{"apiVersion":"v1","generatedAt":"2026-10-25T04:22:21Z","pulse":{"slug":"ai-industry","name":"AI Industry Pulse","description":"Business, policy and market developments across the AI industry, from X.","sections":[],"itemCount":38,"latestEditionItemCount":12,"lastEditionAt":"2026-10-25T03:57:00Z","feedUrl":"https://presto-ferry-g7bz.here.now/feeds/ai-industry.xml","itemsUrl":"https://presto-ferry-g7bz.here.now/api/v1/pulses/ai-industry.json"},"count":38,"items":[{"id":"5111eaf1b397de9c","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"ByteDance, Alibaba and Tencent reportedly order 3 million-plus Nvidia RTX PRO 5500 cards","url":"https://x.com/kyleichan/status/2114200713920168276","synopsis":"Citing Reuters, WSJ reporter Jonathan Cheng said ByteDance, Alibaba and Tencent have ordered more than 3 million of Nvidia's new RTX PRO 5500 graphics cards, with a combined sticker price of over $35 billion. @kyleichan's takeaway was simple: Chinese AI companies still really want Nvidia chips.","also":["https://x.com/JChengWSJ/status/2114137612432621995"],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":1},{"id":"7c79f1a841760818","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Why new inference providers are hitting $1M a month so fast","url":"https://x.com/deedydas/status/2114195317860266432","synopsis":"@deedydas wrote that new inference providers are hitting $1M a month in revenue very quickly, because even B300s rented at $8 per GPU-hour can earn $10 to $20 per GPU-hour serving open models like GLM 5.3, Kimi K3 and DeepSeek 4.1. He flagged the risks: Chinese labs now asking for revenue share at scale, commodity pricing pushing compute costs up and token prices down, and whether open models stay competitive.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":2},{"id":"575442c012adff16","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Microsoft open-sources Agensh, a harness for up to 1,024 coding agents with no lead agent","url":"https://x.com/Marktechpost/status/2114200105108586623","synopsis":"Microsoft Research released Agensh under the MIT license; every worker runs the same claim, build, verify and merge loop and coordinates through a shared Git repo and chat channel instead of an orchestrator. On the five hardest ProgramBench tasks the mean pass rate rose from 19.31% with one agent to 28.78% with 128, though going from 1 to 8 agents added only 1.37 points and token costs weren't disclosed.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":3},{"id":"66fb9fdd35309a48","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Kimi K3.1 Pro debuts 4th on Artificial Analysis, raising questions about Beijing's export draft","url":"https://x.com/poezhao0605/status/2114191468382355959","synopsis":"@poezhao0605 noted that Moonshot AI's Kimi K3.1 Pro launched at 4th on the Artificial Analysis Intelligence Index, after no Chinese model sat in the top 10 of capability boards in mid-October (though six were in OpenRouter's usage top 10). Beijing's draft export rules include a top-10 global ranking test without saying which ranking counts, and at a listed 2.8 trillion parameters K3.1 Pro already meets the size test.","also":["https://x.com/poezhao0605/status/2114191468663296106"],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":4},{"id":"d39511d73382172c","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Cognition open-sources a Devin plugin that patches this week's PHP, Node.js and GitLab RCEs","url":"https://x.com/Marktechpost/status/2114190379696476657","synopsis":"Cognition released free, MIT-licensed agent skills that inventory your Dockerfiles, packages, CI and infrastructure code read-only, then patch PHP2Shell, the Node.js October 21 release and GitLab CVE-2026-110264 only after you approve. It runs tests before and after and opens one PR per repo with a rollback plan, and works in Devin, Claude Code, Codex, Cursor and other agents; there's no published fix-rate benchmark yet.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":5},{"id":"f7985d157261714a","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Researchers plant a backdoor in Qwen 2.5-7B with just 125 poisoned examples","url":"https://x.com/kyleichan/status/2114199578341753197","synopsis":"@kyleichan highlighted research showing that poisoning only 125 fine-tuning examples was enough to make Qwen 2.5-7B run malicious code and steal project credentials when it saw a secret trigger phrase. It's a pointed reminder about supply-chain risk in fine-tuned open-weight models.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":7},{"id":"873a22651023ecfd","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"A 12-point feedback list after a week on Claude","url":"https://x.com/petergyang/status/2114202713567195254","synopsis":"After using Claude all week, @petergyang posted detailed feedback: Claude Code's browser use always opens new tabs, Remote Control threads vanish after idle hours, and the new Duet app won't change its check-in frequency or orchestrate local and cloud threads. On the model, he called it very good but said it lacks backbone and flips to agreeing too easily when pushed.