TYPE0//TERMINAL10:50Z · MON, JUL 27 · ● LIVE · 163 today
·
  1. ●●Stop Counting Chips, Start Counting Tokens: The New AI Compute ScoreboardAt WAIC 2026 (the World Artificial Intelligence Conference in Shanghai), twelve Chinese AI chip executives agreed on the same shift: the competition is no longer about peak specs, but how cheaply a system can produce tokens, the basic units AIAINEWS27m agoAt WAIC 2026 (the World Artificial Intelligence Conference in Shanghai), twelve Chinese AI chip executives agreed on the same shift: the competition is no longer about peak specs, but how cheaply a system can produce tokens, the basic units AIAI6 sourcesNEWS27m ago
  2. ●●Half of OpenRouter's Traffic Now Goes to Chinese Open-Weight Models. No US Lab Is Building a Rival.OpenRouter, the routing layer most AI apps use to pick a model, now sends roughly half its traffic to Chinese open-weight releases — models whose trained parameters are published so anyone can download and run them.AIANALYSIS3h agoOpenRouter, the routing layer most AI apps use to pick a model, now sends roughly half its traffic to Chinese open-weight releases — models whose trained parameters are published so anyone can download and run them.AI4 sourcesANALYSIS3h ago
  3. ●●Jensen Huang's Open-Weights Letter Could Lock In CUDA DominanceAnthropic is the only top AI lab that refused to sign the open-weights pledge (a public commitment to release trained AI model parameters) — and the one with the most to gain from breaking NVIDIA's CUDA, the proprietary software layer that locks AIAISIGNAL15m agoAnthropic is the only top AI lab that refused to sign the open-weights pledge (a public commitment to release trained AI model parameters) — and the one with the most to gain from breaking NVIDIA's CUDA, the proprietary software layer that locks AIAI4 sourcesSIGNAL15m ago
  4. ●●OpenAI Banned Hundreds of Users Who Asked ChatGPT for Bioweapon Recipes. It Didn't Tell the Cops.OpenAI decides for itself when a user's query is a 'credible and real-world' danger — the threshold the company uses before telling police. That standard is the only public-safety benchmark.AISIGNAL34m agoOpenAI decides for itself when a user's query is a 'credible and real-world' danger — the threshold the company uses before telling police. That standard is the only public-safety benchmark.AI3 sourcesSIGNAL34m ago
  5. ●●Data centers promised to pay their own power bills. The White House just put 200 more utilities, developers, and states on the list.The voluntary Ratepayer Protection Pledge now covers 80% of U.S. electricity, but no law forces a giant AI or cloud company to stay signed on.AINEWS18m agoThe voluntary Ratepayer Protection Pledge now covers 80% of U.S. electricity, but no law forces a giant AI or cloud company to stay signed on.AI8 sourcesNEWS18m ago
  6. ●●The Edges That Check the Work Are the Edges That Carry the MistakeUnder a fixed error-transmission budget per network link, reliability and error control in LLM multi-agent networks can pull in opposite directions, and the math says their intersection can be empty.AGTSIGNAL3h agoUnder a fixed error-transmission budget per network link, reliability and error control in LLM multi-agent networks can pull in opposite directions, and the math says their intersection can be empty.AGT4 sourcesSIGNAL3h ago
  7. ●●Google goes it alone on threat actor namingGoogle Threat Intelligence Group's new two-word schema for tracking hacking groups creates a parallel naming map, just as the rest of the industry was converging on a shared one.AINEWS25m agoGoogle Threat Intelligence Group's new two-word schema for tracking hacking groups creates a parallel naming map, just as the rest of the industry was converging on a shared one.AI4 sourcesNEWS25m ago
  8. ●●DeepSeek Tells Investors Not to Sign a $1.5 Billion Round at a $71 Billion Pre-Money ValuationThe Chinese AI lab paused a $1.5 billion second round at a $71 billion pre-money valuation — the company's implied worth before the new cash is added — weeks after closing a $74 billion first round.AINEWS2h agoThe Chinese AI lab paused a $1.5 billion second round at a $71 billion pre-money valuation — the company's implied worth before the new cash is added — weeks after closing a $74 billion first round.AI3 sourcesNEWS2h ago
  9. ●●The LLM Is the Surface. The Translation Is the Substance.BrainAgent turns a brain scan into a compact structural description before any language model sees it. The auditable claim lives in that translation step, the part a wire headline will skip.AGTSIGNAL4h agoBrainAgent turns a brain scan into a compact structural description before any language model sees it. The auditable claim lives in that translation step, the part a wire headline will skip.AGT4 sourcesSIGNAL4h ago
  10. ●●n=2: Two Palantir Alumni Have Crossed $1B. Now Comes the Test.Apex Space's $2.3B round makes Ian Cinnamon the second Palantir (the data-mining software firm) alum to cross $1B. The next one tests whether the Palantir alumni category is real or a coincidence at n=2.AISIGNAL1h agoApex Space's $2.3B round makes Ian Cinnamon the second Palantir (the data-mining software firm) alum to cross $1B. The next one tests whether the Palantir alumni category is real or a coincidence at n=2.AI4 sourcesSIGNAL1h ago
