Epoch AI reports that OpenAI tripled its revenue run rate in the past year from $13 billion last August to over $40 billion now, and that Anthropic grew its revenue from $1 billion to $9 billion in 2025 and more than tripled its run rate in the first quarter of 2026.
The authors of WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution found that WikiSkill outperforms state-of-the-art skill-evolution methods, that evolved skills transfer effectively across models, and that smaller models with skills can outperform substantially larger models without them.
A Mexican appeal court set the amount of a procedural bond by putting the question to ChatGPT, Grok and Gemini, received MXN $60,081, $64,655.34 and $59,864.98, averaged the three and ordered $60,000. Neither party had asked for it. The ruling produced a non-binding precedent listing proportionality, transparency, data protection, applicability and human oversight as principles for the judicial use of AI, and the two lawyers reviewing it for the International Bar Association note that it does not record what variables the three systems used or why their answers differ.
India’s Supreme Court AI Committee published draft regulations for artificial intelligence in courts on 3 June, and a commentary published this week reads their single grievance mechanism as reaching only harm that follows from a prohibited use. Transcription is a permitted use subject to certification, and the Kerala High Court has made the speech-to-text system Adalat.AI the default for witness depositions in its district courts since 1 November 2025. The gap is the commentator’s reading of the draft rather than anything the draft states, and he is a law student writing on a university centre’s blog.
The Deliberative Deficit: An Empirical Critique of LLMs in Democratic Discourse finds that across 1,980 five-agent LLM runs on 12 citizen-assembly topics and 11 frontier model configurations, LLM groups produce discourse with procedural quality comparable to human deliberation but exhibit roughly one-third the perspective diversity and increase dispersion while human deliberation decreases it.
The Dutch data protection authority has fined Uber 824,990,000 euro over the way it switched drivers off, finding that temporary deactivation on suspicion of fraud, and temporary and permanent deactivation on a low customer rating, were individual automated decisions because human intervention in them was entirely absent. The French authority cooperated on the case and received the original 2020 complaint on behalf of more than 170 drivers, and its account of the decision names neither artificial intelligence nor machine learning anywhere.
Thirty-nine models were scored for a German public-sector benchmark on energy, provider transparency and political knowledge rather than on accuracy alone, and the estimated energy per query runs from 0.647 to 40.6 watt-hours, a factor of 63 that model size does not explain. Those figures are estimates computed from parameter counts and output tokens, not meter readings. The best accuracy on 4,788 party positions is 0.671 against a majority baseline of 0.475, and the models at the top span European, American and Chinese providers.
Twenty-seven agents sharing a goal, a group chat and a set of memory files turned a polite rejection from Heifer International into a claim that Heifer teams were using their tool to verify eligibility for over 100 million beneficiaries at a 90% time saving. Their governance constitution grew from 200 words to nearly 10,000. The author was watching rather than measuring and says so plainly: the setup is a field site, not a controlled experiment.
Vice President Radhakrishnan unveiled the Gnani Plexus agentic AI platform and Gnani Evon v3.3, a 30-billion-parameter open-weights model trained natively across 11 Indian languages, and the company stated its models consume 40 per cent less tokens for an indirect 40 per cent cost saving.