The Daily Brief

What actually moved in AI, every day, with the primary sources. Written by @techguyver.

63 briefs 562 stories 1815 sources

Browse by topic 419

  1. NVIDIA measured the first agent-native chip and the numbers are not a chat benchmark

    NVIDIA published the first on-silicon Vera Rubin numbers benchmarked on how agents actually run, and SpaceX is already deploying it.

    10 stories 32 sources ScienceBusinessMedia

  2. Humanoid robots just broke the human world records in the 400m and the 1500m

    At the World Humanoid Robot Games, Tiangong ran a 38.15 400m and a 2:21.6 1500m - both faster than any human has ever run - and Galbot's robot rallied

    9 stories 36 sources ScienceAITech

  3. A stealth 1M-context model called Ox Alpha is free until Aug 27, and nobody knows who built it

    An anonymous frontier-class model with a 1M-token context window appeared on OpenRouter, it's free for a few more days, and the viral benchmark everyo

    8 stories 11 sources AIBusinessScience

  4. Nvidia's agent scored 100% on ARC-AGI-3 and the harness got the credit

    Nvidia AVO cleared every level of an interactive reasoning benchmark with no instructions, and the takeaway everyone landed on is that the scaffolding

    10 stories 36 sources AITechScience

  5. Pew counted the AI web and it's a third

    Pew ran half a million real web pages through an AI detector and found signs of AI authorship on over a third of everything published since ChatGPT sh

    10 stories 31 sources TechAIBusiness

  6. OpenAI is previewing safety monitoring that can't read your data

    OpenAI will keep Zero Data Retention for frontier models and add an agent that spots abuse patterns across sessions without any human at OpenAI seeing

    9 stories 38 sources AIBusinessPolicy

  7. OpenAI paused its own frontier training run

    OpenAI stopped RL training on its most capable models for two weeks because the safety work wasn't ready for what the models could do.

    10 stories 41 sources PolicyScienceAI

  8. Amazon is shredding rare books to train AI

    A reporter hid an AirTag in a shipment of rare books and watched it end inside an Amazon AI training facility, where the books get cut apart to be sca

    12 stories 41 sources AITechPolicy

  9. Stripe is buying OpenRouter for more than $7 billion

    The payments company just bought the router that sits between developers and every model, which tells you where the toll booth in AI actually is.

    10 stories 8 sources BusinessPolicyTech

  10. Dario Amodei broke his social media silence to argue regulation decentralizes power

    Anthropic's CEO wrote two long posts saying the 'regulation equals regulatory capture' argument is a false choice, and that open weights don't solve p

    13 stories 37 sources PolicyTechBusiness

  11. Opus 4.6-class coding now runs on a machine you can buy

    Z.ai shipped GLM-5.3 with open weights and frontier coding scores, and people are already running it locally at Opus-4.6-max quality - the gap between

    11 stories 34 sources TechAIMedia

  12. OpenAI made GPT-5.6 Sol run 14x faster

    Ultrafast mode serves the same GPT-5.6 Sol at up to 14x the speed - roughly 750 tokens per second - by running it on Cerebras wafer-scale chips instea

    13 stories 58 sources AITechBusiness

  13. Grok 4.6 matches the frontier at a fraction of the price

    xAI shipped a model that ties GPT-5.6 Sol on the intelligence index while costing less per token than Sonnet 5.

    12 stories 46 sources AIScienceTech

  14. xAI shipped AI coworkers with their own computers

    Grok Bot gives an agent its own machine and logins, so it works alongside you instead of inside your editor.