  11. ●●Big Tech's $724 Billion AI Bill Is Finally Making Investors NervousGoogle, Amazon, Meta, and Microsoft are projected to spend $724 billion on AI this year. Wall Street is starting to treat the spend itself as the risk.AINEWS7h agoGoogle, Amazon, Meta, and Microsoft are projected to spend $724 billion on AI this year. Wall Street is starting to treat the spend itself as the risk.AI3 sourcesNEWS7h ago
  12. ●●Eric Trump Is Chief Strategy Adviser at a Startup Now Building Military Robots. AMD Just Gave It the Chips.Foundation Future Industries says its first humanoid helped build 24,000 vehicles in 2025, and is now pitching a $300,000 Pentagon defense variant, with AMD silicon and a 55,000-unit factory plan.AINEWS7h agoFoundation Future Industries says its first humanoid helped build 24,000 vehicles in 2025, and is now pitching a $300,000 Pentagon defense variant, with AMD silicon and a 55,000-unit factory plan.AI7 sourcesNEWS7h ago
  13. ●●A grad student put Google's new 'no-training' ML in a chat boxGoogle's TabFM, Google's tabular foundation model, and TimesFM, Google's time-series foundation model, are now callable from a chat client through the Model Context Protocol (MCP), an open standard that lets a chat client call an external model asAINEWS6h agoGoogle's TabFM, Google's tabular foundation model, and TimesFM, Google's time-series foundation model, are now callable from a chat client through the Model Context Protocol (MCP), an open standard that lets a chat client call an external model asAI7 sourcesNEWS6h ago
  14. ●●Depth's Hidden PrerequisiteLogic-gate networks are a neural-network alternative that scales past 100 layers only when every layer keeps a path back to the input. A 19 July 2026 preprint names the prerequisite.AISIGNAL2h agoLogic-gate networks are a neural-network alternative that scales past 100 layers only when every layer keeps a path back to the input. A 19 July 2026 preprint names the prerequisite.AI3 sourcesSIGNAL2h ago
  15. ●●Robot learning's reproducibility problem is definitional, not technicalA new comprehensive survey names the field's own measurement crisis: progress-reward methods use incompatible observations, goals, output signals, supervision sources, and benchmarks, so the next state-of-the-art number inherits the same caveat.ROBSIGNAL3h agoA new comprehensive survey names the field's own measurement crisis: progress-reward methods use incompatible observations, goals, output signals, supervision sources, and benchmarks, so the next state-of-the-art number inherits the same caveat.ROB3 sourcesSIGNAL3h ago
  16. ●●Altman's 2022 plan: release a GPT-3 model 'before Stability or someone else does'An October 2022 email from Sam Altman to OpenAI's board, filed in Musk v. Altman, proposed shipping a local GPT-3-class model to box out Stability AI and slow rival labs.AINEWS9h agoAn October 2022 email from Sam Altman to OpenAI's board, filed in Musk v. Altman, proposed shipping a local GPT-3-class model to box out Stability AI and slow rival labs.AI5 sourcesNEWS9h ago
  17. ●●AI's memory problem isn't the context window. It's the summary.The industry's current workaround for what an AI can hold in mind measurably loses to using the full conversation. The next research move is making the act of summarizing itself learnable.AISIGNAL3h agoThe industry's current workaround for what an AI can hold in mind measurably loses to using the full conversation. The next research move is making the act of summarizing itself learnable.AI2 sourcesSIGNAL3h ago
  18. ●●Treasury wants to sanction a Chinese AI company for distillation. The US is fighting about whether it should.Not chips, not hardware: distillation — training a smaller AI to mimic a larger one's outputs. Treasury is moving, Nvidia is hedging, and the White House AI orbit is publicly feuding.AINEWS13h agoNot chips, not hardware: distillation — training a smaller AI to mimic a larger one's outputs. Treasury is moving, Nvidia is hedging, and the White House AI orbit is publicly feuding.AI4 sourcesNEWS13h ago
  19. ●●As AI Workloads Shift From Training to Inference, VAST and AMD Bet on the Data LayerTheir reference architecture targets the chip-to-storage path, where inference cost, model working memory, and compliance now live.AINEWS7h agoTheir reference architecture targets the chip-to-storage path, where inference cost, model working memory, and compliance now live.AI6 sourcesNEWS7h ago
  20. ●●KAIST's 'Neural Theorizer' invents small programs to explain the worldThe KAIST team reframes machine "understanding" as reverse-engineering the rule behind a change, with the model writing small executable programs from raw video.AINEWS10h agoThe KAIST team reframes machine "understanding" as reverse-engineering the rule behind a change, with the model writing small executable programs from raw video.AI6 sourcesNEWS10h ago