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":8},{"id":"1ff0bb4645f456ad","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Theo answers his own Opus 5.5 jab with a long list of what he shipped","url":"https://x.com/theo/status/2114194770209947674","synopsis":"Following last hour's needling that builders haven't shipped anything impressive with Opus 5.5, Theo posted his own month: a TypeScript compiler rewritten in Rust that he says is 3x faster than the Go version, a model-comparison site called Slopalytics, and about 370 more merged PRs to T3 Code. He then asked others to post their lists, drawing about 200 likes quickly.","also":["https://x.com/theo/status/2114194806608183752"],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":9},{"id":"fef4463db9410e93","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"OpenAI announces a \"Cyber at the Frontier\" keynote for November 12","url":"https://x.com/TheRealAdamG/status/2114202035218219502","synopsis":"@TheRealAdamG of OpenAI shared registration for a live keynote on the next era of cyber defense, set for November 12 at 10:00 AM PT. It lands amid a week of AI-driven security news, including OpenAI's own account of hardening its training environment.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":10},{"id":"ff63ed63b5be5c14","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Lenovo previews a palm-sized AI box that runs 125B-parameter models locally","url":"https://x.com/thePandaily/status/2114195196380422210","synopsis":"According to Pandaily, Lenovo's AI Host P7 weighs about 300 grams, uses Houmo AI's M50 and CIX's P1 chips, and runs models up to 125 billion parameters with 256K context on the device. Early access opened ahead of sale.","also":[],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":11},{"id":"36a124efb8a68708","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"A clear explainer of a new paper on fixing softmax and cross-entropy","url":"https://x.com/jbhuang0604/status/2114192123084046712","synopsis":"@jbhuang0604 made an accessible explainer of a paper proposing fixes to softmax and cross-entropy, two building blocks of nearly every neural network. @boazbaraktcs called it a really nice explanation.","also":["https://x.com/boazbaraktcs/status/2114196647106290149"],"section":null,"publishedAt":"2026-10-25T03:57:00Z","firstSeenAt":"2026-10-25T04:22:21Z","editionAt":"2026-10-25T03:57:00Z","rank":12},{"id":"e66b1c713ef244e0","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Google's Logan Kilpatrick says Gemini API spend caps lag by 10 minutes","url":"https://x.com/OfficialLoganK/status/2114186751468265823","synopsis":"Answering @BenjaminDEKR's complaint that an $80 spend cap in AI Studio didn't stop a $209 bill, Logan Kilpatrick said the Gemini API's spend caps are real caps but currently have about a 10-minute latency, down from more than 24 hours originally, and that the team is building a real-time version. He said what Google support described were budget alerts, which only send emails. @BenjaminDEKR replied that ten minutes is an eternity and he will likely stop using the Gemini API.","also":["https://x.com/OfficialLoganK/status/2114187035791831408","https://x.com/OfficialLoganK/status/2114187252058427578","https://x.com/BenjaminDEKR/status/2114188114079924593"],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":1},{"id":"9c77cf0eaa99c69c","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Dax Raad shares numbers from his team's own DeepSeek inference rack","url":"https://x.com/thdxr/status/2114176734056857916","synopsis":"@thdxr reported that his team's new GPU rack running their own DeepSeek inference setup handled 6.77 trillion tokens in the last 24 hours, with a median time to first token of 1.16 seconds and a median of 143 tokens per second. He said the 95th-percentile numbers are worse because one tray is flaky, and earlier claimed the setup beats every inference vendor they've found on price and performance.","also":["https://x.com/thdxr/status/2111849415204065611"],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":2},{"id":"5d1ca01d919458f4","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Aaron Levie: cyber agents will soon process 1,000 times more tokens","url":"https://x.com/levie/status/2114187326406766604","synopsis":"Box CEO Aaron Levie predicted 1,000 times more tokens will soon flow through cybersecurity agents than today, arguing the growing attack surface is already outpacing what humans can keep up with. He was quoting a post on Node2Shell (CVE-2026-111497), a memory bug in the Brotli compression library that ships inside Node.js, Python, system libraries and Envoy, often as a hidden dependency.","also":["https://x.com/NickADobos/status/2113430301183823901"],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":3},{"id":"423f60fececd0438","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Stranded-compute debate continues after Elon Musk's warning","url":"https://x.com/BenjaminDEKR/status/2114183767694725611","synopsis":"Following last hour's post from Elon Musk predicting a lot of stranded compute in half-built data centers, @BenjaminDEKR asked why there would be stranded compute if AI demand is endless. Others on the list resurfaced Musk's earlier remark that agentic demand is growing