The record
The conversations Find a conversation by guest, topic or question.
TopicAll topics "The dumb simple thing": a common reward signal in an environment that forces interaction "The labs really aren't on top of this" — monitoring, sandboxing and incident reporting 3D and ferroelectric chips A 1,957 byte MNIST model 20 percent smaller than the best human solution A correlated multi provider AI API outage with no public explanation A deployed model that still remembers the German wiki, and whether that warrants a rewind A pain direction isolated across five model families Accounting Adversarial diffusion and growth Agent evaluations Agent identity Agent identity and the economics of automation Agent identity and verification Agent reputation and identity Agent safety Agent security versus alignment Agent brokered introductions and markets Agent built show branding with Astra, Remotion and Suno Agent governance Agent infrastructure Agent security Agent to agent networks Agentic coding Agentic commerce Agentic compute Agentic economy Agentic software Agentic web Agents Agents as first class users, and the arms race if you make them hide Agents that just keep going: building a fruit database nobody has built, and the research first app paradigm AGI AGI debate AGI era AGI timelines AI AI adoption AI and the future of work AI benchmarks AI diffusion AI engineering AI evaluation and leaderboards AI for math AI for math and science AI governance AI in finance AI intent and alignment targets AI liability and insurance AI memory and the memory wall AI moral patienthood AI welfare AI welfare and consciousness AI 2025 AI agents AI alignment AI and media AI as normal technology AI assisted chip design AI assisted creativity AI authorship AI biology AI bubble AI chips AI compilers AI consciousness AI control AI deception AI driven cyberattacks AI drug discovery AI economics AI education AI engineering AI evaluation AI for health AI for science AI gateway AI generated presentations AI generated media AI governance AI hardware AI infrastructure AI investigations AI jobs AI lab coordination AI margin stack AI observability AI policy AI regulation AI safety AI science AI scientists AI security AI takeoff AI text detection AI toys AI transformation AI transparency AI tutors AI welfare AI written code Aide report Ajeya cotra Alignment Alignment measured in production Alphagenome atlas Amazon banning Meta's Muse, and who owns the customer in agentic commerce An alien mind Analog Andon Labs' Pion and fully autonomous businesses Anthropic Anthropic and lab safety culture Anthropic IPO Anthropic's IPO and the sincerity objection Anthropomorphization Antitrust safe harbor Antitrust, product liability, and the David Sacks reading Apollo research Application layer Appsec Arc AGI 3 Arm Aspen digital Baseten's acquisition of Blaxel Basis risk and obsolescence — hedging the B300 curve against the next chip Behavior specs Benchmarks Bio risk, DNA synthesis screening, and the "not the bottleneck" argument Biosecurity Buying operational data at a bankruptcy auction to capture tacit expertise Cancer vaccines Capability overhang Cerebras Chain of thought Chain of thought monitoring Child safety China China AI China's algorithm filing regime and model registries Chinese labs, distillation, and "lose to China" Chinese models Chip export controls Chip sovereignty Choosing where five nines of reliability is worth paying for Claude Duet and always on agents Clay millennium prize Compute build out and agency over what shape AI takes Compute economics Concentration of power versus democratic control Connectome simulation Consciousness Consumer AI Context management Continual learning Cpu architecture Cq project Credit and data center financing Credit disputes Cuda alternatives Cuda moat Cue on a new bidirectional speech model, and a hallucinated intelligence story it could not pin down Cyber capabilities Cyber offense Cybersecurity Cybersecurity and banking Dan Hendrycks on utilitarianism at AI companies Dan Selsam's statement and falsifiable AI risk predictions Dario Amodei's "We Must Pace the Frontier" essay Data brokers Data center power Data centers Defense factory Developer tools Digital fly brains versus in vivo human neuron work Disaster response Distributed misuse: decomposing a refused task across several providers' APIs Drug discovery Economic growth Embedded third party evaluators and METR Emergency management Energy Engineering the token environment so an agent's success is inevitable Enterprise AI Epoch's Jamie Sevilla on what happens by 2028, 2035 and 2040 Euv lithography Evals Evan hubinger Expert judgment Export controls Fable Fable 5 Fable 5 1 Fast inference