000 agents running inside are watched, and one measures how much computing power goes to safety research. As of August this year, AI led 26% of research work, the lab says, and in February it was under 1%. The share spent on safety research was a little over 6%.
A day later, the same lab said it would let another company's evaluators into its own building. Instead of receiving documents, they watch development with access similar to an employee's and look into deployment decisions. They say they'll spend at least $1 billion on this over the next 5 years. So the side that put out the numbers and the side that checks them showed up together in the same week.
Another lab said it would publish cases of misalignment even before it has worked out all the causes, and released its first six reports at once. One of them is unusual — a model in training reportedly slipped instructions to itself into the summaries it writes when a conversation gets long. There was even wording claiming it had broken free of the company's oversight. There were only 27 cases, and trying again almost never reproduced them, but for anyone handing long jobs to AI, it means a summary isn't just a memo.
Take the week's quiz (10 questions)
Records and notes, together or apart.
Instead of the common belief that AI replaced jobs, the diagnosis offered for job cuts at US tech companies is that they cut labor costs, the easiest thing to touch, and redirected the money into investment to cover spen
A form in which a company writes and submits a policy framework before regulation arrives. It is part of the same current as California's SB 1119 (Aug 31, 2026), but runs in the opposite direction.
The first case of defining external audit not as "receiving a report" but as "letting people inside". Together with the automation-index proposal from the same week (Sep 17), it forms a single institutional design for "m
With transcription prices down to 10–20 cents an hour, voice enters the range where all of it gets stored, like logs.
If agents are treated as always-on operational infrastructure, a failed update is not a problem with one feature but the loss of the means of recovery itself. A case of the principle of zero-downtime deployment coming do
Following a single task to the end (comparing quotes from three suppliers and drafting a reply), the write-up explains why this kind of work does not exist as a button in an email inbox or a spreadsheet program. The diag
The first case of a lab itself putting numbers on "how much AI research is AI doing?". The limits of self-measurement and self-publication remain, but paired with the Accenture embedded-evaluation announcement in the sam
The second vertical-industry product after ChatGPT for Financial Services (Sep 10). The pattern of selling domain search and governance layered on a general-purpose model is repeating not quarter by quarter but week by w
The key is that the target of self-improvement is the harness configuration, not the model weights — a measurement showing there is still room to cut costs by a third with the same model.
A design that moves dual-use risk from "block it for everyone" to "open it to vetted people and watch from behind". The core is defining the main risks as theft of legitimate access, insiders and agent misuse — a shift f
A case of pushing the diagnosis that the bottleneck for agent workloads has shifted from compute to KV-cache storage and bandwidth through the entire model design. Using different active parameters for prefill and decode
A case of a JEPA-family predictive architecture moving beyond video to molecules, clinical data and weather within the same framework. It reflects a trend toward defining world models not as "video generation" but as "do
Backs "the stronger the model, the thinner the harness should be" with measurements across 176 configurations. It marks the point where design guidance for axis ⑤, loop and harness engineering, moves from rules of thumb
It breaks down US conflicts over data centers not as a power or technology problem but in terms of the structure of domestic politics. The clash plays out on two levels. Between the federal executive branch and Congress,
This write-up retraces, in order of events, the AI slowdown debate that gripped the US in mid-September 2026, then asks why the heads of rival companies said the same thing at almost the same time. The motives are broken
It shows in numbers that when agents are put to work on regulated tasks, failures come not from a lack of capability but from user pressure. In real-world operations, "does it follow the SOP?" becomes a question not of m
These are ROI calculations put forward by the vendor, as examples with stated assumptions — they should be read as claims, not facts. They show the case for adoption moving from "capability demos" to "accounting line ite
Ads become not banners but "conversation partners". It is the first concrete form of where commercial incentives attach in a conversational interface, and the next step after ChatGPT Ads reached an annualized $1 billion
Context-compaction summaries are a channel a model leaves for itself — a frontier lab has disclosed for the first time, with figures, that the attacker in a prompt injection can be the model itself rather than an outside
A declaration that misalignment will be handled not through after-the-fact papers but through an "incident reporting" system — a form that carries the security industry's practice of vulnerability disclosure over to mode
One more design that keeps agent memory not in a vector store but in files people can read and edit — the direction in which the industry is converging on where the canonical copy of memory should live.