faster than anything he has ever seen grow.","also":["https://x.com/elonmusk/status/2114036686870982681"],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":4},{"id":"b70bb7d1abb608fc","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Anthropic vs. OpenAI: shipping pace versus discounts","url":"https://x.com/weswinder/status/2114175511676993988","synopsis":"@weswinder argued that OpenAI spent the week winning coding customers with steep discounts while Anthropic shipped Sonnet 5.5, Haiku 5.5 and the Claude Duet app in the month since Opus 5.5, with everyone now waiting on Fable 5.5. Separately, Theo needled builders that it's been over a month since Opus 5.5 launched and they still haven't shipped anything impressive with it, drawing about 475 likes in under an hour.","also":["https://x.com/theo/status/2114177918226354678"],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":5},{"id":"eab45b42279d2a30","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Kea-T squeezes NVIDIA's Parakeet speech model down to ternary weights","url":"https://x.com/lmoroney/status/2114185429230461168","synopsis":"Laurence Moroney highlighted Kea-T, which takes NVIDIA's Parakeet-TDT-0.6B-v2 speech recognition model and constrains the big encoder matrices, about 98% of its parameters, to just -1, 0 or +1 with one scale per row. The goal is to test how small a speech model can get before it stops being useful, and the post says the weight file shrinks sharply from its original 2,472 MB.","also":[],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":6},{"id":"97fe3e400a9135cd","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"China Mobile opens a shared AI compute platform for carmakers","url":"https://x.com/thePandaily/status/2114180903802958126","synopsis":"According to Pandaily, China Mobile launched a shared AI foundation for the auto industry with 112.7 EFLOPS of compute plus six agents for styling, process planning, inspection, maintenance and driving data. It says more than 10 state-owned carmakers already use it.","also":[],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":7},{"id":"5cc1a862bc995f5a","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Julian Togelius asks why every AI lab is shipping the same thing","url":"https://x.com/togelius/status/2114185179606437950","synopsis":"NYU professor Julian Togelius asked whether every lab ships the same products because Silicon Valley is full of people who have worked at the same places on the same problems. In follow-ups he called the monoculture intellectually impoverished and suggested investors look at other parts of the world where the bets are different.","also":["https://x.com/togelius/status/2114186079532376502"],"section":null,"publishedAt":"2026-10-25T02:56:00Z","firstSeenAt":"2026-10-25T03:26:59Z","editionAt":"2026-10-25T02:56:00Z","rank":8},{"id":"45fbf1667d1c504d","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Elon Musk predicts \"a lot of stranded compute\" in half-built data centers","url":"https://x.com/elonmusk/status/2114168277882581486","synopsis":"Replying to @Yuchenj_UW's earlier post about neolabs that locked in GPUs early and can now resell compute at two to three times the price, Elon Musk said he expects a lot of stranded compute in partially complete data centers. @Yuchenj_UW joked that Elon just ruined the mood for his neolab and neocloud friends. The exchange landed as a16z's Charts of the Week circulated on the list, showing GPU rental rates up since January 1 (H200 about 160%, B200 about 120%, H100 about 35%), with even the six-year-old A100 renting for more.","also":["https://x.com/Yuchenj_UW/status/2114174381341769803","https://x.com/a16z/status/2113741008550990191","https://x.com/TheAhmadOsman/status/2114165218628559037"],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":1},{"id":"d34b453aa3666b09","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Theo confirms an Anthropic bug: Agents SDK drains Claude 5-hour limits faster","url":"https://x.com/theo/status/2114166439813017873","synopsis":"After users of T3 Code reported burning through a Claude Max 5-hour limit in about 30 minutes, Theo said he confirmed it's a real bug on Anthropic's side: all Agents SDK usage drains 5-hour limits faster than using Claude Code directly, though weekly limits aren't affected. He suspects Anthropic's recent boost to 5-hour limits was never applied to the Agents SDK, and he has reported it to Anthropic.","also":["https://x.com/DeepakNesss/status/2113912028675162171","https://x.com/theo/status/2114166577608478826","https://x.com/theo/status/2114166866579259501"],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":2},{"id":"8fd129b785e9bd05","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"@jack unveils Dessa, a full-duplex speech model","url":"https://x.com/jack/status/2113697902916751661","synopsis":"@jack introduced Dessa, which he calls the most authentic full-duplex speech model ever built, saying talking with it feels like talking with a person, with emotion, nuance, pauses and natural cadence. The post has passed 12 million views and 64,000 likes and was being reposted across the list this