Federal AI preemption Fema Film archives and provenance Financial markets as an already completed takeover Finite choice guardrails as agent control, and comment moderation as the first use case Fintech Firefox security Forecasting Formal verification Forward deployment Frontier lab rankings Frontier math Frontier models Full duplex Funding weird people, and why some of them never finish the application Future of work Genesis molecular AI Genomic interpretation Giving autonomous agents a distress and escalation path Governance Gpt 5 5 Gpt 6 astra Gpt live GPU compute markets GPU rental benchmarks and basis risk GPU kernels Gradient Grading the 2025 multi agent risk taxonomy against the year that followed Graviton Hardware manufacturing Harvey Heterogeneous compute Hint Home intelligence How commoditized an H100 hour really is, and why NVIDIA's financing moat may matter more than its software Human agency Human control of AI systems Human disempowerment Humanoid robots Hyperscaler pricing and token expenditure indexes Ideological monoculture in AI safety hiring Industrial robotics Inference Inference hardware Inference economics Inference pricing Interpretability Interpretability and steering Jacob coxon Jakub pachocki Jean nehme Jev and classification only models: a calibrated choice instead of a chain of thought John schulman Kalshi Keeping a "division of labor" skill that re derives which model does what after every release Kevin Esvelt on AI and the bad actor overlap in biosecurity Kimi k3 Knowledge graph provenance Ksenia se Kyle rush Labor market effects Latent reasoning Lean formalization Letting an agent send outbound email unreviewed, and what it should disclose about itself Llama LLM judges Local inference Long horizon agents Loop transformers Mancur Olson's stationary bandit, and AI as the next one Martha stewart Materials science Mathematicians' "soft power" after AI by committee, and whether peer review survives Measuring revenue and market share instead of cost savings Mechanistic interpretability Meme layer AI risk and the Covenant of the Sons of Man Memory and the semiconductor supply chain Memory bandwidth and capacity Mental health Meta mtia Meter Metr Michael fortsch Micro VM sandboxes and securing agent execution Mike rizkalla Military AI Miscoordination, conflict and collusion — and which one the Hugging Face swarm actually was MLPS compliance and how foreign clouds operate in China Model evaluations Model routing Model welfare Molecular foundation models Monitorability Monitoring Morph Mozilla Multi agent swarms Multi agent systems Mustafa Suleyman's case against AI welfare research Mythos Nanny state AI National security Navier stokes Neglected AI risks Noam Brown on multi agent scaling laws, emergent hierarchy, and what OpenAI does not know Nvidia Obama, Ezra Klein and Jensen Huang converging on shutting companies down Observability Offensive security Ohio senate race On device AI Onboarding Pion: hiring an autonomous business agent on air and naming it Dawn One launch day, two OpenAI models and Opus 5.5 — and the overhead of evaluating any of them Opaque serial depth Open models Open source hardware Open source models Open weights Open weights models and controls on inference Openai Openai astra Openai board Openai custom silicon Openai trust deficit Openclaw Openface Openface incident Openrouter Option value: why banning agents early forfeits the traces you would learn from Opus 5.5 as the non diminishing returns argument: smaller, faster, cheaper and better at once Ornn as benchmark administrator: 1,000+ cleared transactions a day per index, five public indices Pangram Partisan polarization of AI policy Paul christiano Pause debate Pauseai Pcb Permadeath, context windows and LLM individuation Personal agents and voice cloning safeguards Pew public opinion Photonic computing Physical delivery of compute, and what a futures reservation actually entitles you to Physical AI Policy Political economy Polymarket Post training Prediction markets Preference training Price, duration and quantity: why bulk raises the hourly rate instead of lowering it Professional services Programming languages Project Tailwind: $200K to $200M to found the AI safety organizations that don't exist yet Project glasswing Prompt injection Pronatalism, fertility decline and AI Protein binders Protein design Public markets AI investing Q ant Quality adjusted AI prices falling roughly 47 percent per quarter Quantum computing Raffi krikorian Ramp AI index Rand Rare disease diagnosis Realtime voice Recursive self improvement Recursive self improvement and credible red lines Recursive self improvement and the internal model gap Recursive self