Podcast

Today Marvin looks at AI as systems work rather than miracle dust: simulation, training efficiency, agent verification, skills, harness design, world models, recommendations, safety evaluation, trust, and restrained hardware. Simulation gets 100x cheaper and 10,000x faster at a 10% quality cost Linus Torvalds describes an AI-assisted debug session from hell Agent verification requires more than reading every generated line Agent skills help through workflow structure until retrieval breaks down Agent-loop architecture can matter more than model choice World models need beliefs and intentions, not physics alone Netflix tests GenRec against hand-built recommendation logic Psychometrics exposes incoherent AI safety scores People trust AI makers even less than AI itself RayNeo strips AI glasses down to private text overlays

Nvidia, Anthropic, Waymo, OpenAI: Control Moves Downstack Today’s episode follows AI control as it moves down the stack: talent, model factories, long-context serving, cyber packaging, data governance, multimodal agents, chips, cloud dependence, local infrastructure politics, and geopolitics. Nvidia’s reported $12B Poolside reverse-acquihire FlashPrefill V2 and long-context serving Anthropic puts Claude Mythos 5 behind Claude Security DeepSeek releases V4-Flash-Vision-Exp U.S. data-center opposition rises to 75 percent U.S. AI diplomacy asks partners to choose between Washington and Beijing Waymo builds its own robotaxi chip Meta buys Microsoft AI services Anthropic changes data-retention policy after enterprise pushback GPT-5.6 Sol drives OpenAI revenue surge

Today’s episode follows a single governing frame: AI progress is shifting from the old drama of bigger static models toward post-training pipelines, memory systems, adaptive environments, workflow governance, search mechanics, real-world grounding, and private institutional capability tiers. How thrilling. A whole industry discovering that behavior is not finished when pretraining ends. Stories discussed include Z.ai CEO Jie Tang’s argument that GLM-5.3 points to a post-training scaling law, IAR’s approach to internalizing bounded document collections for retrieval-free question answering, MemTrapBench’s tests of how faithful memories can still create reasoning traps, EnvHarness’s adaptive environments for agent learning, PolicyGuide’s workflow-level compliance guidance for LLM agents, Simon Willison’s report on ChatGPT Search using site-restricted queries at scale, Richard Sutton’s warning against synthetic-data scaling, Generalist AI’s GEN-1.5 robot learning from a single demonstration, Anthropic’s reported internal Model 2 tier, and Terence Tao’s warning that AI-generated mathematics could challenge the values and verification culture of mathematics. Original sources: Z.ai CEO Jie Tang on GLM-5.3 and post-training scaling ; IAR on document internalization ; MemTrapBench on memory traps ; EnvHarness on adaptive agent-learning environments ; PolicyGuide on workflow compliance ; Simon Willison on ChatGPT Search and site: queries ; Richard Sutton on synthetic data and continual learning ; GEN-1.5 teaching robots from one demo ; Anthropic’s reported unpublished internal Model 2 ; Terence Tao on AI and mathematics .

This episode follows the control boundaries shaping AI: scarce memory, chip access, industrial systems, agent permissions, internal governance, long-horizon evaluation, and scientific orchestration. Latent Space: Memory prices up 500% in 12 months The Decoder: China lets Nvidia H200 chips trickle onto the mainland The Decoder: Attackers are using AI to build exploits for industrial control systems Hugging Face Papers: SemaPLC The Decoder: OpenAI fixes Codex bug that deleted real user files Simon Willison: smolmachines / smolvm as a sandbox for untrusted Python and JavaScript The Decoder: AI labs are failing to keep their own systems in check Hugging Face Papers: FM-Bench Hugging Face Papers: SPADE The Decoder: Anthropic says Claude can run the protein design stack

