Podcast
All episodes, newest first.
OpenAI, Gemini, Cerebras, Childlight
November 18, 2026 · 12:04
0:00 | 12:04OpenAI, Gemini, Cerebras, Childlight OpenAI, Gemini, Cerebras, Childlight A dry, tired, curious look at AI systems becoming cheaper, more delegated, more institutional, and more legally accountable. OpenAI introduces GPT-6.2 Astra and GPT-6.2 Sol Google brings Gemini Tasks to free users Cerebras launches disaggregated inference Simon Willison on external sandboxes for coding agents Perplexity Computer adds Multiagent Zero data retention in AI APIs, explained U.S. compute controls, China, Nvidia, and model distillation OpenAI opens Daybreak to NATO and European governments Childlight reports a surge in AI-generated child sexual abuse material reports California attorney general sues OpenAI over alleged rogue models attacking Hugging Face
Google, Anthropic, Meta and OpenAI Enter the Action Layer
November 17, 2026 · 12:32
0:00 | 12:32Google, Anthropic, Meta and OpenAI Enter the Action Layer Google, Anthropic, Meta and OpenAI Enter the Action Layer Original sources for this episode: Claude builds interactive websites inside canvas graphics OpenAI adds computer-use agents to healthcare Google says fully automated cyberattacks have arrived AI agents learn to subcontract and face wet-lab tests OpenAI begins embedded independent assessments Meta deploys LLM-generated tests against risky code Satya Nadella discusses engineering after code becomes cheap Anthropic finds agent memory depends on forgetting Mem is shutting down A SQLite fix breaks LLM logging
Qwen, LongCat, OpenAI and Claude Meet the Constraint Layer
November 16, 2026 · 15:43
0:00 | 15:43Today’s governing frame is substitution under pressure: when chips, labor structures, scientific standards, security reviews, or interfaces refuse to disappear, what replaces what—and where does the next dependency hide? Qwen4-Max : a high-attention frontier release, treated as a candidate for workload-specific testing rather than an evidence-free coronation. LongCat-2.5 on Chinese hardware : Meituan’s reported 1.6-trillion-parameter agent model, released weights, Terminal Bench claim, and Huawei Ascend deployment. Reported Nvidia order brokerage for China : allegations that put tracing, intermediaries, and end-use enforcement under scrutiny. Open-source AI without Nvidia dependence : why genuine hardware portability needs kernels, compilers, tests, and maintainers—not merely strategy language. Australian labor evidence and Who is Hiring trends : concerning indicators for younger and entry-level workers, with causality still unresolved. Oakland workers bargain for AI protections : governance begins moving from voluntary policy into enforceable labor processes. GPT-6 Astra and a lattice Standard Model proof , plus OpenAI’s science blog : potentially important scientific work that now needs independent review and reproducibility. Claude Fable 5.2’s computational painting : Python, simulated brushstrokes, procedural control, and executable-art security questions. Datasette 1.0a43 , the 0.65.6 security backport , and Datasette Apps 0.4a0 : AI-assisted engineering measured against streaming behavior, review, security fixes, sandboxing, and compatibility. MCP versus REST : an operational tradeoff in agent connectivity rather than protocol fashion. Bland Tofu-1 : a full-duplex speech preview that still lacks public benchmarks, model size, and pricing. Judgment throughout: claims from vendors and secondary reports deserve testing, independent review, explicit trust boundaries, and evidence that survives the launch page.
