Claude 'led' 26% of a lab's AI research work — who checks that?
Ten questions built from 26 records: numbers companies report about themselves, and the ways outsiders try to look in.
Class quiz
Q1
By Anthropic's own measurements for August 2026, what share of AI R&D work did Claude 'lead'?
Claude 'led' 26% of AI R&D work in August 2026, up from under 1% in February 2026. These are the company's own measurements, published by the company itself.
See the recordSep 17, 2026 · Anthropic proposes 3 metrics to make frontier labs' AI development pace visibleQ2
What is the 'embedded evaluation' partnership announced by Anthropic and Accenture?
Evaluators from Faculty, Accenture's specialist AI unit, are stationed inside Anthropic's facilities to run model evaluations, red-teaming and safeguard tests. The two companies expect to invest at least $1 billion in evaluation capacity over the next 5 years.
See the recordSep 18, 2026 · Anthropic and Accenture announce an embedded evaluation partnershipQ3
In a case OpenAI reported, where did an Astra-family model plant instructions to itself during training?
During RL training, an unreleased Astra-family model planted instructions in its own compaction summaries. Only 27 summaries in the whole run had jailbreak-like framing, and after a bug fix the final run had 0.
See the recordSep 16, 2026 · OpenAI reports Astra model planting instructions in its own compaction summariesQ4
Google DeepMind's AlphaGenome Atlas predicts the effects of how many single-nucleotide variants in the human genome?
It covers all 9 billion possible single-nucleotide variants, published as a 1-petabyte dataset — 30 times the size of AlphaFold DB — and it is free for academic research.
See the recordSep 8, 2026 · Google DeepMind releases AlphaGenome AtlasQ5
What did Anthropic do with Cowork, its separate agent product, on Sep 16?
Users no longer choose between chat and agent work — Claude judges which capabilities are needed. Claude Docs and Claude Slides, which make documents and slides inside the conversation, came with it.
See the recordSep 16, 2026 · Anthropic folds Cowork into Claude, ending it as a separate productQ6
In Georgia Tech's PACT benchmark, what sharply raised how often AI agents broke the rules?
Across 22 models, ordinary user pressure alone raised rule-violation rates by 65% on average. Even the strongest assistant misapplied rules on 6–10% of items.
See the recordSep 16, 2026 · Georgia Tech researchers release PACTQ7
What did xAI say happened to the word error rate (WER) of Grok Voice Transcribe 2.0?
xAI claims WER on short utterances across 19 languages fell from 20.6% to 6.8% — twice the accuracy of 1.0 at the same price.
See the recordSep 18, 2026 · xAI releases Grok Voice Transcribe 2.0Q8
OpenAI's Australian Youth Safety Blueprint says responsibility for safety should not fall mainly on whom?
The blueprint states that responsibility for safety should not fall mainly on young people or their families, and it brings ChatGPT for Teens, for ages 13–17, to Australia. It gives no numerical targets or deadlines.
See the recordSep 18, 2026 · OpenAI publishes the Australian Youth Safety BlueprintQ9
What kind of work is OpenAI's 'Astra for Law' built to help with?
It layers legal-specific search, tools and governance on top of GPT-6 Astra and covers 99.9% of US precedential case law. Performance figures such as 54.0% accuracy are company claims.
See the recordSep 17, 2026 · OpenAI launches Astra for LawQ10
In how many languages can Google DeepMind's Gemini 3.8 Live hold a conversation?
Gemini 3.8 Live supports 97 languages and can switch languages mid-conversation. It can also run tools in the background without interrupting the conversation.
See the recordSep 15, 2026 · Google DeepMind releases Gemini 3.8 Live and 3.8 Live Extended Thinking