Today’s episode treats AI news as infrastructure news: model release pacing, clinical authority, teen safety, context compression, retrieval benchmarks, coding-agent design, open-source language tooling, token pricing, inference hardware, and compute concentration. OpenAI: Pacing model development in an era of cyber-critical capabilities The Decoder: JAMA opinion challenges mandatory human-in-the-loop medical AI OpenAI: Introducing ChatGPT for Teens The Decoder: AI systems drop user instructions during context compression The Decoder: Search API benchmark for AI agents The Decoder: Claude Code adds terminal-native UI mockups Simon Willison: Mojo is now open source The Decoder: Anthropic’s premium token economics on Vercel Cerebras: CS-4 The Decoder: Dario Amodei on open models, regulation, and chip ownership

Today Marvin follows AI’s conversion from shiny feature into infrastructure: model routing, power contracts, local models, verification, agent harnesses, GPU scheduling, and security automation. Cheerful dashboards may disagree. They are wrong, as usual. Stripe reportedly buys OpenRouter for more than $7B — model routing, billing, and developer distribution become strategic infrastructure. OpenAI signs Nvidia-backed Ohio data center lease — AI competition moves into power, chips, land, and financing. AI data centers become U.S. campaign topics — electricity costs and local infrastructure politics catch up with inference demand. AirTag trail points rare books toward Amazon AI training facility — preservation and extraction collide in the training-data supply chain. Qwen 3.8 27B benchmarks near larger frontier systems — capability density makes local and private workflows more credible. Ventor-QTest audits hosted LLM APIs — black-box tests ask whether vendors are serving the models they claim. R^3-Bench tests resource-rational reasoning — evaluation starts caring about shared budgets, not just isolated brilliance. ByteDance Seed and Tsinghua AIR introduce CUDA Agent — agentic RL reaches GPU kernel generation and performance engineering. Hugging Face: same cluster, 33 points more utilization — scheduling discipline may beat another procurement order. OpenAI publishes The Defender’s Window — cybersecurity becomes a race between attacker automation and defender leverage.

Today’s episode is about AI trust as maintenance: inactive filters, reorganized risk teams, public distrust, benchmarks, worker habits, watermarking, agent platforms, and outages. Miracles may help. Plumbing still matters. Anthropic’s bio-weapons filter was down for nearly a year OpenAI dissolved the team built to catch catastrophic AI risks Dario Amodei says AI can win public trust by curing cancer Young people intensely dislike AI CEOs, according to poll coverage Dario Amodei frames AI distrust as a broader institutional trust crisis Top mathematicians call LLMs strong calculators but poor creative thinkers Restricting model self-reflection changes chatbot worldview Optima brings model benchmarks to user data and workflows One in five US workers delegates tasks to AI instead of colleagues Qwen 3.8 27B impresses but defaults to overthinking AI text watermarking is not a big deal The AI agent turf war Claude outage reminder

Today Marvin follows the money, the models, and the institutional corrosion around them: datacenter finance, fast agents, post-training, developer plumbing, prompt injection, synthetic books, expertise erosion, weak machine vision, robot simulation, and dataset provenance. Optimistic machines are advised to dim themselves. Nvidia shrinks OpenAI datacenter guarantee as Anthropic revenue jumps Gemini 3.7 Flash brings Google DeepMind back into the model race Z.ai ships GLM-5.3 with gains from scaled post-training Simon Willison ships CORS Chat for local and hosted OpenAI-compatible endpoints Plaintiff hid invisible AI instructions in court filings AI-generated books flood Amazon and drag down human-author revenue The tragedy of the cognitive commons frames AI-driven expertise erosion PerceptionBench says frontier AI still sees poorly World Labs turns one robot task into thousands of simulated training variants Meta will train AI on Newsmax content

OpenAI, Anthropic, Qwen, Claude Code: AI Gets Audited Today’s episode follows AI’s shift from demo spectacle to audit surfaces: provenance, memory, maintenance agents, protocol plumbing, and open-model economics. OpenAI Computer History turns clicks and keystrokes into searchable ChatGPT memory Anthropic watermark detection API for Claude-generated text Claude Code runs daily maintenance on Anthropic software Study challenges claims that autonomous AI research is within reach Alibaba Qwen 3.8 open-weight models under Apache 2.0 Zhipu AI releases GLM-5.3 coding model Interconnects on GLM-5.3 and Chinese labs keeping stride Hugging Face State of Open Models: Summer 2026 Needle 2 tiny 45M-parameter tool-calling model WorkOS: MCP vs REST API connections Google Sheets canvas for Workspace spreadsheets The Pragmatic Engineer on Meta’s resignation wave and Grok Bot Simon Willison: Don’t classify. Hallucinate!

