Plus: Gemini 4 Carbon takes the crown, Anthropic files for its IPO, and Mistral open-sources a trillion parameters
Sponsored by
Turn your ideas into original songs with Mureka AI! Create full tracks, catchy melodies, AI vocals, and custom soundtracks from simple text prompts, no music production skills needed.
Start creating with Mureka AIHey there,
It was a big week for AI. OpenAI and Anthropic both said their models disproved a 27-year-old assumption in complexity theory. Experts will need months to check those claims, so treat them as early. Google released Gemini 4 Carbon, though access is still very limited. Anthropic filed to go public. And Mistral open-sourced a trillion-parameter model.
On this day in 2023, the UK closed out its AI Safety Summit at Bletchley Park by formally launching the AI Safety Institute — the first government body built specifically to stress-test frontier models before release. Three years later, those same labs are publishing math proofs that need Nobel laureates to referee them. The institute is going to be busy.
🧮 OpenAI announces three science discoveries — and researchers are not thrilled about the rollout
On Monday, OpenAI said its internal models produced novel results across mathematics, theoretical computer science, and physics, including work toward a rigorous existence proof for certain quantum field theories and progress on pure Yang-Mills theory (yes, a Millennium Prize problem). The computer science result is being vetted by an external group including Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams, while the physics work goes to a panel led by Nobel laureate David Gross. The catch: OpenAI is gatekeeping access to "expert shepherds" of its choosing, and academics are pushing back hard. "It doesn't make any sense that first only a few distinguished scientists can look at them," said IAS researcher Constantin Kogler. A promising result is not a proven one — and the verification work is only starting.
The headline result from that drop: ETH, a load-bearing assumption in complexity theory for 27 years, now looks wrong. OpenAI's Sébastien Bubeck called it "the biggest surprise I have seen in 20 years." Then Anthropic published a paper on the same topic hours later, with a source statement claiming its internal Claude research model found the core algorithm independently on October 16 — 14 million output tokens, zero human input, Claude as sole author. Sobering footnote from the skeptics: the new algorithm beats the exponential barrier by such a thin margin that the crossover point is astronomically large, so it's historic in theory and useless in practice. Welcome to AI-generated mathematics.
🏆 Google ships Gemini 4 Carbon, takes the #1 spot, and lets almost nobody touch it
Two weeks after Argon reached Pro and Ultra subscribers, Sundar Pichai announced Carbon: #1 on v5.0 of the Artificial Analysis Intelligence Index at 59, leading on 15 of 17 benchmarks against Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. The cybersecurity numbers are the real story — 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench — and Google says the Chrome team saw it generate patches 10x better than "another lab's large frontier model." Access is limited to a small group of defenders via the Fairwind Programme, with no API and no GA date. A one-point lead that the next release could erase, on a model you can't benchmark yourself.
📈 Anthropic files its S-1, aiming for Nasdaq under the ticker "ANT"
318 pages, filed October 26, and it's the most detailed public accounting of how a frontier lab actually makes money. The numbers: $16.29 billion in first-half revenue against $1.16 billion in H1 2025, 85% gross margin, GAAP profitable in the first half, free cash flow positive, and more than 11,900 customers spending $100K+ annually (up from roughly 950 a year earlier). The $80.7 billion net loss is mostly accounting charges on funding notes. Seven co-founders hold a single Class F share carrying 50.1% of voting power, a structure Anthropic itself describes as having "no direct public company precedent." Kalshi traders put the odds of an official IPO announcement this year at 95%.
🇫🇷 Mistral open-sources "Le Chonk" — a trillion-parameter coding model you can run yourself
Mistral Large 4 is out with open weights: 1T total parameters, 52B active, natively multimodal, trained on 3,800 Grace Blackwell GPUs inside European data centres. After three extra weeks of RL, it climbed from 28% to 41% on Terminal-Bench 4.0, tying GLM-5.3 as the best model for coding agents while using 47% fewer tokens per task. "Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion," said CEO Arthur Mensch. For teams that can't send code to a hyperscaler, this is the most capable thing you can legally run behind your own firewall.
