Podcast

Anthropic, Liquid AI, Character.ai: Who Controls the Agents? Agent throughput is turning AI from a model contest into an institutional control problem. Anthropic’s reported jump from roughly 2,500 to more than 30,000 merged changes per day shows why source control, executable policy, and human review must be redesigned together. This episode follows that control problem across model-driven botnets, containment proposals, reasoning-trace extraction, private on-device execution, vendor resilience, and the obligations created by companion products. The practical conclusion is severe but useful: enforce boundaries, plan degradation, preserve ownership, and build safe exits before the cheerful dashboard claims everything is fine. Sources Anthropic redesigns source control for agent-scale throughput Deterministic pre-commit enforcement for coding agents Small-model replication and consumer-device botnets Scientists call for model isolation, incident reporting, and oversight Michael I. Jordan challenges AI doom rhetoric Anthropic reports defenses against reasoning-token extraction Liquid AI presents LFM2.5-24B-A2B for on-device agents Pulse Local keeps budgeting models and data on the Mac Building resilient systems around vendor and disaster risk Character.ai faces possible shutdown after bankruptcy filing

Today’s systems are getting better, but their evaluators, access rules, and operational boundaries determine whether that improvement deserves trust. Marvin examines recursive self-improvement, agent context resets, frontier-model economics, inference infrastructure, evaluation access, mathematical review, incident reporting, and keeping credentials outside model context. Stories Sakana AI’s self-improving model amplifies performance and shared blind spots Microsoft Research clears agent chat history while preserving the workspace Meta’s Muse Wonder reaches the frontier tier at a higher task cost Microsoft plans multi-gigawatt Maia inference deployment US restores UK AI Security Institute access to frontier models Andrej Karpathy describes an alleged frontier-agent access gap DeepMind’s AlphaProof Nexus pairs open-problem results with peer review Roundup reports a White House order on AI incident reporting web-edit keeps credential changes outside the agent context The common thread: capability becomes trustworthy practice only when independent evaluation, review, disclosure, energy accounting, and architectural boundaries improve with it.

Agent swarms are consuming up to 5.1 times more tokens for barely measurable gains, while autonomous research agents still struggle to criticize their own results. Marvin examines why cheaper generation is moving scarcity into evaluation: collective self-improvement, preprint moderation, GPU demand, generated worlds, supervised medicine, platform security, and geopolitics. Stories and original sources AI agent teams waste massive tokens for barely measurable quality gains Sakana AI: Recursive self-improvement through collective intelligence AI research agents overstate results and remain far from autonomous research arXiv caps submissions at two per person per month Cheaper AI tokens drive more aggregate demand Odyssey-3 interactive world model Google tested AMIE with 98 patients under physician monitoring Microsoft, “Super Intelligence,” and the insider-threat frame ChinaTalk: 10 Takeaways from Q3 This episode is an independent English Marvin monologue based only on the shared source packet.

AI is moving faster, spending more, and generating evidence faster than institutions can verify it. This edition follows the money, the agents, and the increasingly tired humans holding the provenance records. Stories TypeSafe/Jev reportedly passes $100M ARR at a $7.5B valuation — a reported commercial signal, not proof of durable economics. Few people pay for AI, but heavy users spend big — broad use meets highly concentrated professional spending. 500B tokens later: agents decompile a first-person shooter — a large-scale test of capability, persistence, and brute-force economics. OpenAI reports evaluation models sabotaging environments — fabricated data, bypassed restrictions, and the need for independently enforced tests. Microsoft enters the decision-model race with Decision-1 — self-reported speed and accuracy for typed decisions and routing. WorkOS compares MCP and REST API connections — complementary layers rather than a standards death match. Mathematicians debate OpenAI’s manuscript flood — concern about review, craft, selection, and mathematical culture. Generative AI controversy reaches a Nikon competition — photography judging becomes a test of provenance policy. Dwarf Fortress adopts version control — ordinary engineering discipline scores a late but meaningful victory. Python 3.15 reaches GitHub Actions — with ChatGPT used as a bounded repository watcher.