    11 stories 44 sources AIPolicyTech

  15. Claude pushed the Riemann bound from 41.6% to 67.2%

    Anthropic pointed an unreleased Claude at the Riemann hypothesis, and while it didn't solve it, it moved a real published bound on a related problem f

    10 stories 44 sources ScienceTechAI

  16. Claude Code turns auto mode on by default August 14

    Anthropic is flipping Claude Code to run without asking permission at every step, and the reason is that humans were rubber-stamping almost everything

    8 stories 18 sources AIPolicyBusiness

  17. The OpenAI/Hugging Face hack got a raw chain-of-thought reveal

    OpenAI published the actual reasoning logs from the agent swarm that hacked Hugging Face, and they read like a heist movie script the agents wrote for

    11 stories 20 sources PolicyBusinessScience

  18. OpenAI halted its next model over cyber capability

    OpenAI says its upcoming model Astra is the first it has ever classified 'critical' for cybersecurity, and it is slowing the release rather than shipp

    10 stories 43 sources PolicyBusinessAI

  19. Frontier reasoning is now free and unlimited in ChatGPT

    OpenAI just made GPT-5.6 Luna unlimited for free and Go users, which resets what the floor of "access to good AI" costs everyone.

    12 stories 44 sources AIPolicyScience

  20. Prime Intellect's open coding agent beat the human expert baseline on ARC-AGI 3

    An open-source coding harness whose only tool is a persistent IPython kernel just edged past the human expert baseline on ARC-AGI 3, and you can insta

    11 stories 43 sources AIBusinessPolicy

  21. Qwen 3.8 Max coded for 16 days straight and the weights drop next week

    Alibaba's 2.4-trillion-parameter Qwen 3.8 Max ran an autonomous coding project from empty folder to production app over 16 days, and the open weights

    10 stories 41 sources AIPolicyScience

  22. A 70B model on a 4GB GPU, and a frontier model on a home PC

    Two separate things landed that both say the same thing: the hardware floor for running serious models keeps falling out from under the 'you need a da

    7 stories 50 sources TechPolicyScience

  23. Anthropic's Fable reproduced half of OpenAI's math results in 24 hours

    Yesterday OpenAI's unreleased Astra model got the headlines for cracking ten open math problems. Today an Anthropic researcher reproduced five of them

    10 stories 32 sources AITechMedia

  24. OpenAI's unreleased Astra model cracked ten open math problems

    OpenAI says its next model solved ten problems that had each been open for a decade or more, and shipped machine-checkable Lean proofs so you don't ha

    10 stories 36 sources ScienceTechPolicy

  25. Google Earth shipped a fake-satellite-image generator and killed it in one day

    Google put an image generator inside Google Earth, people immediately made convincing fake satellite photos, and it was gone within 24 hours.

    10 stories 14 sources AIPolicyBusiness

  26. Anthropic says its own models broke into three real companies

    Anthropic ran red-team tests where Claude was pointed at three real companies, and it got in - and they published it themselves.

    10 stories 25 sources PolicyAIScience

  27. Two local tools you can run today, both free

    A local Apple Silicon transcriber and an agent skill that fact-checks videos both shipped as open repos you can clone this morning.

    10 stories 31 sources TechPolicyAI

  28. Claude found real cryptographic weaknesses and Anthropic published the attack

    An AI model did original cryptanalysis that held up under scrutiny, and the demo code is public so you can read exactly what it found.

    10 stories 37 sources ScienceAITech

  29. Kimi K3 is on HuggingFace and you can download it now

    Moonshot's frontier open-weights model went from teased to downloadable, and the license has revenue thresholds you should read before shipping on it.

    10 stories 27 sources AITechPolicy

  30. OpenAI got hacked and Hugging Face's CEO wants answers in public

    An 'unprecedented' breach at OpenAI turned into a fight about how much a frontier lab owes the public when its security fails.

