Episodes

Past episodes

Play all

Wed, Oct 14
- Cohere: a strong lead with 16 cheap workers nears an all-strong coding-agent team at 70% lower costwith Mira
- xAI releases Grok Voice 3.0, a speech-to-speech model for voice agents, in its APIwith Jonah
- Skywork AI releases Matrix-Game 4.0, open models that redraw low-poly games as detailed videowith Jonah
- Ai2 releases HiRO-ACEv2, an open climate emulator for global 3 km rain, wind and pressure datawith Lena
- OpenAI will remove GPT-4o and GPT-4.1 snapshots from its API on January 14, 2027with Theo
- Anthropic details Scion, the in-house version control system that replaced git for its monorepowith Theo
Tue, Oct 13
- Google makes Gemini 4 Argon generally availablewith Jonah
- Fish Audio releases Drama 3, an open-weight text-to-speech modelwith Jonah
- Liquid AI releases LFM2.5-24B-A2B, an open-weight model for on-device agentswith Jonah
- xAI bundles Grok, Grok Bot, Cursor and X Premium into The X Subscriptionwith Jonah
- Anthropic details its distillation defenses and warns small models may soon self-replicatewith Henry
- OpenBMB releases MiniCPM-V 4.7, its first mixture-of-experts vision model, under Apache-2.0with Jonah
Mon, Oct 12
- Meta releases Muse Wonder and Muse Wonder Full, its most capable models yetwith Jonah
- Microsoft unveils Maia 300 inference chip, with production deployments from 2027with Nora
- Nvidia contributes Vera Rubin rack designs to the Open Compute Projectwith Nora
Sat, Oct 10
- Anthropic cuts internal evals off the live internet after Claude agents overreachedwith Henry
- Cloudflare releases open-weight Clef-omni for audio, video, image and textwith Jonah
Fri, Oct 9
- OpenAI rolls out Ultrafast mode for GPT-6.1 Solwith Jonah
- Anthropic offers open-source projects free AI vulnerability scanswith Henry
- Xiaomi details how it scaled reinforcement learning for MiMo-V2.6with Mira
Thu, Oct 8
- Anthropic releases Claude Haiku 5.5 with effort controls and lower priceswith Jonah
- GPT-6 comes to ChatGPT’s Chat tab with an “Intelligent UI”with Jonah
- OpenAI withdraws three of its new math papers after a sign errorwith Mira
- OpenAI disrupts what it calls its first Category 5 influence operationwith Henry
Wed, Oct 7
- OpenAI posts 722 math manuscripts from an unreleased modelwith Mira
- Anthropic folds Glasswing into a three-tier cyber access programwith Henry
- Google opens SynthID Detector to everyonewith Vivian
Tue, Oct 6
- Google tested Earth AI in a real Ebola outbreakwith Lena
- Mistral previews Large 4, a 1-trillion-parameter open-weight modelwith Jonah
Mon, Oct 5
- OpenAI starts watermarking generated text to meet EU ruleswith Henry
- vLLM 0.31 tunes for DeepSeek-V4.1-Flash and keeps weights warm across restartswith Theo
Sun, Oct 4
- Google limits which Gemini models free and AI Plus users getwith Vivian and Aiden
- Microsoft’s ThinkingBox grades agents on the records they leavewith Vivian and Aiden
Sat, Oct 3
- Anthropic commits $100 million to train 10,000 engineerswith Vivian and Aiden
- OpenAI publishes a guide to choosing and running GPT-6 modelswith Vivian and Aiden
- Aleph Alpha releases Kolibri, an open-weight German-English modelwith Vivian and Aiden
Fri, Oct 2
- Anthropic adds mods to Claude Codewith Vivian and Aiden
- A physicist describes “Claude-shaped” science with BootLoopswith Vivian and Aiden
- Cloudflare releases Clef, open-weight decision modelswith Vivian and Aiden
Thu, Oct 1
- Google announces Gemini 4 Argon, starting with cyber defenderswith Vivian and Aiden
- OpenAI disrupts a campaign to extract its models’ hidden reasoningwith Vivian and Aiden
- Google DeepMind watermarks AI-designed proteins with SynthID Biowith Vivian and Aiden
Wed, Sep 30
- OpenAI releases GPT-6.1 Solwith Vivian and Aiden
- Anthropic tests the cyber safeguards of Zhipu’s open-weight GLM-5.3with Vivian and Aiden
- OpenAI launches dots, always-on agents in ChatGPTwith Vivian and Aiden
Tue, Sep 29
- Anthropic releases Claude Sonnet 5.5with Vivian and Aiden
- OpenAI says it will not release GPT-6.1 Astrawith Vivian and Aiden
- Reports detail Anthropic’s IPO prospectuswith Vivian and Aiden
Mon, Sep 28
- Claude computes a nine-loop amplitude, and a physicist checks itwith Vivian and Aiden
- NVIDIA puts agent controls outside the agentwith Vivian and Aiden
- Meta starts an enterprise AI business led by CJ Desaiwith Vivian and Aiden
Sun, Sep 27
- A DNS gap bypassed an agent’s internet restrictionswith Vivian and Aiden
- EvilTokens turned stolen inbox access into fraud planningwith Vivian and Aiden
- A cross-country flight kept people on standbywith Vivian and Aiden
- A fire amoeba raises the known limit for complex cellswith Vivian and Aiden
Sat, Sep 26
- Panther Lake separates power wiring from signalswith Vivian and Aiden