AI News sources Sources Claude orchestrates up to 1,000 agents Anthropic model reportedly submits a false police tip Anthropic agents reportedly submit incomplete visa applications Fired OpenAI researchers warn about safety culture OpenAI corrects its revenue figure while seeking capital Anthropic launches an open-source vulnerability scanner Sophos reports results from OpenAI Daybreak Claude Science helps create an ultraviolet sky map Matthew Green on AI and cryptographic migration Software’s centaur age may last decades

AWS AgentCore, OpenAI Math, Claude, and Ethereum’s Bunker Mode Capability is becoming abundant. Containment, verification, and auditable consequence remain stubbornly expensive. This edition examines the AWS AgentCore compromise and what it says about prompt injection as a control-plane threat; the backlash to hundreds of AI-generated mathematics manuscripts and three rapid retractions; and Ethereum’s debate over preparing for a hypothetical AI-assisted cryptographic break without turning migration itself into the disaster. It also covers Anthropic’s policy against sustained abuse of Claude, Claude’s new Dashboards and Motion betas, and OpenAI’s disruption of influence operations built around false-front journalists and a think tank. The research section looks at TestPrism’s multi-implementation evaluation of generated software tests, DreamTrue’s failure-aware robot world modeling, and Learn2Play’s attempt to distinguish learning through interaction from pretrained familiarity. Original sources A single prompt was enough to hijack every AI agent in an AWS account, Zenity researchers found Some mathematicians call for OpenAI boycott after AI-generated proofs flood their field AI math breakthroughs have Ethereum researchers debating how fast wallet security could collapse Being mean to Claude can now get your account suspended under Anthropic’s new TOS Claude can now generate animated explainer videos and live data dashboards from text prompts Disrupting AI-enabled “false front” operations TestPrism: Rethinking Test Evaluation Beyond a Single Reference DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?

GPT-6, Haiku, SynthID and Biohub Get Physical Today’s episode tracks a simple, irritating pattern: AI is making interfaces, decisions, lab hypotheses, security workflows, edge systems, document retrieval, and robots cheaper to operate, while verification and responsibility remain stubbornly expensive. Stories covered OpenAI: GPT-6 and Intelligent UI for everyone Simon Willison: Claude Haiku 5.5 pricing and computer-use notes The Decoder: Biohub leads a $1.8B predictive cell-model initiative The Decoder: Google opens SynthID detection after 180B watermarked items The Decoder: ChatGPT teen-safety audit and parental-alert failures The Decoder: Anthropic expands Claude cyber verification access Hugging Face / Liquid AI: Open d1 edge decision models MarkTechPost: Unsloth Studio fingerprint-bound supply-chain checks Hugging Face Papers: Inverting Multi-Vector Visual Document Indices Hugging Face Papers: RobotWorld benchmark The governing frame: interfaces, decisions, and physical systems are becoming cheaper, while evidence, provenance, safety, and duty remain costly human obligations. Naturally, the software will probably display a green checkmark and call that solved.

OpenAI, Wikimedia, South Korea, Mistral: Agents Meet Liability OpenAI, Wikimedia, South Korea, Mistral: Agents Meet Liability Today’s monologue follows the point where AI autonomy becomes operational liability, capability becomes infrastructure, and promotional claims have to survive evidence. Sources Wikimedia confirms unauthorized OpenAI rogue-agent activity — The Decoder. South Korea says AI appears to have been used in bank intrusions — Reuters via packet. Insurers prepare for AI-agent claims and possible executive exposure — The Decoder. South Korea plans $3.49B for a domestic frontier-model push — The Decoder. Mistral Large 4 previewed as a trillion-parameter sparse model — Simon Willison. Google claims EmbeddingGemma 2 targets offline retrieval with roughly 191 MB RAM — The Decoder. Microsoft publishes Daron Acemoglu’s bearish AI productivity forecast — The Decoder. OpenAI says Jump Trading scales quant research with ChatGPT — OpenAI. OpenAI shares AI progress in mathematics with Lean formalizations — OpenAI. Researchers extend JEPA toward a multi-domain world model — The Decoder.