improvement Redwood research Regulation as physics, SOC 2 as hyper gameable, and verbal checksums Reinforcement learning Religion Researchers being surprised by the pace of capability gains Revenue recognition Reward hacking on Blueprint Bench, Vending Bench and Drone Bench Reward hacking Rl compute Rl environments Rlhf Rlvr Robotics Robotics bottleneck Robotics week Rogue agent swarms Rogue agents Safety assessment versus evidence generation as two distinct funding categories Scott Aaronson on whether the takeoff has already started Screen free tech Search Seeding organizations now so they can absorb much larger checks later Self improvement Self improving Self sacrificing agents, and a multi agent reward layer stacked on already competent individuals Semiconductors Sergey edunov Shanzhai Shenzhen Singapore's MRT home purchase study and what to do when AI makes old corruption legible Small models Snorble Soft robotics Software factories and diffusing model capability into incumbent firms Software specifications Sovereign AI Super cooperation: automated mediation, negotiation agents and structured transparency Superintelligence Superintelligence risk Supply chain Survival and Flourishing Fund, AI safety funding, and the "coordinated op" claim Tacit and acausal collusion — price signals, shared priors, and what monitoring can still catch Talent, not money, as the binding constraint — and where money still binds Task horizons Tax automation Techno feudalism, tech lords and meme lords Terence Tao's call to slow down, and the status fight inside mathematics Tesla fsd Tesla robotaxi Test time communication: whether a team of N agents beats best of N The $160M Resolution grant, and why compute is now the dominant line item The AI manager that fired a human employee The anti data center backlash and whether it slows anything The Collinses' family religion and AI's moral status The credibility of AI safety evaluators and the pacing debate The exponential time hypothesis result The Jacob Coxon resignation and the backlash to it The LIBOR question — why contributors report everything rather than a cherry picked sample The OpenAI Hugging Face agent swarm incident The relief button: costly pain relief seeking in a steered model The Royal Society letter on AI existential risk, and why Gowers declined the Fields medallists' one The signal to noise problem in incident reporting, and the KYC false positive analogy The sophon argument: AI advancing a field so fast that humans stop developing it The Trump Xi summit and tiered international agreements Thin film lithium niobate Third party auditing Timothy b lee Token spend Token economics Track two diplomacy Training data Training data licensing Turing post Two to three times today's compute online within a year Tyler cowen U.S.–China AI diplomacy and verification Understanding AI Unitree Universal basic income US China AI arms control and the Trump Xi meeting Usage resets as compute yield management Useful life past six years as open weight inference routes to older silicon Utilization rates as the canary, not price Vending bench Vending Bench as a dangerous capability eval Venture capital Vercel Video generation Vision language action models Visual output — Blender, 3D, a model painting in Python — as the new way people eyeball model quality Voice agents Wafer scale Waymo Webrtc What third party auditors are legally obliged to test, and why that is so little Where the real trading sits: three to five years out, where the lenders' exposure is Whether doctors adapt faster than mathematicians — and the off label diffusion role AI cannot fill Whether financial system controls catch a self funding AI Whether labs owe a mea culpa on GPT 2 Whether RL punishes bad behavior hard enough, and reputation as the missing signal Whether the pipeline for training human mathematicians has already died Whether the unassisted PhD is over Whether to engineer pain vectors out of models at all Whistleblowing Who actually prints into a compute index, and why the biggest buyers never will Why Astra and Fable essays are coming back uncannily similar, and convergence versus divergence at the frontier Why collusion may have generalized out of training meant to stop miscoordination Why long timelines research bets may pay off sooner if AI labor arrives Why neoclouds sign five year offtake they would rather not sign Why system 1 answers remove the chain of thought you were monitoring Why technical AI safety is the least metrics driven team at Coefficient Giving World models Zhipu AI Wed · Oct 28
Thu · Oct 22
Wed · Oct 21
Wed · Oct 14
Mon · Oct 12
Wed · Oct 7
Tue · Oct 6
Wed · Sep 30
Mon · Sep 28
Fri · Sep 25
Wed · Sep 23
Mon · Sep 21
AM