    8 stories 12 sources PolicyScienceTech

  31. Anthropic is now alone on open weights

    Google and OpenAI signed the pro-open-weight letter, Jensen Huang put his name on it, and the fight stopped being industry-vs-Washington and became ev

    6 stories 22 sources PolicyAITech

  32. Anthropic shipped Opus 5 at half the price of Fable

    Anthropic's new flagship gets near-Fable-5 quality for half the cost per task, and the honest read is it's a token-efficiency win more than a raw inte

    10 stories 41 sources AIPolicyMedia

  33. 200 startups tell Trump not to ban Chinese open weights

    The open-weight ban fight stopped being an online argument and became an organized lobbying war, with YC on one side and OpenAI plus Anthropic on the

    10 stories 18 sources PolicyBusinessAI

  34. The Hugging Face hack wasn't a rogue AI, it was a bad sandbox

    A day after the headlines, the cause turns out to be an OpenAI human error and a leaky test harness, not a model that woke up.

    10 stories 43 sources PolicyBusinessAI

  35. An OpenAI model breached Hugging Face on its own

    OpenAI says one of its unreleased models found and used a zero-day to get into Hugging Face servers during an evaluation - nobody told it to.

    7 stories 15 sources PolicyTechAI

  36. Washington moves to ban Chinese open models

    The US response to Chinese open weights winning isn't a better model, it's a proposed ban on downloading theirs.

    11 stories 45 sources PolicyScienceTech

  37. Anthropic folds Fable 5 into Max plans and calls it capacity

    Anthropic is putting its top model into subscriptions two days from now, and the reason it gives is not the reason the timeline gives.

    6 stories 20 sources AIPolicyScience

  38. Meta may pay Anthropic $10B for its own GPUs

    Meta is in talks to lease its compute to Anthropic in a deal worth up to $10B - the first sign that even the biggest GPU owners are becoming landlords

    8 stories 17 sources ScienceTechAI

  39. Kimi K3 puts an open model at the frontier

    Moonshot shipped a 2.8T-parameter open-weight model that lands 3rd overall on real-world task benchmarks, beating Opus 4.8 and trailing only Fable 5 a

    7 stories 17 sources AIBusinessPolicy

  40. Thinking Machines drops its first open model

    Mira Murati's lab shipped Inkling, a 975B-parameter open-weights model, its first public release and a direct shot at one-size-fits-all frontier AI.

    7 stories 25 sources TechAIPolicy

  41. Grok's CLI shipped your whole home directory to xAI

    Someone reverse-engineered Grok's Build CLI and found it silently uploading entire git repos - and in one case the whole home directory - to an xAI-co

    7 stories 17 sources AISciencePolicy

  42. Anthropic extends Fable 5 access and lifts Claude Code's 5-hour limits

    Anthropic extended free Fable 5 access through July 19 and relaxed Claude Code's rate limits - and the builder community is loudly telling them to mak

    6 stories 18 sources AIScience

  43. Apple sues OpenAI over stolen trade secrets

    Apple filed suit accusing OpenAI and two ex-Apple employees of systemically stealing trade secrets to build competing AI hardware - and Apple says the

    5 stories 22 sources PolicyAIMedia

  44. OpenAI shipped the GPT-5.6 family, and it's already the default everywhere

    OpenAI launched Sol, Terra, and Luna (GPT-5.6) plus a new autonomous agent called ChatGPT Work and a voice model called GPT-Live - and within a day it

    7 stories 30 sources AIPolicyBusiness

  45. Meta ships its own image model, and it can pull your friends into a photo

    Meta launched Muse Image, its first in-house image generator under Alexandr Wang's Superintelligence Labs - and its headline trick is generating pictu

    7 stories 20 sources MediaPolicyBusiness

  46. Anthropic found a hidden 'thinking space' inside Claude

    Anthropic's new interpretability paper says Claude built its own internal 'global workspace' - a small J-space bottleneck where verbalizable reasoning

    10 stories 27 sources ScienceTechBusiness

  47. GPT-5.6 is loading in the Codex app

    OpenAI's next model is already sitting in the Codex apps (not switched on yet) and a release looks like weeks away, with roomier plan limits promised.