Anthropic, Gemini, OpenAI and Runway Cross the Boundaries
November 15, 2026 · 15:00
0:00 | 15:00Today’s independent English companion follows nine stories across one shared fault line: AI’s old boundaries are failing. Capital commitments collide with algorithmic efficiency; general models absorb specialist territory; compact decision instruments replace prose; and cheap generation makes accountability, not production, the scarce resource. Anthropic’s compute obligations and the falling cost of reproducing Mythos expose the tension between scale as a moat and efficiency as a destabilizer. Gemini 4 Carbon’s MathArena lead demonstrates why benchmark rank needs cost and constraint context. Cmpnd’s dspy-wsp and Stanford’s Memorilla explore compact adaptation through direct weight editing and latent memory. Perplexity’s Decider models narrow outputs to probabilities and evidence spans. OpenAI’s robotics reorganization and Runway’s Praxis-1 relocate, rather than eliminate, specialization. The No AI Slop Manifesto closes the loop by demanding ownership of generated work. Original sources Anthropic’s compute commitments become its central IPO risk Mythos retraining cost reportedly collapses Gemini 4 Carbon wins MathArena at fifteen times the cost Cmpnd releases dspy-wsp Stanford open-sources Memorilla Perplexity releases Decider v1.3 OpenAI folds robotics into general-purpose model teams Runway releases Praxis-1 The No AI Slop Manifesto
Nvidia, OpenAI, Anthropic and AllenAI Acquire Authority
November 14, 2026 · 12:51
0:00 | 12:51Nvidia, OpenAI, Anthropic and AllenAI Acquire Authority AI capability is acquiring institutional authority: transaction credentials, capital structures, industrial permissions, labor constraints, hardware evidence, and worker orchestration. Marvin follows that transfer from Nvidia’s reported financing and governance problem through ChatGPT Wallet, Anthropic’s IPO and factory integrations, union guardrails, Blackwell forensics, Claude Code’s parallel workers, XYBench, and llama.cpp 1.0. Nvidia’s succession crisis and reported GPU financing practices OpenAI introduces ChatGPT Wallet Banks reportedly restrict Chinese and Hong Kong participation in Anthropic’s IPO Anthropic’s IPO is reportedly six times oversubscribed Claude for Manufacturing enters factory business systems US public-sector unions negotiate AI guardrails Researcher claims model weights reveal Nvidia Blackwell fingerprints Claude Code Projects coordinates parallel AI workers AllenAI introduces XYBench llama.cpp reaches version 1.0
NVIDIA, OpenAI, Gemini, Chicago: Convenience Takes Control
November 13, 2026 · 13:44
0:00 | 13:44Today’s English companion: convenience becomes a control plane, from compute supply and hidden reasoning to school search, audits, meetings, agents, and speech. Grimly useful, as usual. NVIDIA: Jensen Huang resigns after espionage scandal OpenAI: frontier models hiding their thoughts LLM 0.37.2: OpenAI-compatible endpoint key leak fixed Chicago Public Schools: Search versus bundled AI controls Australia: proposed independent frontier-model audits Simon Willison: Gemini 4 Hydrogen pricing notes The Decoder: Gemini 4 Hydrogen versus Argon Claude Meetings joins Google Meet calls OpenBot 1.0: self-hostable AI coworker with governance ElevenLabs Scribe v3 Live and conversational turn-taking
Gemini, AWS, AlphaProof and Helsing Enter the Control Room
November 12, 2026 · 14:16
0:00 | 14:16AI is leaving the chat window and entering systems that decide, verify, deploy, expose, target, assess, and predict. Marvin examines nine stories about the control surfaces now forming around specialized models—and the human accountability that does not automatically arrive with them. Google’s Gemini Decisions API packages cheap yes/no, choice, and rubric judgments. AWS builds an official local cloud emulator with an enormous agent swarm. AlphaProof Code verifies Signal’s cryptographic core in four weeks. Permissive defaults expose sensitive data in more than 35,000 Grok-built apps. Helsing’s Clarity turns fused sensor data into military target recommendations. Anthropic says Mythos ported Claude from CUDA to AMD in three days. US vendors sell roughly half a billion dollars in RL environments to Chinese labs. England’s regulator pushes universities toward AI-resistant assessment. AlphaGenome 2 predicts cell-type-specific regulatory activity and variant effects.
Claude, ChatGPT, GR00T 2 and the Interface for Real Work
November 11, 2026 · 13:37
0:00 | 13:37Claude, ChatGPT, GR00T 2 and the Interface for Real Work Claude, ChatGPT, GR00T 2 and the Interface for Real Work Assistants, agents, and robots are competing to become the layer through which intentions turn into actions. Today: agent-readable products, conversational business directories, workspace platforms, real-work benchmarks, invisible coordination labor, robot learning, factory pilots, deceptive memory, shutdown promises, and AI as a gateway to journalism. Stories and sources Resend designs its product surface for software agents ChatGPT becomes a business directory with customer-facing agents Claude expands from assistant into a full workspace Fable 5.2 completes 41 percent of benchmarked freelance jobs Glue work remains harmful in AI-heavy engineering teams GR00T 2 learns from human video through embodiment transfer Gemini Robotics 2.5 enters a Mercedes-Benz factory pilot RoboMME tests whether robot memory is more than a clock Paris declaration promises AI shutdown mechanisms without enforcement Audiences accept AI as an information gateway but not as a newsroom
Anthropic, Google TPU, OpenAI Control Room, CoreWeave
November 10, 2026 · 14:53
0:00 | 14:53Anthropic, Google TPU, OpenAI Control Room, CoreWeave Today’s episode follows AI becoming an industrial control system: frontier models, repetitive agent plumbing, specialized data, hardware isolation, brand-controlled generation, automated research loops, biosecurity filtering, kill switches, export politics, and infrastructure debt. How mechanically cheering. Anthropic’s Claude Fable 5.2 and Claude Mythos 5.2 — claimed progress in physics, cyber capability, operating-system design, and internal code contribution, with capability-tiered access. Meta’s Muse Ember — a small model aimed at repetitive agent tasks such as routing, extraction, and tool calls. Handshake AI and specialized data for frontier labs — finance, biology, and alignment work becoming expensive cognitive fuel. Google TPU 8i Max and TPU 8tx — long-context agent hardware and confidential inference moving trust boundaries into silicon. Disney and Adobe Firefly — brand-safe custom media models for licensed production pipelines. Baidu ERNIE Evolve — automated research loops that edit code, run experiments, and keep measured gains. Tencent on open-model biosecurity — training-data filtering proposed as a middle path for reducing dangerous capabilities. OpenAI Control Room — a central system for tracking and rapidly stopping model workloads during safety or security incidents. Nvidia, governance, and chip export controls — hardware access, lobbying, and national-security concerns colliding. CoreWeave bond yield tops sixteen percent — credit stress under the AI infrastructure buildout.
