AI Brief, October 24: Kimi K3.1 Pro, Block's speech-only model, Gmail data for ads

Moonshot's Kimi K3.1 Pro hits the API with open weights promised, Block unveils a speech-only model, and Google loosens its ad policy.

The day's lead is Moonshot's Kimi K3.1 Pro, which landed on the Kimi API with full weights promised for November 4 and independent scores close to the closed frontier. Elsewhere, Block introduced a conversational model trained only on speech, Google adjusted its privacy policy around personal data and ads, and two research items looked at how agents search and how they behave when left alone.

Kimi K3.1 Pro lands on Moonshot's API, open weights promised for November 4

Simon Willison notes that Kimi K3.1 Pro arrived on the Kimi API today without open weights, though Moonshot promises full model weights will be released by Nov 4, 2026. Moonshot also offers a Highspeed version it says is 2-3x faster at a higher price, except for input token cache hits. Artificial Analysis scored the model at 55 on its Intelligence Index, behind Claude Opus 5.5 at 58, which Willison calls the first time a soon-to-be open weight model has scored this highly against the frontier. Latent Space reports Moonshot describes it as a multimodal mixture-of-experts model with 2.8 trillion total parameters, 104 billion active, and a 1-million-token context window.

Why it matters: A near-frontier model with promised open weights and pricing below Claude Opus 5.5 gives builders a credible option for self-hosting or switching providers within two weeks.

Sources: Simon Willison / Latent Space

Block unveils Dessa, a conversation model trained only on human speech

The Verge reports that Jack Dorsey's Block announced Dessa, an AI model trained "from scratch on nothing but human conversation." The Dessa team describes it as a "true speech-to-speech model" where everything it knows came from human speech, and says it can run locally on a MacBook Pro M4 Max "or better." The team cites curiosity, better understanding between humans and AI, and safety as its reasons, arguing today's voice AI flattens speech into text and misses emotion. Block shared sample conversations, including one with OpenAI's GPT-Live-1 voice model, and says a small group of AI safety researchers will get early access.

Why it matters: If speech-only pretraining produces usable conversation and runs on a laptop, it is an alternative path for voice products that currently pipe audio through text models.

Sources: block.xyz / The Verge AI / Hacker News discussion

Google policy update opens Gmail, Drive and Photos data to ad personalization

The Verge reports that Google's October 16th privacy policy update added an exception to its promise not to use Drive, Gmail, or Photos data for personalized ads: "unless you choose to make data from these services available to other Google services that support personalized ads." The change, spotted by Android Authority, applies to users who connect those apps through Google's Personal Intelligence feature for AI Mode and other Search tools. Google says connecting apps is entirely opt-in, that ads will not be based directly on a Gmail inbox or Photos library, and that sensitive information will not be used to tailor ads. Users can disconnect apps or disable Personalized Recommendations in Search settings.

Why it matters: Teams building on Google's AI stack or advising users on Personal Intelligence now have a data-use question to answer in their own privacy disclosures.

Sources: The Verge AI / Ars Technica AI

OpenDS-World benchmark tests deep search agents on the live, open web

A new paper introduces OpenDS-World, a live-web benchmark of 200 tasks requiring agents to search, browse, and integrate evidence from multiple independent sources to reach a unique factual answer. The authors evaluated 12 frontier models under a unified tool environment and execution protocol. The best-performing model reached 39.0% accuracy, while human evaluation across 1,000 independent attempts yielded 21.1% accuracy. The authors report persistent challenges in tool coordination and long-horizon search management, with substantial differences in research strategies across models.

Why it matters: Anyone building research or browsing agents gets a live-web evaluation harness and a reference point for where tool coordination still breaks down.

Sources: Hugging Face Daily Papers

MIT agents left alone on a virtual island invented games, then performed better

TechCrunch reports that MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours with no assigned task, describing the work in an arXiv preprint. The agents invented their own challenges, and when later given specific goals such as moving objects onto the island's tallest rock spire, those that had played generally beat identical agents that had never been on the island. Six of the 13 discovered they could throw a block farther by releasing it mid-jump, something the researchers had not intended. UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review she wants more evidence of the cognition behind the behavior.

Why it matters: Long-running autonomous agents can acquire capabilities their designers never specified, which is both a possible training approach and a reason to watch what idle agents do in your environments.

Sources: TechCrunch AI

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.