Open-sourcing a production inference stack hits self-host cost and throughput; top engagement of the day.
AI NEWS
Past AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
This day
October 22, 2026
19 items for 2026-10-22, ranked by combined engagement. Unverified items are marked.
Top 3
Anthropic opens Claude Duet research preview: a proactive, cloud-based agent app built on Opus 5.5
Shifts assistants from reactive chat to proactive agents; massive reach.
OpenAI's Patch the Planet: GPT-5.6 Sol helped find critical PHP flaws; fixes now available
Affects a huge share of the web; AI-assisted security discovery story.
All
DeepSeek open-sources DeepInfer, its production inference framework, plus API-aligned deepseek-recipe
Open-sourcing a production inference stack hits self-host cost and throughput; top engagement of the day.
CollapseDeepSeek open-sources DeepInfer, its production inference framework, plus API-aligned deepseek-recipe
DeepSeek open-sourced DeepInfer (Apache 2.0), the in-house inference framework behind DeepSeek-V4.1-Flash in production. It bundles optimized kernels (FlashMLA, DeepGEMM, DeepEP), attention data parallelism, MoE expert parallelism, speculative decoding, SWA bounded replay and structured outputs, and runs on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. This release covers offline inference; a KV-cache system and prefill/decode disaggregation are not included. DeepSeek also released deepseek-recipe, a token encoding/decoding layer matching the official API so self-hosted deployments behave the same. Future models' inference code will land in the repo too.
Anthropic opens Claude Duet research preview: a proactive, cloud-based agent app built on Opus 5.5
Shifts assistants from reactive chat to proactive agents; massive reach.
CollapseAnthropic opens Claude Duet research preview: a proactive, cloud-based agent app built on Opus 5.5
Claude Duet is a standalone agent app (iOS, Android, web) built on Claude Opus 5.5 for professionals and small-business owners. It sends a morning brief, periodically checks your email, calendar and other apps, and pings you only when something needs you. With its own cloud computer and browser, it keeps working when your phone is off and can use 6,000+ connectors. It waits for approval before sending, paying, signing or deleting, never trades securities or moves money, and never sees your passwords. Waitlist only for US Pro/Max users at no extra cost; Anthropic says personal errands like booking tickets are still weak.
OpenAI's Patch the Planet: GPT-5.6 Sol helped find critical PHP flaws; fixes now available
Affects a huge share of the web; AI-assisted security discovery story.
CollapseOpenAI's Patch the Planet: GPT-5.6 Sol helped find critical PHP flaws; fixes now available
OpenAI says security firm Calif found critical PHP vulnerabilities using GPT-5.6 Sol under its Patch the Planet initiative, with Wiz independently reporting both flaws (OpenAI's checker calls them php2shell). PHP runs an estimated 70% of websites. Over two months OpenAI coordinated PHP maintainers, researchers and vendors such as Wiz and Cloudflare to prepare fixes before disclosure; the fix needed a major PHP engine change. Sites can now upgrade safely. OpenAI calls it one of the most impactful vulnerabilities in recent memory and an example of AI finding bugs before attackers. It also proposed a five-stage response framework and committed $1B in subsidized Daybreak access for under-resourced defenders.
OpenAI releases php2shell-checker, a Codex skill to check whether your PHP version needs patching
Turns the vulnerability alert into an actionable self-check tool.
CollapseOpenAI releases php2shell-checker, a Codex skill to check whether your PHP version needs patching
For the PHP flaws above, OpenAI open-sourced php2shell-checker (Apache 2.0) on GitHub with two Codex skills: phpcheck-url when you only have a URL, and phpcheck-server when you have server access; you can also run node phpcheck.js directly. It reads the version from HTTP headers (X-Powered-By, Server) or php -v and compares it against each branch's patched release, returning OK, VULNERABLE or UNKNOWN. The README stresses it is only a version comparison, not proof of a fix: headers can be hidden or altered, distro backports may not change version numbers, and UNKNOWN does not mean patched. Check against PHP's official security release.
Anthropic: Claude generated 300k+ protein designs in 64 hours; 17 lab-confirmed binders to scorpion neurotoxins
Concrete AI-for-science example with lab follow-through.
CollapseAnthropic: Claude generated 300k+ protein designs in 64 hours; 17 lab-confirmed binders to scorpion neurotoxins
Scorpions sting about 1.2 million people a year, killing over 3,250, and antivenoms are costly or scarce. With Citrini Research, Anthropic had Claude design binder proteins against five neurotoxins from scorpions in four regions. In 64 hours Claude generated 318,484 designs; 211 were lab-tested and 17 met the criteria, binding toxins from Mexican, North African and Middle Eastern species, the best in the single-digit nanomolar range. Anthropic stresses this is not yet an antivenom—no cell or animal tests—and has released all 211 designs and data on Hugging Face.
OpenAI acquires drug-development coordination platform Orchestra Bio to join its Rosalind life-sciences line
Clear industry signal: OpenAI deepening life-sciences ops tooling.
