DeepSeek open-sources DeepInfer production inference stack
The serving stack behind a live model was open-sourced, and the post led 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 21, 2026
20 items for 2026-10-21, ranked by combined engagement. Unverified items are marked.
Top 3
OpenAI: latest models boost COBOL, Fortran, and Visual Basic work
The official post names legacy languages that still run core systems.
Anthropic: Claude designs scorpion-toxin protein binders in 64 hours
The post publishes the design, lab, and hit counts, and says this is not an antivenom.
All
DeepSeek open-sources DeepInfer production inference stack
The serving stack behind a live model was open-sourced, and the post led the day.
CollapseDeepSeek open-sources DeepInfer production inference stack
DeepSeek released DeepInfer, the in-house inference framework powering DeepSeek-V4.1-Flash at scale, as fully open source—aimed at high throughput, low latency, and lower compute/energy costs for the community.
OpenAI: latest models boost COBOL, Fortran, and Visual Basic work
The official post names legacy languages that still run core systems.
CollapseOpenAI: latest models boost COBOL, Fortran, and Visual Basic work
OpenAI says its latest model advances bring stronger capabilities for building and migrating software in low-resource languages such as COBOL, Fortran, and Visual Basic—still powering banks, governments, and businesses.
Anthropic: Claude designs scorpion-toxin protein binders in 64 hours
The post publishes the design, lab, and hit counts, and says this is not an antivenom.
CollapseAnthropic: Claude designs scorpion-toxin protein binders in 64 hours
Anthropic’s Science Blog: working with Citrini, Claude generated 300k+ protein designs against scorpion neurotoxins over ~64 hours; 17 of 211 lab-tested met binder criteria. Not an antivenom yet; designs and data released.
OpenAI acquires Orchestra Bio to streamline biopharma R&D
The deal is a drug-development coordination product, tied to Rosalind in the post.
CollapseOpenAI acquires Orchestra Bio to streamline biopharma R&D
OpenAI acquired Orchestra Bio, a platform that helps biopharma teams coordinate drug development, to build on Rosalind and its life-sciences AI tools.
Sam Altman: Aleksander to lead safety/security for superintelligence
A public thanks for someone taking safety and security as the lab talks about superintelligence.
CollapseSam Altman: Aleksander to lead safety/security for superintelligence
Sam Altman posted gratitude that Aleksander accepted OpenAI’s request to take on safety and security challenges as the company confronts superintelligence.
DeepMind: Gemini 4 Argon builds 3D periodic table synced to Tom Lehrer
An official demo of an interactive 3D periodic table timed to a song.
CollapseDeepMind: Gemini 4 Argon builds 3D periodic table synced to Tom Lehrer
Google DeepMind showed Gemini 4 Argon building a 3D periodic table in Antigravity—animating orbiting electrons, atomic data, and light spectra synced to Tom Lehrer’s “The Elements.”
Dario Amodei replies to Nathan Calvin on Anthropic’s safety record
The CEO answered a safety-versus-race critique in public and listed costly safety steps.
CollapseDario Amodei replies to Nathan Calvin on Anthropic’s safety record
Anthropic CEO Dario Amodei posted a long reply thanking Nathan Calvin for scrutiny, defending Anthropic’s “race to the top” strategy and listing costly safety steps while acknowledging communication slips.
Karpathy: “Yeah I’d say the RL is working now all right”
One line that reinforcement learning is working now drew the thread. The post has no method.
CollapseKarpathy: “Yeah I’d say the RL is working now all right”
Andrej Karpathy replied that reinforcement learning “is working now all right,” sparking engagement across the AI training discourse.
Logan Kilpatrick: ~25T tokens in first 72 hours of Gemini 4 Argon
A 72-hour token figure was posted and amplified, including by Demis Hassabis.
CollapseLogan Kilpatrick: ~25T tokens in first 72 hours of Gemini 4 Argon
Logan Kilpatrick said Gemini 4 Argon processed ~25 trillion tokens in its first 72 hours—demand for intelligence “still under appreciated.” Demis Hassabis and others amplified it.
swyx: DeepSeek Zhihu page has more V4.1 Flash inference alpha
Richer V4.1 Flash inference detail is pointed at a Zhihu page rather than the short post.
Collapseswyx: DeepSeek Zhihu page has more V4.1 Flash inference alpha
swyx noted DeepSeek posted richer technical detail on Zhihu than on Twitter—covering a high-throughput, low-latency large-scale inference service for DeepSeek V4.1 Flash.
