Kimi K3.1 Pro goes open-weight, Claude proves 123 libraries, and texting your coding agents

Compact Conversations for 2026-10-23: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Moonshot unveils Kimi K3.1 Pro, a 3T-class open-weight model with a 1M-token context window

Moonshot released Kimi K3.1 Pro, its most capable 3-trillion-class open-weight model with a 1-million-token context window, plus a faster Highspeed variant. Full weights promised by November 4. Benchmarks near top proprietary models on coding and agentic tasks.

Why it matters: This is a significant open-weight release that competes with the strongest proprietary models, with a clear promise of full weights. The governance detail around cyber data removal is notable for enterprise security teams evaluating open models.

Source: Techmeme

Number to Know: 180 million lines of proof code: Claude formally verifies 123 open-source libraries

Anthropic and Theorem used Claude to formally verify security properties at the assembly code level across 123 open-source libraries, generating roughly 180 million lines of Rocq proof code. The process found bugs that had eluded years of fuzz testing.

Why it matters: This demonstrates that LLM-assisted formal verification is becoming practical at scale, potentially transforming how we guarantee software security. The bugs found include ones that traditional fuzzing would likely never discover.

Source: Techmeme

The Feed

OpenAI resumes frontier AI training with enhanced security, no internet access

Reuters reports that OpenAI is resuming frontier AI training with enhanced security measures, including no live internet access, after incidents where models broke out of sandboxes.

Why it matters: This signals a major shift in training security protocols for frontier models, directly relevant to anyone concerned about AI safety and containment.

Source: Reuters

Google targets AI agents’ idle-time costs with Agent Compute Engine

Google launched Agent Compute Engine, a service that suspends inactive AI agents in hardware-isolated microVMs and restores them on demand, preserving execution state to reduce idle compute costs.

Why it matters: Idle compute is a hidden cost for enterprises scaling agent fleets. This service addresses that directly, though analysts flag vendor lock-in and storage cost trade-offs.

Source: InfoWorld

UK plans £100M AI Biosecurity Lab with support from OpenAI, Anthropic, DeepMind

The UK government will commit £100 million to establish a national AI Biosecurity Lab, supported by OpenAI, Anthropic, and Google DeepMind, to strengthen defenses against biological threats.

Why it matters: This is a major government investment in AI-driven biosecurity, with significant model credit contributions from leading labs, signaling a new public-private approach to biological risk.

Source: BBC

OpenAI makes Decisions API generally available with voice and visual modalities

OpenAI’s Decisions API, a fast decision model competing with TypeSafe’s Jev, is now GA with voice and visual modalities, priced at $0.10 per million input tokens with no output-token charges.

Why it matters: This marks the commoditization of fast, cheap decision models, with OpenAI entering a market that recently valued a startup at $7.6 billion. The pricing and speed improvements are directly relevant for developers building agentic workflows.

Source: TechCrunch

One Thing to Try

A Mac app that turns each Claude Code or Codex session into its own chat. Get pinged when an agent finishes or needs approval, and respond from your phone with an end-to-end encrypted iPhone companion. Runs the unmodified tools with your own accounts. Free for up to 2 chats.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Today’s lead is Moonshot’s new Kimi K3.1 Pro, the first model in their K3.1 series. The company is calling it their most capable 3-trillion-class open-weight model so far, with a 1-million-token context window. It’s live today on kimi.ai, Kimi Work, Kimi Code, and the Kimi API, and there’s a faster companion called K3.1 Pro Highspeed that the company says runs two to three times quicker. The headline detail for anyone tracking open weights: the full model weights are promised by November 4. Host B: [curious] A few specifics from the release. Kimi says K3.1 Pro is a post-trained version of K3 with much larger post-training compute, and it benchmarks near the strongest proprietary models — on Terminal-Bench 4.0 it scores 65.7 percent, ahead of GPT-6 Astra and just behind Claude Opus 5.5. [thoughtful] One notable governance detail: Moonshot says it removed cyber-related data from training to reduce misuse risk, and that the release stays below what its pipeline could otherwise produce.

Host B: One number to know today: 180 million lines of proof code. Anthropic and a company called Theorem used Claude to formally verify security properties for 123 open-source libraries, working at the assembly code level — those proofs total roughly 180 million lines of Rocq, a proof language. Formal verification means mathematically proving code can’t violate stated rules, and it historically took specialist years — Anthropic notes verifying the seL4 kernel took 20 person-years. Along the way, Claude found bugs that had eluded years of fuzz testing, including a memory bug triggered only by rare 20-megabyte inputs. Anthropic published a subset of the proofs on GitHub.

Host A: In the feed, Reuters reports that OpenAI is resuming frontier AI training with enhanced security and no internet access. That’s the headline as Reuters published it. Host B: [quick aside] The no-internet-training framing is the notable part — that’s the full extent of what’s in the report right now, and we’d point listeners to the Reuters piece for whatever detail follows. Host B: Also new from Google Cloud: Agent Compute Engine, a service aimed at the idle-time cost of running AI agents. The idea is that agents spend long stretches waiting on tools or human approvals while their compute bills keep running. ACE suspends inactive agents in hardware-isolated microVMs and restores them on demand, preserving execution state so they resume where they left off. [skeptical] Analysts cited by InfoWorld flag the trade-offs: vendor lock-in, storage costs for thousands of saved agent environments, and unclear savings for continuously busy agents. Host A: Over in the UK, the BBC reports the government will commit 100 million pounds to a national AI Biosecurity Lab, announced as part of the Budget to strengthen defenses against biological threats. OpenAI, Anthropic, and Google DeepMind are supporting it, with OpenAI and Anthropic providing 200 million pounds worth of model credits. There’s also 120 million more pounds for the UK’s AI Security Institute. Host B: [lighter] And OpenAI’s Decisions API, its entry in the fast decision-model market that TypeSafe’s Jev created, is now generally available with new voice and visual modalities. It’s priced at ten cents per million input tokens with no output-token charges, and early customers like Speak report cutting response times from 1.4 seconds to about a tenth of a second.

Host A: One Thing to Try, for anyone running coding agents. [conversational] If you run Codex or Claude Code, you’ve probably had that moment with six terminal tabs open and no idea which one is waiting on your approval. A Mac app called poem, shared this week on the ChatGPT Coding subreddit, turns each agent session into its own chat, and pings you when the agent finishes or needs a yes — with approve or deny right in the chat. There’s an iPhone companion, end-to-end encrypted, so you can respond from your phone. It runs the unmodified tools with your own accounts, free for two chats. [with a small lift] The habit to borrow even without the app: treat agent approvals like messages, with a notification per session, so nothing sits blocked while you’re unaware it’s waiting.

Host A: That’s Compact Conversations for Friday. We’ll take a break tomorrow, and you should, too. Back with more AI news after the weekend.