💰 @Meta just lined up to $95B in equity for the AI buildout. CNBC tallies underwritten offerings plus a planned at-the-market program as up to $95 billion to fund AI infrastructure and "personal superintelligence." - Near-term: 250M Series A depositary shares (mandatory convertible preferred) plus $22.5B of Class A common via forward sales; CNBC frames the initial underwritten raise around $35B - Later: an ATM for up to $60B of Class C, not expected until Class C trading in 1Q 2027 - NVIDIA, AMD and Broadcom indicated non-binding interest in up to an aggregate $5.0B of the Class A common; Meta guided full-year capex to $135B–$145B Same day as Microsoft's voluntary exit offer, another hyperscaler is tapping public markets at AI-factory scale. Stress-test your 2027 inference budget against buyers who are also raising tens of billions in fresh equity. Carlo Edoardo Ferraris (@carloAI) of Natura Umana and Apheth D Almeida of Lucen Robotics join us live today at 9:45 and 10:10 AM PT.

🌿 @mirendil just put 10M tokens on the board with Fern 0.1. Its first model, pretrained from scratch, opens early access with a context window of up to 10M tokens, about 23 MB of code or ~0.5M lines. - Aimed at high-volume agentic R&D work: audit training data, debug training logs, read full agent traces - Mirendil says Mirendil-Base was built by one researcher in 3.5 months with less than $10M in compute, agents writing over 99% of the code, public data only - Early vendor scores include SWE-Bench Verified 73.4, but Terminal-Bench 4.0 only 6.1 versus Opus 5.5 at 66.4 and Astra at 57.9 on the same table Context length is a product category again, and this is not your daily coding agent yet. If you live in training logs and monorepos, request early access; everyone else should wait for the next Fern versions.

🔬 @OpenAI just posted three AI-produced science results, with outside shepherds. On Oct 26 it said internal models produced discoveries in harmonic analysis, complexity theory and quantum field theory, framed as raw research data, not finished papers. - Math: a claimed proof of Sogge's local smoothing conjecture in every spatial dimension at least two, with a Lean writeup on GitHub that OpenAI says also resolves Kakeya and Stein restriction in all dimensions - CS: a claimed refutation of the Exponential Time Hypothesis, under review by a group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams - Physics: a claimed existence proof for interacting QFTs including QCD, with materials held for now by a group led by David Gross Yesterday Anthropic filed an S-1. Today OpenAI is competing on scientific credibility, not just chat share. If you sell AI for science, your bar just moved from demos to Lean-checked claims and named outside reviewers.

💻 @OpenAI just shipped GPT-6.1 Terra for everyday computer use. A fast mid-tier model for visual reasoning, precise computer use and clearer writing, live in ChatGPT Work and Codex and on the API as gpt-6.1-terra at $1 / $5 per million tokens. - Standard API price is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, with cached input at $0.05 per million - On Agents' Last Exam, medium effort scores 15 points higher than GPT-5.6 Terra at about one-third the cost, and matches GPT-6 Astra at one-tenth the price per token while finishing over 3× faster than Astra and Sol on average - Terminal-Bench 4.0 at max effort hits 50.3%, more than double GPT-5.6 Terra's 21.5%; OpenAI still points the hardest coding to Sol and Astra This is the make-the-workday-agent-cheap bet, not the pay-for-top-coding bet. Point one high-volume computer-use workflow at gpt-6.1-terra this week and keep the stack that finishes the task.

🚪 @Microsoft just opened voluntary exits to around 60,000 people. TechCrunch reports Microsoft confirmed the eligible pool is around 60,000, about 27% of its 223,000-person workforce. That is eligibility, not a confirmed exit count, and Microsoft's own blog prints no headcount. - The Voluntary Exit Offer covers eligible employees in participating orgs in Canada, India, the UK and the US at level 70 and below - Employees must decide in early December and leave Dec. 31 if they take it; TechCrunch says the package is similar to Microsoft's standard severance - Bloomberg via TechCrunch names Xbox, Windows, Copilot and Finance among participating divisions; data centers, sales and marketing are among those excluded One hyperscaler is asking ~60k people if they want a package while AI capex stays hot. Do not confuse eligible with exiting. If you sell into Microsoft or hire ex-MSFT talent, map a backup champion before January.

dario as soon as the S-1 hits

oh no!

🐕 A startup with no dogs yet just raised $200M to engineer military working dogs. Warhound, conceived by @BenLamm of @colossal and @PalmerLuckey of @anduriltech, came out of stealth on Oct 22 at a $1.2B valuation, per @TechCrunch. - It plans to clone and gene-edit proven working dogs for the US military, law enforcement and emergency responders, with first puppies expected in early 2027 - Anduril will supply gear "on top of the dog, not in the dog," such as vests with data sensors for handlers - Backers include In-Q-Tel, @sequoia, @ARKInvest, Franklin Templeton, @8vc and @Lux_Capital, with no lead investor named and no price per dog announced TechCrunch says Colossal has been using AI for gene cloning and gene identification, so this is AI-assisted biology and hardware, not just a dog story. Read it next to Isaac Lab: NVIDIA is building simulation so robots can learn, while Lamm told Fortune no machine today matches a dog's nose or its judgment. Look for the broken physical supply chain in your industry that nobody has modeled yet.

💬 A US lab just trained an AI to know when to stay quiet in your group chat. @humansand released Ampersand v0.1 on Oct 20, a "collaborator model" built for groups instead of one-on-one chat, and two days later launched Boba, a free assistant powered by it. - After every message the model chooses between two tools, send_message and pass, and it's trained with RL on long simulated group chats - It's trained from @Zai_org's open-weight GLM 5.3, with API pricing at $1 per million input tokens and $5.20 per million output - Every benchmark is vendor-reported and based on simulated users, and it scores slightly below GLM 5.3 on SWE-bench Pro v2 (82.4% vs 84.3%) Almost every assistant is trained to answer when spoken to, and this one is trained on when to pass, which is the difference between a teammate and a noisy bot. Give your agent an explicit way to stay quiet, then add Boba to one low-stakes group chat and note when it speaks up. Alex Newman (@thedotmack), founder of Claude-Mem, joins us live today at 10:10 AM PT, and memory is exactly what I want to ask him about.

🤖 NVIDIA says its new robot trainer is about 3x faster, and it's still open source. @NVIDIARobotics announced Isaac Lab 3.0 on Oct 22, its open-source framework for training and testing robots in simulation, with the GA tag on GitHub on Oct 19. - NVIDIA's demo reel says it runs about 3x faster than Isaac Lab 2.3, and in its own chart a vision-based Kuka Allegro lift task on 8 L40S GPUs hit 521.5K environment steps a second versus 171.8K - One task can now run on Newton, OVPhysX or Isaac Sim PhysX for physics, and Isaac Sim is now optional - It adds LEAPP for moving trained policies toward real robots NVIDIA built the picks and shovels for language models, and by keeping this one open it's trying to become the default for physical AI. It's a major release with breaking changes, including a new quaternion order and Python 3.12, so read the 2.x migration guide first. Run one task on the Isaac Lab 3.0 quickstart this week and compare it to your current setup.