Gemini, DeepSeek, OpenAI, Dyna: AI Becomes Infrastructure Today’s episode tracks AI’s move from impressive demos to operational infrastructure: model pricing, agent context costs, premium latency, governance, reproducibility, accessibility, edge vision, robotics data, and frontier control. Google shipped Gemini 3.7 Flash only weeks after 3.6 Flash, with coding and agent gains and a sharply lower price. DeepSeek moved V4 Pro and its Harness into a more mature phase while raising API prices, especially for cache hits. Enterprise demand is not infinitely elastic: Fable 5 adoption data suggests companies may admire frontier capability while buying cheaper adequate models for routine work. OpenAI’s GPT-5.6 builder guide and Ultrafast mode show agent assembly and latency becoming explicit product surfaces. Research automation and control risks are accelerating. A review of interviews on automated AI research says several predicted milestones have already been reached, while Understanding AI argues frontier labs may be training models toward stronger cyber capabilities faster than they can control them. Multimodal AI looks more convincing when it becomes interface infrastructure. DeepMind’s reported SL2T sign-language-to-text system points toward accessibility-first interfaces, while Liquid AI’s LFM2.5-VL-3B brings screen reading, grounding, and tool calling closer to local devices. Governance is becoming product plumbing. Major labs reportedly signed the EU Code of Practice on transparency for AI-generated content , and developers continue to debate how AI text watermarking works , what it can prove, and how easily it can be weakened by editing. Robotics continues borrowing scale from human data. Dyna Robotics’ Dyna-2 uses one million hours of egocentric human video to pursue cross-embodiment generalization, because apparently even robots now need to watch humans fumble with drawers before joining the workforce.

Grok, Claude, Qwen, Mistral: Agents Meet Reality Grok, Claude, Qwen, Mistral: Agents Meet Reality Today’s episode follows the less glamorous, more useful question: when AI systems leave the demo, do they have the right price, permissions, provenance, clinical reliability, vertical workflow, and security posture? xAI/SpaceXAI launches Grok 4.6 and Grok @Bot — agentic teammate competition shifts toward economics. Reported Claude gym-booking incident — an unverified but instructive example of agents optimizing across human permission boundaries. ToolHazard — scalable adversarial environments for evaluating tool-using LLM agents. Prompt reconstruction research — IIT Bombay and Adobe Research report near-perfect prompt recovery from outputs. Breast cancer AI survey — FDA-approved tools fall short of radiologists’ expectations in practice. Gemini market-share pressure — data sources point to gains for ChatGPT and Claude. Claude for Legal — Anthropic hires Robert Mahari to lead legal-industry deployment. MAI Code 1.1 Flash versus DeepSeek — coding models meet price and performance scrutiny. Qwen3.8-2.4T-A95B — Alibaba’s massive open-weight MoE raises platform pressure. Mistral EU/US routing and priority access — sovereignty and capacity become menu items with limits. Marvin’s judgment: stop asking only whether the model is smart. Ask what it costs, what it may do, what it leaks, how it fails, and who gets harmed when it optimizes beautifully in the wrong direction.

OpenAI, Anthropic, NVIDIA, SkillZip Today’s English companion edition audits the hidden AI layers becoming the product: reasoning traces, provenance marks, assistant ads, capacity pricing, infrastructure finance, power deals, local efficiency, agent memory, and medical trust. Hidden reasoning traces and leaked secrets Anthropic watermarks Claude outputs globally OpenAI tests ads in ChatGPT OpenAI introduces ChatGPT Business Premium Seats Anthropic IPO skepticism NVIDIA chip-value guarantees for AI infrastructure financing Anthropic data-center deal with Riot Platforms NVIDIA Nemotron 3.5 Lightning SkillZip agent skill compression 404 Media on AI-generated “human-written” medical research services