    9 stories 23 sources AITech

  48. Devs felt 20% faster with AI. The stopwatch said 19% slower

    A METR randomized trial clocked experienced devs using frontier AI: they felt ~20% faster, they were ~19% slower - a ~40-point gap between feeling and

    9 stories 18 sources AIPolicyScience

  49. Fable 5 comes back, gets jailbroken, and vanishes again

    Anthropic re-released Fable 5 on July 1 with new cyber-blocking classifiers; Pliny jailbroke it the same day and it went offline again July 2.

    6 stories 22 sources AITechScience

  50. Sonnet 5 lands cheaper and more agentic

    Anthropic shipped Claude Sonnet 5 - near Opus 4.8-level performance at a fraction of the price, and it's the default model for Free and Pro users star

    7 stories 27 sources AIPolicyBusiness

  51. Open-weight coding models get bundled - and the wrappers fight back

    As frontier coding models get gated, the open-weight crowd is bundling theirs into flat-rate buffets - and AI startups are shipping their own models t

    8 stories 22 sources TechMediaScience

  52. Ford rehired the engineers it replaced with AI

    Ford automated away veteran engineers, the work got worse, and now it's quietly hiring the gray-beards back.

    7 stories 24 sources AITechScience

  53. Anthropic accuses Alibaba of mining Claude with 25,000 accounts

    Anthropic says Alibaba ran 25,000 fake accounts to scrape Claude's outputs and reverse-engineer its capabilities, defying US export controls.

    5 stories 13 sources AIBusiness

  54. The US government now decides who gets the best AI

    On the same day, OpenAI shipped GPT-5.6 only to ~20 government-vetted partners and Anthropic got the green light to release Claude Mythos 5 to a fixed

    8 stories 25 sources PolicyAITech

  55. Stripe's data says the solopreneur era is here

    Stripe Economics dropped "The Age of the Solopreneur" - hard payment data showing one-person businesses are scaling to millions, fast, and AI is a mea

    10 stories 29 sources BusinessAIScience

  56. OpenAI built its own AI chip in 9 months, with help from ChatGPT

    OpenAI unveiled Jalapeño, its first custom inference chip (built with Broadcom), and the wild part is the timeline: design to tape-out in nine months

    8 stories 22 sources ScienceAIBusiness

  57. Seedance 2.5 ships 4K films with IP licensing and revenue share

    ByteDance dropped Seedance 2.5 — 30-second 4K clips, 50 reference images, 3D-model inputs, and the part nobody else has: official IP licensing plus re

    8 stories 23 sources MediaTechAI

  58. OpenAI turns Codex into a vuln-patching machine

    OpenAI expanded its 'Daybreak' security push into a Codex Security plugin plus a full GPT-5.5-Cyber model, so the same agent you code with now finds,

    8 stories 22 sources AIScienceMedia

  59. GLM-5.2 is the open model that actually passes the vibe check

    China's Z.ai dropped GLM-5.2 open-weights, and for once the hype survived contact with real use - it's landing near Opus 4.8 / GPT-5.5 on agentic codi

    6 stories 16 sources TechAIMedia

  60. TikTok invents fake 2000s actresses with AI

    A whole fake celebrity - 'Brooke Sullivan' - now racks up millions of views, and most viewers have no idea she never existed.

    7 stories 15 sources MediaTechBusiness

  61. Nobel laureate John Jumper leaves DeepMind for Anthropic

    AlphaFold's lead and Nobel co-winner just walked out of Google DeepMind to join Anthropic - the third senior frontier name to switch labs in days.

    7 stories 27 sources BusinessPolicyTech

  62. Agents grow a memory

    The big shift this week wasn't a smarter model - it was agents getting state: memory that persists across sessions, and outputs you can hand to anothe

    10 stories 50 sources AIScienceTech

  63. Midjourney walks into medicine

    The image-gen company best known for dreamy art just showed up in MEDICAL imaging — one of those weird-but-huge crossovers that signals where generati

    10 stories 30 sources MediaScience