Nathan Calvin’s long critique: is Anthropic safety-first or racing?
A long critique lists chips, compute, and RSP changes, and it drew the CEO’s reply.
CollapseNathan Calvin’s long critique: is Anthropic safety-first or racing?
Safety advocate Nathan Calvin listed Anthropic’s chip plans, compute deals, RSP V3 changes and more, arguing outsider evidence looks more like racing to superintelligence than a safety-first org—prompting Amodei’s reply.
[Unverified] Screenshot claims Anthropic S-1 warns of AI time travelUnconfirmed
The screenshot traveled widely.
Collapse[Unverified] Screenshot claims Anthropic S-1 warns of AI time travel
A viral screenshot claims Anthropic’s IPO S-1 warns that advanced AI could transmit information/matter backward in time. High engagement, but counter-checks say the confidential draft reviews don’t mention time travel—treat as unverified rumor.
Report buzz: OpenAI seeks ~$1.4T valuation raise ahead of Anthropic IPO
A valuation and raise figure spread through market-wire posts and Chinese discussion.
CollapseReport buzz: OpenAI seeks ~$1.4T valuation raise ahead of Anthropic IPO
Market wire posts say OpenAI is racing to raise ~$30B at a valuation over $1.4T before Anthropic’s expected November IPO, claiming ~$50B annualized revenue. Widely discussed in CN/intl finance-AI circles.
Simon Willison asks for the prompts behind soulless reply bots
A public ask for the prompts behind empty reply bots opened a long thread.
CollapseSimon Willison asks for the prompts behind soulless reply bots
Simon Willison asked whether anyone has seen the prompts used by “awful reply bots,” offering to tolerate one more reply if they share—drawing wide discussion.
Jenson李: Why is Kimi cooler with GPT/Claude but quieter than Doubao/DeepSeek?
The question sets ability next to GPT and Claude against quieter domestic buzz than Doubao and DeepSeek.
CollapseJenson李: Why is Kimi cooler with GPT/Claude but quieter than Doubao/DeepSeek?
kale243222 asked why Kimi—seen as one of few CN models that can hang with GPT/Claude—stays cooler in domestic buzz than Doubao and DeepSeek. 320k+ views and a large reply thread on distribution vs capability.
Simon Willison builds browser RTS Battle Drones mainly with Claude Opus 5.5
A browser RTS built mainly with Claude Opus 5.5, with multiplayer described in the post.
CollapseSimon Willison builds browser RTS Battle Drones mainly with Claude Opus 5.5
Simon Willison shipped Battle Drones, a Command & Conquer–style RTS built mainly with Claude Opus 5.5, running in-browser with device-to-device audio multiplayer (later WebSocket online play).
DHS order to exclude/remove DeepSeek from civilian federal systems (viral JP post)
A Japanese summary of the DHS exclude-and-remove order drew more engagement than the Chinese post on the same order.
CollapseDHS order to exclude/remove DeepSeek from civilian federal systems (viral JP post)
A widely shared Japanese post summarizes DHS’s order excluding DeepSeek (and affiliates) from civilian federal procurement and requiring removal from existing IT systems.
Artificial Analysis: intelligence-index cost collapses after thresholds are hit
A concrete multiple: at an index of 30, cost per task is about 86× lower than when first measured.
CollapseArtificial Analysis: intelligence-index cost collapses after thresholds are hit
Artificial Analysis reported that once Intelligence Index thresholds are reached, cost to access that capability falls fast—reaching 30 now costs ~86× less per task than when first measured.
xiaohu: DHS requires civilian agencies to purge DeepSeek products
A Chinese retelling of the same removal order, with less engagement than the Japanese post.
Collapsexiaohu: DHS requires civilian agencies to purge DeepSeek products
Chinese tech account xiaohu summarized DHS requiring civilian federal agencies to stop buying DeepSeek products/services and immediately remove them (and DeepSeek-based third-party products) from government IT systems.
Trending story: Claude fewer users, higher revenue-per-user vs ChatGPTUnconfirmed
A trending page debates user scale and revenue per user. The page has no post-level counts.
CollapseTrending story: Claude fewer users, higher revenue-per-user vs ChatGPT
X News trending story summarizing debate that Claude’s smaller user base may still outpace ChatGPT on revenue run-rate via higher paid/professional ARPU. Aggregated topic page—verify figures independently.