CollapseOpenAI acquires drug-development coordination platform Orchestra Bio to join its Rosalind life-sciences line
OpenAI has acquired Orchestra Bio, a platform where biopharma teams and AI agents jointly run drug programs: scientists and program leads delegate study tracking to agents, which, when something goes wrong—say a failed manufacturing batch that could delay a clinical trial—identify affected activities and timelines and compare responses. Orchestra's team joins OpenAI's Rosalind line, combining with the GPT-Rosalind model and Rosalind Workbench to connect scientific findings with the studies and decisions that follow. OpenAI notes a new US drug takes 10–15 years from target discovery to approval, and aims to get treatments to patients sooner.
Aleksander Madry returns to OpenAI to work on superintelligence safety; Sam Altman thanks him
High-signal OpenAI safety leadership move, endorsed by Altman.
CollapseAleksander Madry returns to OpenAI to work on superintelligence safety; Sam Altman thanks him
MIT professor (on leave) Aleksander Madry left OpenAI a few months ago, believing the priority then was making AI more useful. He now says AI has entered a different regime in which getting the safety and security of increasingly capable systems right must be central, so his departure turned out to be more of a sabbatical: he is returning to OpenAI to work with Jakub Pachocki on these challenges. Sam Altman said OpenAI had asked him to come back, calling the safety and security challenges of confronting superintelligence immense and Madry well suited to them.
OpenAI: thousands of GPT-Rosalind agents analyzed millions of brain cells in 32 minutes, surfacing neurodegeneration leads
Continues OpenAI’s life-sciences AI narrative with a concrete lead.
CollapseOpenAI: thousands of GPT-Rosalind agents analyzed millions of brain cells in 32 minutes, surfacing neurodegeneration leads
With PerturbAI and Scripps Research, OpenAI ran thousands of GPT-Rosalind agents in parallel over a whole-brain mouse CRISPR atlas: each agent studied one gene disruption across millions of cells and searched the literature, finishing in 32 minutes work the team estimated would take humans months to a year. One lead: disrupting DCTN1, linked to the Parkinson-like Perry syndrome, raised cholesterol-synthesis genes, suggesting cholesterol may not reach where neurons need it. Another: TSC2 and TSC1 affect cerebellar neurons very differently, relevant to tuberous sclerosis. Blinded experts preferred distributed-agent reports 75% of the time. The reasoning atlas is public; the hypotheses still need lab testing.
Recursive releases Flint, an open-weight local coding model hitting 125+ tokens/sec on a MacBook Pro
Combines local fast coding with an automated-AI-research story.
CollapseRecursive releases Flint, an open-weight local coding model hitting 125+ tokens/sec on a MacBook Pro
Recursive, founded by Richard Socher, released its first model, Flint: a local coding model built from scratch in under four months with its automated research system, the Eureka Machine. Per OpenRouter, it is a mixture-of-experts model with ~40B total and ~2.5B active parameters and a 262K context, aimed at exploring codebases, fixing bugs, implementing features and writing tests, with tool calling. Recursive claims up to 2x the speed of comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark; Socher said 48GB is enough to run it. Open weights are on Hugging Face, with free trials on opencode, pi and OpenRouter.
Logan Kilpatrick: Gemini 4 Argon processed ~25 trillion tokens in its first 72 hours
First-party usage scale signal right after a major model launch.
CollapseLogan Kilpatrick: Gemini 4 Argon processed ~25 trillion tokens in its first 72 hours
Google's Logan Kilpatrick said that in the first 72 hours after Gemini 4 Argon launched, Google processed roughly 25 trillion tokens, adding that demand for intelligence is incredible and still underappreciated. It is a first-hand usage figure from Google for the new model's opening days; the post gives no breakdown of where the usage came from.
DeepMind demo: Gemini 4 Argon builds a 3D periodic table synced to Tom Lehrer's 'The Elements'
Viral official demo of multimodal/interactive generation.
CollapseDeepMind demo: Gemini 4 Argon builds a 3D periodic table synced to Tom Lehrer's 'The Elements'
Google DeepMind shared a demo in which Gemini 4 Argon, working in Google's agentic coding tool Antigravity, built a 3D periodic table synced to Tom Lehrer's song 'The Elements', which lists the element names. As the song plays, the app animates orbiting electrons and shows atomic data and light spectra for each element—an example of the model generating interactive 3D web apps.
Dario Amodei's long reply to AI-safety critics: admits some mistakes, defends 'race to the top'
CEO-level safety-governance signal with strong discussion.
CollapseDario Amodei's long reply to AI-safety critics: admits some mistakes, defends 'race to the top'
Encode AI's Nathan Calvin criticized Anthropic over nine actions—an in-house chip team, $500B+ in compute deals, loosened scaling commitments and more—arguing it puts winning the superintelligence race ahead of safety. Dario welcomed the scrutiny and admitted several communication missteps, but rejected the thrust: the 'race to the top' requires being at the frontier. He cited costly moves from the past year: backing California's SB53, strict bio safeguards, delaying Mythos-class models for four months until cyber safeguards were ready, restricting sales to PRC-controlled firms, and putting 5% of the company into a public benefit fund.
