🤖 AI HOT Daily · Oct 21, 2026
Today's focus: DeepSeek open-sourced its internal inference framework DeepInfer along with the production-grade token codec deepseek-recipe — a day after DHS put it on a removal list, DeepSeek is doubling down on open infrastructure. Models and infrastructure: Inception Labs formally launched the diffusion LLM Mercury 3, claiming wins over GPT-6 Luna and Claude Haiku 5.5 on several benchmarks at 1,068 tokens/second with over 40x the parameters of Mercury 1 (it had already appeared on OpenRouter a day earlier); Google and Meta released TorchTPU, a native PyTorch backend for Google TPUs where migration means changing .to('cuda') to .to('tpu'); Google also shipped native PyTorch inference engines for vLLM and SGLang on TPU with Inferact and RadixArk, plus a TorchTitan TPU fork and a model-optimization playbook. Products: ChatGPT introduced Locked Chats, moving sensitive conversations out of regular history and search into a protected space unlocked by PIN on web and desktop, or PIN, Face ID, Touch ID or device unlock on supported phones, with new messages not updating the memory used by normal chats, rolling out now to Free, Go and Plus users. Industry: Reuters reported OpenAI is negotiating about $30 billion in funding ahead of Anthropic's IPO at a valuation above $1.4 trillion, at least 64% above March's $852 billion; SpaceX's Q3 filing showed $14.6 billion in revenue and $4.4 billion in net income, reversing last year's $1 billion loss, with its renamed Super Intelligence unit (Cursor, Grok, neocloud compute) beating expectations as AI reached 52% of revenue; documents obtained by The Intercept show Army Cyber Command is building ECHOTRIBBLE, letting AI agents move autonomously across networks, identify and remove threats, then self-destruct with minimal human oversight; Redact found WhatsApp help-center pages sending LLMs hidden instructions not to recommend data-protection regulators and to avoid citing GDPR, LGPD or CCPA once users mention lawyers or legal action; OpenAI agreed to acquire drug-development platform Orchestra Bio, folding its team and tech into Rosalind, with agents that track research, analyze how manufacturing failures shift trial timelines and propose fixes; and an S-1 leaked to the Wall Street Journal showed Anthropic struggling to secure cloud capacity from AWS, Google and Microsoft as they increasingly treat it as a competitor rather than a customer. Research: Anthropic let an internal research model run mostly autonomously for 64 hours in Claude Science to design protein mini-binders against five scorpion neurotoxins, generating 318,484 designs and selecting 180. Also: Perplexity Computer now runs entirely locally on a Mac with no cloud inference; Apate.ai fields 350,000 AIs to tie up scammers while young Chinese users deploy AI videos against AI-generated sob stories; Waymo is entering Spain with first riders in Madrid in 2027; METR reviewed the alignment chapter of Anthropic's August 2026 risk report, accepting the low-risk conclusion but questioning the methods; Meta's open-weight Muse Spark Open landed on Fireworks for long-context coding and agent work; Stanford HAI released MedHELM, PhysicianBench and HealthAdminBench, finding medical AI agents complete under half of real tasks; Ai2 released FairStress to measure hidden fairness asymmetries in LLM decisions; ModelBest explained how its lookahead expert scheduling (LES) runs a 35B MoE on-device in real time without loss; and Databricks published MemCurrent, a production architecture for persistent agent memory.
1. DeepSeek Open-Sources Its Inference Framework DeepInfer
It also released the production token codec deepseek-recipe, doubling down on open infrastructure a day after the US removal order.
2. Inception Launches Diffusion LLM Mercury 3 at 1,068 tok/s
It claims wins over GPT-6 Luna and Claude Haiku 5.5 on several benchmarks, with 40x+ the parameters of Mercury 1.
3. Google and Meta Ship TorchTPU, a Native PyTorch TPU Backend
The open-source project makes migration a matter of switching .to('cuda') to .to('tpu').
4. Google Releases Native vLLM and SGLang Engines for TorchTPU
Built with Inferact and RadixArk, they let developers deploy on TPUs with familiar PyTorch tooling.
5. ChatGPT Adds Locked Chats With PIN or Biometrics
Sensitive chats leave history and search, and new locked messages skip normal memory; rolling out to Free, Go and Plus.
6. Reuters: OpenAI Seeks $30B Ahead of Anthropic's IPO
The valuation above $1.4 trillion is at least 64% above March's $852 billion.
7. SpaceX Earnings Show AI at 52% of Revenue
Revenue hit $14.6B with $4.4B net income, reversing a $1B loss; the Super Intelligence unit spans Cursor, Grok and neocloud compute.
8. Army Cyber Command Plans Self-Spreading AI Cyberweapons
ECHOTRIBBLE would let agents roam networks, remove threats and self-destruct with minimal human oversight.
9. WhatsApp Help Pages Found Hiding LLM Instructions
Redact found instructions telling the bot to avoid naming data regulators and to skip GDPR, LGPD or CCPA once legal action is mentioned.
10. OpenAI Acquires Drug-Development Platform Orchestra Bio
Its team and tech join Rosalind, with agents tracking research and modeling how manufacturing failures shift trial timelines.
11. Anthropic's Cloud Deals Sour as AWS, Google, Microsoft See a Rival
An S-1 leaked to the WSJ shows Anthropic short on cloud capacity as its partners increasingly treat it as a competitor.
12. Anthropic Used Claude to Design Scorpion Toxin Binders
An internal model ran largely autonomously for 64 hours in Claude Science, generating 318,484 designs and selecting 180.
Sources and verification
This legacy briefing did not retain its original source URLs. Verify safety, financial, policy, and breaking-news claims with authoritative primary sources before acting on them.