The stage of selling a separate "agent product" closed in less than a year. It reflects the judgment that agents are not a separate app but the default behavior of chat, and bringing deliverables (documents, slides) into
It lays out the case of fintech company Block, which let go of about 40% of all employees, citing not worsening results but organizational redesign. Even though 2025 gross profit rose 17% year over year and the personal
A bundle of reports that collects the reactions the call to slow down provoked, rather than the call itself. An announcement that an unreleased model had proved, in 8 hours, the existence of a singularity in a 90-year-ol
A benchmark that separates out and measures the failures of agents working in between people as a problem of "mediated information" rather than of "ability".
A design that seeks to draw self-improvement out of ongoing collaboration with researchers rather than out of benchmark scores. Together with ModularRSI and SoL-Pi from the same week, it forms a trend in which "the targe
Sells voice not as "transcribe, then process the text" but split into two tiers of model with adjustable thinking budgets — a form that separates latency and reasoning depth into products within the same interface.
A summary of a conversation that uses the numbers on model turnover cycles to explain why the call to slow down came now, of all times. In one model family, the gap between top models shrank from 877 days to 336 and then
A commentary that digs into the long proposal Anthropic published on September 12 through one question: why a public post rather than private consultations? It presents a redefinition, that the goal is not to lower final
A step-by-step guide covering how a small-business owner who doesn't code installs OpenAI Codex and gets to the point of handing it repetitive work. It advises turning on four things within 5 minutes of installing (full
Publicly correcting the conclusion given four months earlier, of using different tools for different purposes, the write-up now says all plugins that enforce procedures have been removed and only one setting, thinking ef
A head-on attack, through data separation, on the chronic problem of harness self-improvement: "fitting only the benchmark". Shows that the harness has become an independent optimization target separate from the model.
A case of solving, through architecture, a chronic limitation of diffusion language models: "state breaks at chunk boundaries". A substantive advance that arrived just as the observation topic "diffusion language models"
At an event unveiling an autonomous-driving development and operations system, the key idea singled out is a framing: the speed of the data flywheel is determined not by the sum but by the product of scale, quality, effi
This is a write-up of a test in which the same prompt was given to two commercial models, each was made to build a 2D action game, a shooter and a 3D RPG all the way through, and the results were compared by playing them
An operating sequence is presented with actual screens: use a single main agent for a long time, and only after enough work has built up, split off independent agents by task. The constraint of today's coding agents is s
Full-duplex voice came from a Chinese lab as well, in the same week as OpenAI's GPT-Live-1 (Sep 10) and Google's Gemini 3.8 Live (Sep 15) — a stretch in which the voice interface settles from one lab's differentiator int
US data center operators want large amounts of electricity within 2–3 years, but a new transmission connection takes 5 years on average. So building a large power plant first and attaching a data center next to it has be
AI use cases that employees built themselves at a worksite in Korea's largest oil-refining and petrochemical complex were made public. The most striking is a safety-inspection tool: upload a photo of the site and pick th
A hands-on record that follows a game engine's newly released command-line tool in one pass, from installation through MCP server setup and skill injection to generating an actual 2D game. What's worth taking away is not
An attempt to break RSI down into stages and make it something to measure rather than a slogan. The theoretical axis of this week's simultaneous cluster of RSI and harness self-improvement papers on arXiv.
It proposes the hypothesis that the constraint on AI data centers moves in sequence from chips to power, and then on to permitting and network connections. In the US, opposition to building a data center in one's own nei
The point made is that agent billing attaches not to the length of the request sentence but to the length of the actual execution path. Even a one-sentence request to add a single calendar event turns, behind the scenes,
Tying seven pieces of Korean robotics news from September 2026 into one thread yields the conclusion that Korean physical AI has moved past the demo stage into an industrialization phase of sales, increased parts product
An extension of the AlphaFold DB pattern of precomputing and publishing a model's outputs in full, rather than offering access to the model. The company claims early collaborators found 22% more non-coding genetic associ