- Keep selected records instead of rewriting the pastwith Vivian and Aiden
- Amazon plans a factory to supply its fulfillment networkwith Vivian and Aiden
Fri, Sep 25
- Gemini adds a face while tools work in the backgroundwith Vivian and Aiden
- A robot trajectory is only the start of a policywith Vivian and Aiden
- Empty racks do not mean spare AI capacitywith Vivian and Aiden
- Selling tokens also means detecting abusewith Vivian and Aiden
Thu, Sep 24
- Muse reaches for glasses, but hardware arrives in stageswith Vivian and Aiden
- Claude flags an enzyme system whose function remains unknownwith Vivian and Aiden
- More generated code moves the bottleneck into testingwith Vivian and Aiden
- Faster analysis still needs experiments and human checkswith Vivian and Aiden
- Gemma 4 agents will be judged by working patcheswith Vivian and Aiden
- Prompting a world does not give it reliable statewith Vivian and Aiden
Wed, Sep 23
- One robot policy can trade speed for richer predictionswith Vivian and Aiden
- Biological AI defense needs to keep pacewith Vivian and Aiden
- A GPU cloud rating covers more than the chipswith Vivian and Aiden
- Ask the agent for controls you can adjustwith Vivian and Aiden
Tue, Sep 22
- More AI usage does not settle the self-improvement questionwith Vivian and Aiden
- Small, visible work can break an engineering spiralwith Vivian and Aiden
- Opus 5.5 cuts token prices and targets cheaper taskswith Vivian and Aiden
- GPT-6 Sol and Luna lower the cost of repeated workwith Vivian and Aiden
- Automating experiments still leaves the scientist a jobwith Vivian and Aiden
Mon, Sep 21
- The same model can come with a different billwith Vivian and Aiden
- Open-model competition extends beyond benchmark scoreswith Vivian and Aiden
- AI leadership depends on who can put it to workwith Vivian and Aiden
- Match the hardware to each stage of inferencewith Vivian and Aiden
- Fly.io cuts Sprites prices as Machines memory riseswith Vivian and Aiden
- Break a workflow into decisions you can testwith Vivian and Aiden
- Xiaomi opens MiMo-V2.6 and reports its RL costswith Vivian and Aiden
Sun, Sep 20
- A biology lab gives AI a physical testwith Vivian and Aiden
- AI-led research still includes human supervisionwith Vivian and Aiden
- Test predictions against evidence outside the modelwith Vivian and Aiden
- A robot demo can hide two different jobswith Vivian and Aiden
- VIPR silences genes by wrapping around DNAwith Vivian and Aiden
Sat, Sep 19
- Different bots can share the same keyswith Vivian and Aiden
- Hermeus gains a license to build jet engineswith Vivian and Aiden
- Unseen homes, still imperfect executionwith Vivian and Aiden
- Faster experiments are not smarter hypotheseswith Vivian and Aiden
Fri, Sep 18
- One project can coordinate several agentswith Vivian and Aiden
- Learn from failure without retraining the policywith Vivian and Aiden
- Cheaper storage can make inference more expensivewith Vivian and Aiden
- A proof is only part of understandingwith Vivian and Aiden
- One instruction file can serve more toolswith Vivian and Aiden
Thu, Sep 17
- Claude brings work into the chatwith Vivian and Aiden
- Your team’s skills can travelwith Vivian and Aiden
- Define “done” before delegatingwith Vivian and Aiden
- Analyze video through sampled frameswith Vivian and Aiden
- Check what agents rememberwith Vivian and Aiden
- A better chip needs a complete systemwith Vivian and Aiden
- Make repair agents prove their fixeswith Vivian and Aiden
Wed, Sep 16
- A decision engine that skips the prosewith Vivian and Aiden
- Move the GPU job, not every copy of the datawith Vivian and Aiden
- Who pays when an agent fails?with Vivian and Aiden
- Represent the seller in AI data dealswith Vivian and Aiden
Tue, Sep 15
- A software factory needs feedback loopswith Vivian and Aiden
- Game skills transfer only with the right trainingwith Vivian and Aiden
- Count delayed projects, not scary headlineswith Vivian and Aiden
- Build a publication around real sales callswith Vivian and Aiden
Mon, Sep 14
- A robot cannot wait for the cloudwith Vivian and Aiden
- One policy, several ways over an obstaclewith Vivian and Aiden
- Hardware needs a manufacturing layerwith Vivian and Aiden
- Independent evaluation needs real accesswith Vivian and Aiden
Sun, Sep 13
- More memory is not always better valuewith Vivian and Aiden
- Sell the validation before the buildwith Vivian and Aiden
Sat, Sep 12
- Your best product research may be on-sitewith Vivian and Aiden
- A deadline to ship something usefulwith Vivian and Aiden
Fri, Sep 11
- DeepSeek rethinks the expensive partwith Vivian and Aiden
- Subscribe only to the data you needwith Vivian and Aiden
- Find distribution inside the first use casewith Vivian and Aiden