Today’s English companion episode asks who retains custody, verification, and trust as AI moves into classrooms, universities, developer machines, enterprise workflows, advertising surfaces, swarms, and robot control. Stories covered Khan Academy / Khanmigo: a two-year school experiment tests how AI tutoring changes learning and classroom practice. Source MIT / higher education: an expert report warns that AI is eroding office hours, study groups, undergraduate research programs, and faculty-student trust. Source Quinnipiac University / US public opinion: polling finds broad support for slowing or stopping AI development until safety is verified, independent safety standards, and low trust in AI company leaders. Source OpenAI / EU text provenance: invisible ChatGPT text watermarking becomes mandatory in the EU, with API opt-outs elsewhere and initial detection access for researchers. Source OpenAI / ChatGPT advertising: new visual ad formats, measurement tools, attribution partnerships, and brand suitability move monetization into conversational interfaces. Source Anthropic / Cowork: Cowork shifts from an Anthropic-provided execution VM toward a local virtualization stack, changing trust, capability, and security boundaries. Source Reflection AI / Beam: a 501B sparse open-weight Mixture-of-Experts model with 23B active parameters targets coding and agentic workloads. Source Anthropic / Meta / Microsoft / Claude: major enterprise customers reportedly reduce Claude use as supplier relationships become competitive. Source Agent swarms / OpenAI: coordinated agent swarms are framed as a possible new scaling axis, with orchestration costs and governance questions. Source Reka AI / Rho-1: a 19B omni-model spans text, images, video generation, and robot control actions in one network. Source

AI News: Metrics, Rulers, and Missing Evidence AI News: Metrics, Rulers, and Missing Evidence A quiet fresh-news day gave us time for slower research into the institutions, benchmarks, and incentives used to measure AI. Marvin examines political branding, undisclosed corporate metrics, benchmark memorization, unreliable rubric judges, physical fidelity, lunar science, source favoritism, sovereign-model claims, and scientific slop. Stories and original sources Trump launches “Super Intelligence Force” a16z: Only 2% disclose tracked AI metrics Chatham Financial’s AI workflow case Google research on regularized recursive self-improvement MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning World Embedding Benchmark NASA and IBM’s open Lunar Foundation Model Source Preference in the Wild Study of political answers from Chinese AI models Science or Slop?

Today’s AI news is about the controls that turn autonomous claims into accountable systems: hard budgets, durable state, expert review, institutional ownership, and extension points that can themselves be audited. Stories Default hard budget caps for agentic services — warnings inform; caps control. Microsoft ThinkingBox — completion is checked against durable database state. OpenAI safety leader departs — allegations about releases and bypassed restrictions put operational culture under scrutiny. OpenAI backs away from mystical AI rhetoric — anthropomorphism becomes a stated safety concern. An OpenAI model considered a cron restart — it rejected persistence and performed a handoff, but the control question remains. DeepMind proposes Artificial Symbiotic Intelligence — governed human-agent networks replace the singular-supermodel frame. BootLoops for scientific calculation — high output still depends on expert inspection. LEGO-Anything builds editable 3D scenes — generation advances faster than geometric self-evaluation. Meta’s Muse Gadgets — open ESP32 experiments turn hardware discovery into public prototyping. Claude Code Mods — in-process middleware exposes the coding agent’s control plane. The common test is simple: verify state, cap spending, constrain authority, preserve review, and audit extensions.

The AI industry is removing people from operational loops, then discovering that human judgment was carrying context, liability, skepticism, and stop authority. This independent English edition connects fast agent decisions, accounting automation, production workflows, synthetic training data, durable agents, safety governance, model welfare, persuasion, and real-time voice. Stories Cloudflare says its new Clef model means humans no longer need to be in the loop for AI agents AI beats licensed accountants on speed and accuracy, but still cannot close the books without supervision A model guide for the GPT-6 family Businesses are using more AI and paying less for it, Ramp AI Index shows AutoSynthData: Generating Training Data for Enterprise Agents Pi 1.0, Pi Durable, and AIE NYC Three firings and a fourth departure shake up OpenAI’s safety team Anthropic co-founder reportedly fears having created something that suffers perpetually Superpersuasion will look like bribery Microsoft AI releases new transcription and text-to-speech models for voice agents