Desk
Broad artificial intelligence coverage.
Luca Guadagnino’s OpenAI comedy-drama Artificial premiered at the New York Film Festival after AmazonMGM dropped it following Amazon’s $50 billion OpenAI deal. WIRED’s review says the film portrays AI’s leaders as vain and reckless, and frames danger warnings as part of the sales pitch.
EmbeddingGemma 2 brings search across photos, video, audio and code onto a phone under a permissive license, which is great for privacy and a little unsettling for the same reason.
Brett Adcock’s startup Hark has widely released Hark Pro, a free full-screen assistant built on a model trained for computer use. It pitches itself as an ad-free alternative to Meta and OpenAI, but the privacy case so far rests on stated intent rather than published technical detail.
Look past the trillion-parameter headline: Le Chonk's sharpest claim is that it will do defensive security work closed models refuse, and that claim carries both promise and risk.
WIRED reports OpenAI told people it plans to post hundreds of results on long-standing math problems to GitHub, after about 40 mathematicians asked in August for proper papers. OpenAI says a new model has resolved more than 100 open problems and that no release time is set.
San Francisco startup Mirror Particle says it is building a model from scratch to track why people act as they do and how that changes, rejecting LLM persona role-play. It targets market research first, with a long-term aim of individual-level prediction that raises consent and persuasion questions.
On the Microsoft Research Podcast, Jennifer Neville described findings that models perform significantly worse when a task is spread across several clarifying turns, and that repeated document edits accumulate small errors that subtly erode meaning, failures single-turn benchmarks rarely catch.
Atlassian and OpenAI expanded their partnership so OpenAI models including GPT-6 Astra and the GPT-5.6 series power agents across Atlassian’s platform and Rovo, grounded in its Teamwork Graph. The firms are exploring Jira features to assign work to agents, raising permission and accountability questions.
Anthropic expanded Claude for Startups: qualifying companies get a free year of Claude Team with up to five premium seats, $1,000 in API credits, Claude Marketplace access and office hours with its Applied AI team. It lowers costs for founders and is also a bid to become their default vendor.
NVIDIA's green contexts let one program carve a GPU into dedicated lanes, and the latency win for urgent work is big enough that inference and robotics teams should take notice.
Mistral released its trillion-parameter Large 4 through a guardrailed endpoint only, planning to publish weights in about three weeks after safety testing and work with partners and governments meanwhile. It says training used 4,000 Nvidia GPUs; benchmark results were still pending.
Google is limiting free Gemini users to Flash Lite from October 9 and removing Pro from its $4.99 AI Plus plan. Pro and Deep Think will require AI Pro at $19.99 or Ultra at $99.99, a sign that giving away top models to win users is ending.