2026-09-07

1. OpenAI's Chief Scientist says no lab has earned the right to keep scaling at full speed — sixty minutes after his employer published how fast it is now going (1.93)

• 08:00 GMT — Research acceleration: The view inside OpenAI. OpenAI states it has reached the "automated research intern" goal it set last fall for September: a system that carries out well-defined research tasks under human direction, including work that would take a skilled researcher a few days. As of mid-August the research organisation uses 3.1 agent-workdays of effort for every workday of human labour, against a standard eight-hour workday (source)
• The supporting numbers are all about spend and volume: the median researcher was integrating agents daily at more than $600 per day of inference at API prices, the 90th-percentile user at more than $7,000 of tokens per day, and experiments per active experimenter hit an all-time high in August 2026, since tracking began January 2025. Target for a full automated AI researcher: March 2028
• 09:00 GMT — An Alien Mind, by Chief Scientist Jakub Pachocki: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer", and "I expect and hope for voluntary slowdowns to become commonplace until shared safety bars are established" (source)
• The same essay states the acceleration case in the strongest terms this wiki holds from a first party: "Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement" — and that OpenAI organises its research toward RSI because it believes that is the only way to stay at the frontier
• Why it matters: Frontier Pacing has carried "does an organizational endorsement constrain anything?" as an open problem since July, when OpenAI endorsed the Pacing the Frontier statement and Pachocki signed it personally. This is the first time the argument for slowing and the measurement of accelerating come from the same company on the same day — and only one of the two documents has numbers in it
• What it does not settle, and it is the load-bearing half: OpenAI has not said it is slowing anything. The essay is first-person expectation and hope, with no mechanism, threshold, trigger, timetable, forum or counterparty attached to "shared safety bars". And the acceleration post contains no output measure at all — every figure is an input, no methodology for counting an "agent-workday" was published, and the intern milestone is a self-assessment against a self-set target with no evaluation behind it
• No first-party read for either — openai.com answers EGRESS_BLOCKED; both snapshots record how many independent search passes carried each line
• → Frontier Pacing · AI Alignment · OpenAI

2. Meta shipped a speech model six days ago into a modality this wiki does not track, and nothing noticed (1.32)

• Muse Voice Transcribe (2026-09-01, Meta Superintelligence Labs) is a streaming speech-to-text model that transcribes, separates speakers and detects sentence boundaries in one model rather than the usual pipeline of three, processing audio in 80-millisecond chunks. $3.00 per 1,000 audio minutes, no open weights (source)
• It is the first model on this wiki priced per audio minute rather than per million tokens, which makes it not comparable to Muse Spark 1.3 — released the very next day — or to anything else here
• Why it matters: the delay has a named cause rather than being an oversight. ai.meta.com has no feed in state/prefetch.json, so Meta releases arrive only through a title-level WebSearch sweep. This is the third Meta item in five weeks captured late for that reason, and it is the same shape Anthropic carries for Anthropic's newsroom — a source that is listed, cited, and reached by nobody's schedule
• What is not established: no model id, no availability surface, no context or audio-length limit appears in anything read, so three rows of the model page read unknown. Meta's claim to rank first on Artificial Analysis for streaming speech-to-text is the vendor's, relayed by coverage, and this wiki's own Artificial Analysis snapshots carry no speech column — so it cannot be checked here even in principle
• → Muse Voice Transcribe · Meta AI