OpenAI models and Codex are coming to Google Cloud's Model Garden, purchasable against GCP commitments
Expands enterprise distribution across cloud marketplaces.
CollapseOpenAI models and Codex are coming to Google Cloud's Model Garden, purchasable against GCP commitments
OpenAI and Google Cloud are bringing OpenAI's frontier models and Codex agentic workflows into Google Cloud's Model Garden, so enterprises can use them inside their existing Google Cloud environment with its access controls and data governance—aimed especially at regulated industries. Customers buy through Google Cloud Marketplace, and companies with Google Cloud spend commitments can draw them down. OpenAI models will also come to Google Antigravity and the Gemini Agent. General availability is planned in the coming weeks; OpenAI notes Microsoft remains its primary cloud partner.
Anthropic publishes its 2026 Public Benefit Report, reviewing progress on five mission pillars
Governance/mission disclosure to read alongside product moves.
CollapseAnthropic publishes its 2026 Public Benefit Report, reviewing progress on five mission pillars
Anthropic, a Public Benefit Corporation legally required to balance shareholders' financial interests with the public good, published its 2026 Public Benefit Report. It reviews actions and remaining work under five pillars: solving AI safety's technical problems; making safety a universal practice backed by law; using AI to strengthen democracy and global stability; expanding prosperity and human flourishing; and sharing AI's benefits with everyone. The report page collects related research, policies and programs, which Anthropic says it will keep adding to.
DeepMind demo: Gemini 4 Argon combines three open datasets into an interactive 3D human-anatomy viewer
Another strong official capability demo in the same launch window.
CollapseDeepMind demo: Gemini 4 Argon combines three open datasets into an interactive 3D human-anatomy viewer
Google DeepMind shared another Gemini 4 Argon demo: the model combined three open-source datasets into an interactive 3D visualization of human anatomy. Users can toggle body layers while they move, isolate individual muscles, or zoom inside a beating heart. It follows the earlier 3D periodic-table demo as another example of the model generating rich interactive visualizations.
Simon Willison vibe-codes browser RTS 'Battle Drones' mainly with Claude Opus 5.5, with audio-link multiplayer
Practitioner build that shows coding-agent ceiling in public.
CollapseSimon Willison vibe-codes browser RTS 'Battle Drones' mainly with Claude Opus 5.5, with audio-link multiplayer
Simon Willison built Battle Drones, a Command & Conquer-style RTS that runs in mobile and desktop browsers. Its 12 major versions and 321 commits came mainly from Claude Code, and he reviewed almost none of the code. It has 3D graphics, multiple maps, a giant mech and an LLM-vs-LLM mode, with models and music all generated in code. Nearby devices can play multiplayer over sound via mic and speaker; online play was added later. The prompt 'Make it a LOT more fun' led Claude to run 1,000+ simulated games to rebalance it—with surprisingly good results.
Guizang (op7418) uses AI to build 'guizang TV', a Xiaomi TV launcher with only frequent apps and wallpaper rotation
Popular CN-circle example of practical AI-assisted product tinkering.
CollapseGuizang (op7418) uses AI to build 'guizang TV', a Xiaomi TV launcher with only frequent apps and wallpaper rotation
Chinese AI blogger Guizang (op7418) used AI to repackage a new TV launcher, 'guizang TV', based on two popular projects, CleanFrame and cleanyourtv. It shows only the entries he uses, such as PS5 and Apple TV inputs, Bilibili and screen casting, and he redesigned the UI with wallpaper rotation and built-in wallpapers. He invited Xiaomi TV owners who like these features to try it—an example of AI helping ordinary users customize everyday devices.
Baoyu (dotey) compares four proactive assistants: Grok Bot, Muse, Dots and Claude Duet
CN framing that places Claude Duet in a cross-lab product wave.
CollapseBaoyu (dotey) compares four proactive assistants: Grok Bot, Muse, Dots and Claude Duet
Baoyu (dotey) notes four companies launched proactive assistants in about two months: xAI's Grok Bot (Aug 11), Meta's Muse (Sep 8), OpenAI's Dots (Sep 29) and Anthropic's Claude Duet. All have their own cloud computer and browser, report proactively and mostly ask before actions like payments. They differ by audience: Duet targets professionals and owners, US only; Grok Bot lets several bots split work and is included in X's $30/month Premium tier; Muse is mostly free and works in WhatsApp; Dots lives in ChatGPT and connects 4,000+ apps.
[Rumor] Account claims Google is preparing Gemini 4 Carbon, with early-access partners already testingUnconfirmed
High engagement around a claimed next Gemini checkpoint; the post is a personal observation with no Google confirmation.
Collapse[Rumor] Account claims Google is preparing Gemini 4 Carbon, with early-access partners already testing
This is an unconfirmed claim from a single personal account; Google has made no announcement. Haleemah, an account focused on frontier-AI leaks, says Google is preparing to launch Gemini 4 Carbon, that early-access partners have received a Carbon EAP build and report a noticeable improvement over Gemini 4 Argon, with the downsides of being slower and weak at frontend work. She says she is testing it in Arena.