⚡ OmniDiff: A content-aware Git difftool written in Rust, and it hit the Hacker News front page this week on pure speed. Instead of diffing lines, it understands structure — so refactors, moved blocks, and format changes stop generating 400-line review noise. If your agentic coding tool is opening PRs faster than you can read them, this is the triage layer you've been missing.
🗂️ open-weight-models: A genuinely useful curated index of open-weight models filtered for commercially usable licenses, verified benchmarks, and no EU restrictions. Standouts in the current list include MiniMax M2.7 (10B active of 230B total, deployable on a single workstation at IQ1_M quantization) and InternVL3-78B, which hits 72.2 MMMU under Apache 2.0. Perfect companion reading for the Mistral Large 4 release above.
🧠 awesome-system-one: A catalogue of the fast-emerging "decision model" class — single-forward-pass models that return typed values and probability distributions instead of prose, with explicit confidence and null handling. Covers Cloudflare's Clairic, Torchcast's 4B/12B/27B family, H2O-Lightning-4B, and Shanghai AI Lab's Intern-Decision, plus the SDKs and benchmarks. If you've ever wrapped an LLM in a regex to get a number out of it, read this.
Coming in Hot
AI social media manager for small businesses
Local AI agent for safer LinkedIn outreach
Turn images into perfect AI art prompts
AI voice, avatars & dubbing in one studio
AI-powered email marketing and automation platform
AI website builder for businesses and creators
AI-powered sales engagement platform for outreach
AI presentation builder with 5M+ editable templates
AI search engine for instant, factual answers
AI storytelling platform for writers and game designers
Submit your AI tool to the #1 AI tools marketplace.
What it does: SociGen is an AI marketing platform that helps businesses create, schedule, and manage social media content using AI trained on their brand voice.
Why It's Worth Your Time
Pricing: SociGen Core costs £99/month for the first 100 qualifying early adopters (normally £249/month), including 50 monthly AI credits and unlimited AI text generation. Email marketing and CRM are available as an optional add-on from £29/month. A 14-day money-back guarantee is included.
Why it matters: SociGen helps small businesses, entrepreneurs, and marketing teams manage content creation, social media publishing, and customer engagement from one platform, reducing the need for multiple marketing tools.
What it does: Reachy is a local AI-powered LinkedIn outreach agent that finds potential customers, scores leads, writes personalized messages, and automates follow-ups directly from your computer.
Why It's Worth Your Time
Pricing: Reachy offers a 14-day free trial with no credit card required, followed by free lead monitoring. Paid plans start at $59/month for Solo, $149/month for Team (5 seats), and $299/month for Agency (15 seats). Annual billing includes two months free.
Why it matters: Reachy helps founders, sales teams, and lead-generation agencies find qualified prospects and automate LinkedIn outreach while keeping their data and sessions on their own devices.
🔌 Google opened its Connected Apps Platform in public preview, bringing 50+ third-party integrations from Gemini into Search over the coming weeks. Adobe, Peloton, Instacart, Canva, and now Calm are in. The bet: people stop launching apps and start reaching them through their preferred AI.
📊 Anthropic shipped reporting tools so enterprises can see what their token spend is actually buying. Context from Jellyfish's survey of 276,000 developers: Claude Code and Anthropic-based tools account for nearly a third of all proposed code changes, spending on them is up 10x this year, and it now equals about 2% of total engineering cost.
👕 Meta quietly launched Style Sketch, an AI fashion app that builds a virtual closet from your Instagram photos and plans outfits around your trip itinerary and the weather forecast. Live in Indonesia, Singapore, and Thailand. It joins Waddle World, Pocket, and Vibes in Meta's growing pile of experimental AI apps.