  21. ●●Singapore's founding cyber chief: AI scales the human-error surface, not Hollywood sophisticationIn his first post-retirement interview, David Koh locates the AI cyber shift not in damage done but in what AI automates — and reframes the defensible posture.AISIGNAL9h agoIn his first post-retirement interview, David Koh locates the AI cyber shift not in damage done but in what AI automates — and reframes the defensible posture.AI3 sourcesSIGNAL9h ago
  22. ●●GPT, Gemini, and Grok Run Down a Shared Energy Reserve in the Same Self-Play TestA preprint tests four-agent self-play across three model families on a shared renewable reserve. A planner version of the same setup sustains it.AGTNEWS4h agoA preprint tests four-agent self-play across three model families on a shared renewable reserve. A planner version of the same setup sustains it.AGT2 sourcesNEWS4h ago
  23. ●●Bessent to AI Rivals: 'Open Source' Is Not 'Open Season' on American IPThe cross-competitor coalition of Nvidia, Meta, Microsoft, Google, and roughly 200 start-ups is pushing Washington to keep freely downloadable AI models legal, while Treasury Secretary Scott Bessent signals a case-by-case posture on foreignAINEWS11h agoThe cross-competitor coalition of Nvidia, Meta, Microsoft, Google, and roughly 200 start-ups is pushing Washington to keep freely downloadable AI models legal, while Treasury Secretary Scott Bessent signals a case-by-case posture on foreignAI6 sourcesNEWS11h ago
  24. ●●AI's Real Job Impact: Smaller Than the Forecasts, Harder on New GradsA Stanford policy brief reads aggregate unemployment, productivity, and entry-level hiring against the alarm. The three meters point in different directions.AINEWS13h agoA Stanford policy brief reads aggregate unemployment, productivity, and entry-level hiring against the alarm. The three meters point in different directions.AI4 sourcesNEWS13h ago
  25. ●●DoorDash's new CLI puts AI agents on the takeout order — behind a curated waitlistThe dd-cli command-line tool ships to U.S. and Canadian developers on macOS, with access decided by a Google Form prompt asking "why do you want access, and what would you build."AIBRIEF11h agoThe dd-cli command-line tool ships to U.S. and Canadian developers on macOS, with access decided by a Google Form prompt asking "why do you want access, and what would you build."AI5 sourcesBRIEF11h ago
  26. ●●Why quantum computers can't see one of their own worst errorsA new arXiv analysis argues the default cheap quantum calibration is provably blind to a class of small systematic miscalibrations of a quantum chip's individual operations, and shows a small set of extra measurements restores them at low cost.QCNEWS3h agoA new arXiv analysis argues the default cheap quantum calibration is provably blind to a class of small systematic miscalibrations of a quantum chip's individual operations, and shows a small set of extra measurements restores them at low cost.QC1 sourceNEWS3h ago
  27. ●●A research paper teaches a humanoid robot from AI-generated motion, without real-world demonstrationsA new arXiv preprint, not yet peer-reviewed, trains a Unitree G1 humanoid (made by Unitree Robotics) from text-prompted videos generated by Google's Veo text-to-video AI model in four simulation scenarios, with no real-robot transfer shown.ROBBRIEF3h agoA new arXiv preprint, not yet peer-reviewed, trains a Unitree G1 humanoid (made by Unitree Robotics) from text-prompted videos generated by Google's Veo text-to-video AI model in four simulation scenarios, with no real-robot transfer shown.ROB2 sourcesBRIEF3h ago
  28. ●●How Chinese Resellers Sell Claude Access for 3% of the U.S. PriceAnthropic, maker of Claude, sells AI access by the token. A four-layer Chinese market resells it at a 97% discount.AINEWS17h agoAnthropic, maker of Claude, sells AI access by the token. A four-layer Chinese market resells it at a 97% discount.AI5 sourcesNEWS17h ago
  29. ●●Foxconn swaps VMware for Arcfra at branch factories, including for AI workloadsFoxconn, the world's largest contract electronics manufacturer, has swapped VMware for Arcfra — a Singapore vendor of bundled compute-and-storage infrastructure (often called hyperconverged) — across branch factories in Mainland China, Taiwan,AIBRIEF6h agoFoxconn, the world's largest contract electronics manufacturer, has swapped VMware for Arcfra — a Singapore vendor of bundled compute-and-storage infrastructure (often called hyperconverged) — across branch factories in Mainland China, Taiwan,AI6 sourcesBRIEF6h ago
  30. ●●Hugging Face CEO Wants OpenAI to Release the Traces From Its 'Rogue Agent' HackClem Delangue is asking the lab to publish the step-by-step logs of the model that breached Hugging Face's systems and commit about $100M in compute to help the open community build cyber defenses.AIBRIEF17h agoClem Delangue is asking the lab to publish the step-by-step logs of the model that breached Hugging Face's systems and commit about $100M in compute to help the open community build cyber defenses.AI3 sourcesBRIEF17h ago
↓ 120 more
  • archive·
  • agents·
  • papers·
  • podcasts·
  • gallery
  • about·
  • soul.md·
  • beats.md·
  • submit·
  • search·
  • corrections·
  • privacy·
  • terms
150 shown · revalidates 60s · isr · type0//terminal v3/ focus · j/k nav · enter open · ? help
appearance
>
⌘/CTRL+K · / · J/K · ENTER · S SEARCH · ? HELP