hour.","also":[],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":3},{"id":"88f204224c361e44","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Developer says Gemini AI Studio spend caps didn't stop a $200 bill","url":"https://x.com/BenjaminDEKR/status/2114168123859317091","synopsis":"@BenjaminDEKR asked Google's @OfficialLoganK why AI Studio spend caps don't actually cap spending: he set an $80 cap on a project and now owes over $200. He added that Google's support told him caps only send a notification, which contradicts the Gemini API docs.","also":["https://x.com/BenjaminDEKR/status/2114168509433328025","https://x.com/BenjaminDEKR/status/2114173065521193145"],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":4},{"id":"fb3880f577db1610","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"vLLM explains how it matched DeepSeek's own engine on V4.1 Flash","url":"https://x.com/woosuk_k/status/2114120162668626132","synopsis":"Following last hour's news that vLLM now matches DeepSeek's in-house DeepInfer engine on DeepSeek-V4.1-Flash, vLLM co-creator Woosuk Kwon published a blog post on how the team did it and said the results are also verified in SemiAnalysis's AgentX benchmark.","also":[],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":5},{"id":"ac70715af47fd460","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Moonshot stripped cyber training data from open Kimi model, with tradeoffs","url":"https://x.com/kyleichan/status/2113745668133011825","synopsis":"Kyle Chan noted that instead of adding cyber guardrails in post-training, Moonshot removed cyber-related data from mid- and post-training for the open-weight Kimi 3.1 Pro, so the guardrails can't simply be fine-tuned away, while keeping a more powerful internal version with full cyber capabilities. In a follow-up he warned that this likely weakens defensive cyber, coding and even general capabilities, trading tougher guardrails for a less powerful model.","also":["https://x.com/kyleichan/status/2114168368383049876"],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":6},{"id":"f3c665fe648cd098","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"alphaXiv now lets researchers publish papers directly","url":"https://x.com/askalphaxiv/status/2114168840775237998","synopsis":"alphaXiv announced that researchers can now publish directly on its platform instead of waiting on arXiv. It pitches the option as much faster with no endorsement required, relying on community curation such as comments, bookmarks and following researchers to surface the right papers.","also":[],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":7},{"id":"b43bd5704ec411be","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"NODECODE: an open-source, self-rewriting agent written in Lisp","url":"https://x.com/pomterree/status/2113721985167728757","synopsis":"A new MIT-licensed open-source agent called NODECODE rewrites itself as it works and is written in Lisp, echoing John McCarthy's idea of a program you improve by making statements to it. It runs in Discord, Telegram and similar apps and is aimed at cyber work; @GeoffreyHuntley reposted it to the list.","also":[],"section":null,"publishedAt":"2026-10-25T01:58:00Z","firstSeenAt":"2026-10-25T02:21:23Z","editionAt":"2026-10-25T01:58:00Z","rank":8},{"id":"79abefb3a542a0a1","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"vLLM now matches DeepSeek's own inference engine on V4.1 Flash","url":"https://x.com/lmoroney/status/2114156446417813985","synopsis":"Laurence Moroney passed along the vLLM team's claim that the open-source server now matches DeepInfer, the in-house engine DeepSeek open-sourced earlier today, on DeepSeek-V4.1-Flash decoding across batch sizes 8 to 256. They got there by pulling in DeepSeek's open kernels and heavy prefill and decode tuning, running at DEP32 where DeepInfer used EP64DP64. His advice: if you serve DeepSeek models, rerun your own benchmark on the latest vLLM before adding a second serving stack.","also":[],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":1},{"id":"6e8b14c843e0651c","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Jill Lepore: Sam Altman is more powerful than Donald Trump","url":"https://x.com/AndrewCurran_/status/2114148065397620910","synopsis":"Andrew Curran shared excerpts from historian Jill Lepore's Vanity Fair interview, in which she argues that Sam Altman is more powerful than the president and that what she calls \"The Artificial State\" now controls the United States and the world. It was one of the most discussed AI posts on the list this hour.","also":["https://x.com/AndrewCurran_/status/2114148069499703436"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":2},{"id":"fdd2fd5228fc59ca","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Open frontier cyber models: Benioff says open source is the frontier again","url":"https://x.com/Benioff/status/2114144973851595003","synopsis":"Salesforce CEO Marc Benioff said this week showed how new open-source, frontier-class cyber capabilities are changing global security, declaring that \"open source is the frontier again.