Anthropic files S-1 IPO prospectus with $112.4B cash
38 sources · 249 items · ≈22 hours of content · 3 min read
📌 OpenAI Proves Computational Complexity Limits
OpenAI and Anthropic used Lean to refute the Exponential Time Hypothesis, releasing 722 papers on local smoothing and partial Yang‑Mills, while Anthropic’s Claude autonomously generated a 14‑million‑token ETH‑disproof and outperformed human designs in 7 of 8 quantum circuit primitives.
🚀 Google Unveils Genie 4 World Generator
Google DeepMind’s Genie 4 creates real‑time interactive audio worlds from text, with Gemini Omni preserving consistency and Gemini 3.8 Flash guiding multi‑step logic; sessions last two minutes and suffer object drift, currently in research preview for trusted testers and Ultra subscriber waitlist.
📌 Anthropic Files S1 IPO Prospectus
Anthropic’s S‑1 shows 900 % net revenue retention, $112.4 B cash, $16.3 B H1 2026 revenue, $600‑$800 B compute commitments (up to $431 B non‑cancelable) and $19.2 B spend since 2023 (70 % internal R&D), while warning of catastrophic AI risk uninsurability and dependence on rivals Google and Amazon.
🔬 Anthropic Expands Into Biotech Drug Discovery
Anthropic acquired Mithril Medicine and partnered with six biotech firms to accelerate drug discovery, targeting human trials by late 2028.
🤖 OpenAI Launches GPT-6.1 Terra Computer-Use Model
OpenAI’s GPT‑6.1 Terra, priced $1/$5 per M tokens, runs three‑times faster than Sol, achieves 50.3 % coding score on Terminal Bench 4.0, cuts hallucinations to 38 %, and ships with API credits ($200‑$500) for Pro and Business users.
🚀 Google Debuts Gemini 4 Carbon & Argon Models
Google launched Gemini 4 Carbon (top on Artificial Analysis index) and Argon, both able to autonomously discover, validate, and patch software vulnerabilities—Argon uncovered a global healthcare flaw—and lead benchmarks (SEC Bench Pro, Exploit Gym, CWE‑bench 68 %, Terminal‑Bench 77.9 %); priced $2/$10 per M tokens with 95 % cached‑input discount, initially rolled out to trusted cyber defenders and US government via Fairwind.
🤖 Engineers Deploy Secure AI Agent Frameworks
Engineers hardened AI agents by moving safety logic from prompts to code, deploying zero‑trust MCP proxies, execution hooks, validation layers, and preflight harnesses; libstrix delivers up to 31× faster 64 k‑context prefill, and Tadpole shows a 66 MB LoRA‑compressed Qwen3‑4B tool agent.
🏥 Claude Generates Playable Games & 3D Worlds
Claude Opus 5.5 now builds complete games—Tree Appreciator, Emberwick RPG, Bloodbench strategy—while DeepSeek V4.1 Flash achieves 1,286 tps and renders hyper‑realistic 3D assets such as a mechanical watch, a Fall Guys clone, a folding‑phone model, and a PDF reader app.
🤖 Regulators Advance AI Governance & Cybersecurity Laws
Black October exposed PHP flaws affecting ~70 % of sites, breaches at Quest Diagnostics and Railway, and AI‑driven ransomware, prompting FBI/NSA alerts; Altman calls traditional patching obsolete, urging AI defense shields; NY Gov. Hochul signs AI chatbot and baby‑monitor rules, banks move to block unauthorized third‑party AI agents, and Lightwell/Akrites launch coordinated open‑source patching alongside Harvard’s AI safeguard governance effort.
🤖 Legora Deploys AI Contract Review Agents
Legora’s OpenAI‑powered agents cut NDA review time from 20 % to <1 %, slashed a Fortune 100’s contract turnaround by 90 % (saving $20 M and reclaiming 9.5 hrs per lawyer weekly), and can query 10 k contracts in minutes with linked citations for lawyer verification.
📌 Creators Share ComfyUI Tricks & Community Insights
Community tips show character LoRAs can be trained on 4‑15 curated images, expanded with editing models like F2K9B; the Clown Context Suite halves video degradation; ComfyUI‑MOSS‑QwenVL‑Layout converts text to BBOX prompts; Claude Opus 5.5 scripts BBOX prompts for multi‑character LoRAs; MOSS‑Image offers more variance than Krea 2 Turbo, which retains higher photorealism.
📌 Quick Roundup of Smaller Updates
Recent micro‑updates include DeepMind’s Genie 4 launch, DeepSeek V4.1 Flash 1017 self‑feedback learning, Anthropic’s OneGov 40 % Claude discount for US agencies, OpenAI’s reported TPU chip purchase from Google, Stanford’s Next‑Concept Modeling for rapid sketching, Anthropic’s president denouncing AI romantic partners, NVIDIA’s Nemotron 3.5 Transcribe for high‑accuracy transcription, and AssemblyAI’s new lossless‑wav TTS model.
SOURCES
Anthropic · 2 posts
AK · 5 posts
- RT @ruizhaocv: @_akhaliq Here’s a simple demo of Hy-Nomas running in the Space 🤗 https://t.co/NqgZwkrvgn
- RT @hynomas: Thanks for @_akhaliq for setting up the HuggingFace demo space for our Hy-Nomas interactive video world mod
- Hy-Nomas Towards Free-form Interactive World Generation paper: https://t.co/by82MMznL9 app: https://t.co/clt7ItswBD ht
- Representing the Unseen Rethinking Visual Imagination for Spatial Reasoning paper: https://t.co/SgjDTxldcm https://t.c
- ThinkWorld World Imagination Reinforces Chain-of-Dynamics Reasoning paper: https://t.co/DCyVCUs3wa https://t.co/vb94kp
Simon Willison · 1 post
OpenAI · 2 posts
Rowan Cheung · 1 post
Yann LeCun · 1 post
…and 237 more