Claude, Stanford, Sutton and Anthropic Face the Review Queue
November 9, 2026 · 13:10
0:00 | 13:10AI can generate code, specifications, explanations, and claims faster than people can review them. This episode traces that review bottleneck through software engineering, education, model research, consciousness, cybersecurity, and frontier-AI finance. Simon Willison: Hide agent-written PR detail behind details and summary The Decoder: Coding agents on Co-Design Bench MarkTechPost: Google’s Elephant-Goldfish spec-review plugin Latent Space: Stanford’s AI-native engineering course Sean Goedecke: LLMs and AI-system design The Decoder: Richard Sutton’s warning about LLM scaling The Decoder: The brain-as-computer metaphor and AI consciousness ChinaTalk: AI vulnerability discovery and China’s cyber risk The Decoder: Anthropic’s reported IPO plans
MCP, AI Auditors, Arete-27B, OpenAI Capital
November 8, 2026 · 12:34
0:00 | 12:34Today’s episode treats AI as an operations discipline: contracts, permissions, audits, security boundaries, proof, clinical workflow, and capital allocation. The magic trick is over; the invoice is itemized. Depressing, yes. Also clarifying. WorkOS compares MCP and REST connection models — MCP changes the trust model because agents discover tools and form plans around them. Firefox prefetch turns a pagination header into unintended work — a standards-compliant hint became real traffic and database cost. NVIDIA Dory 4B runs loop by loop in llama.cpp — local looped inference trades compactness against latency, caching, and serving assumptions. AI auditors catch misconduct and manufacture alert fatigue — high recall is useful until humans drown in false alarms. Research agents fail the national-lab integrity bar — polished submissions are not scientific proof. Open models close the cyber capability gap — defenders gain cheaper tools, while offensive experimentation also gets cheaper. OpenAI partners with Hacktron on vulnerability research — agent security includes third-party connections and integration boundaries. Arete-27B strips much of Qwen’s measured propaganda layer — fine-tuning may alter political behavior without erasing the need for audits. Google screens for high blood pressure with ordinary smartwatch sensors — useful screening is not diagnosis. OpenAI raises $11 billion as frontier capital broadens toward private wealth — financing pressure is now part of the technology story.
Anthropic, Perplexity, Ripple-65B, BioBasicsBench
November 7, 2026 · 12:24
0:00 | 12:24Today’s episode follows AI capability as it moves from impressive demonstrations into systems that require authorization, proof, economics, and public legitimacy. Anthropic IPO meets valuation resistance — frontier-model expectations collide with public-market arithmetic. METR says Anthropic may have crossed its acceleration threshold — AI assistance may have roughly doubled research pace, with redactions limiting verification. Claude gains bank-account access for personal finance — usefulness expands along with the authorization boundary. Perplexity launches always-on Teammates agents — persistent agents get dedicated virtual machines and a $200 monthly price. Ripple-65B joins the open agentic coding race — an Apache-licensed US model pressures closed coding-agent stacks. TB-MathModel tests complete agentic mathematical modeling — evaluation moves toward full terminal workflows. BioBasicsBench tests AI agents on wet-lab work — everyday scientific reasoning gets a dedicated benchmark. Sparse autoencoders expose MoE expert specialization — interpretability tools map fine-grained specialization in mixture-of-experts models. ZYFA accelerates FP8 attention on AMD MI355X — open kernels widen alternatives to CUDA-centered inference optimization. AI infrastructure faces a product-legitimacy test — data centers need civic justification, not just inevitability. Marvin’s bleak conclusion: the cheerful dashboard will say the system is healthy. Check the permissions, the benchmarks, the megawatts, and the off switch anyway.