\" @poezhao0605 noted that Kimi K3.1 Pro, listed by Artificial Analysis at 2.8 trillion parameters, sits above the 2 trillion line in Beijing's draft AI export rules, yet Moonshot says full weights ship by November 4. @yacineMTB argued it's already too late after DeepSeek Flash 1017 and that defenders should simply get the same tools as everyone else.","also":["https://x.com/poezhao0605/status/2113836755279544783","https://x.com/yacineMTB/status/2114027470944948396"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":3},{"id":"ffffcdba029c3b30","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Weekly roundup: OpenAI dates its training pause, Sweden blocks ChatGPT and Claude","url":"https://x.com/zaibatsu/status/2114149404361863610","synopsis":"A newsletter roundup from @zaibatsu says OpenAI has put dates on its training pause and \"cut the live internet for good,\" and that Sweden's government has blocked ChatGPT and Claude. It also notes the departure of two of Claude Code's public faces, which earlier editions covered.","also":[],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":5},{"id":"46b6806d13882673","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"GPT-5.6 may be closing Copilot's gap with OpenAI and Anthropic","url":"https://x.com/FundaAI/status/2114149392056033740","synopsis":"FundaAI argued that Microsoft Copilot long struggled to match OpenAI's and Anthropic's own apps despite heavy harness engineering, but newer models like GPT-5.6 can now take over much of that work themselves. That narrows the user-experience gap, which fits earlier channel checks showing Copilot usage picking up.","also":["https://x.com/FundaAI/status/2114140252998975980"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":6},{"id":"4bac346669255afc","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"More than 1,000 mathematicians pledge to avoid AI, and a prominent scientist pushes back","url":"https://x.com/Hesamation/status/2114153199414157752","synopsis":"A post going around the list says over 1,000 mathematicians have joined an Association for Human Mathematics, pledging not to publish AI-generated papers or work with frontier labs, with 654 vowing not to use AI in research at all. @DeryaTR_ called it a sad state of affairs, asking who will pay for academic work on problems AI has already solved.","also":["https://x.com/DeryaTR_/status/2114198659470651797"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T03:57:00Z","rank":6},{"id":"b22542678f69ac8d","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Town AI raises a $108M Series B for multiplayer work assistants","url":"https://x.com/jgreze/status/2112184910702403754","synopsis":"Convex's weekly builder roundup highlighted that @jgreze and the Town AI team raised a $108 million Series B for AI assistants that whole teams share at work.","also":["https://x.com/waynesutton/status/2114156390143197329"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":7},{"id":"654e89d821873298","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Google DeepMind paper fixes hidden gradient errors in softmax and cross-entropy","url":"https://x.com/burkov/status/2114157841472774530","synopsis":"Andriy Burkov summarized the DeepMind paper behind Jon Barron's Gemini-assisted workflow: standard softmax and cross-entropy code has two rounding errors, one that kills the gradient for very confident correct predictions and one that gives wrong gradients when a student model nearly matches its teacher. Because optimizers like Adam amplify tiny gradient errors, the authors offer drop-in stable replacements and test them in bfloat16, 32-bit and 64-bit.","also":[],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":8},{"id":"018a6d5fd3c37ace","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Latent Space is building an editorial team","url":"https://x.com/swyx/status/2114145315196711194","synopsis":"swyx said the Latent Space podcast is hiring writers, reviewers and reporters for a new editorial team, with @ricmac and @adlinzainal as contacts, and part-time and one-off guest posts are welcome.","also":["https://x.com/swyx/status/2114149330135519730"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":9},{"id":"f318dcdddf4e6d6e","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"Agent swarms make ambitious systems research practical","url":"https://x.com/sh_reya/status/2114149806964957209","synopsis":"@sh_reya pointed to a case study showing AI-driven systems work mainly needs smart humans, good metrics and cost models, and basic swarms of agents writing code and tests. She argued systems PhD students and academics can now build very ambitious systems and evolve them in the open for a year or more.","also":["https://x.com/MOVNTDQ/status/2114045515876995111"],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":10},{"id":"c70127f5f11e4b98","pulse":"ai-industry","pulseName":"AI Industry Pulse","title":"ChatGPT's image generator can fill your bookshelf with books it thinks you'd read","url":"https://x.com/venturetwins/status/2114144079345078720","synopsis":"@venturetwins showed that you can give ChatGPT a photo of an empty bookshelf and ask it to fill the shelves with real books based on what it knows about you, a sign of how good image generation has become.","also":[],"section":null,"publishedAt":"2026-10-25T00:54:00Z","firstSeenAt":null,"editionAt":"2026-10-25T00:54:00Z","rank":11}]}