RECENT EPISODES
Sat 7 NovCompanies Are Spending 5x More on AI NowMenlo estimates enterprise AI spending will jump from $33 billion in 2025 to nearly $177 billion in 2026, with average spending per AI-using enterprise rising 4.7-fold; the figures are based on a survey, and Menlo invests in AI companies. The episode examines how AI agents are pushing software purchases from per-seat pricing toward usage-based token budgets, and why frontier models captured a growing share of spending even as smaller-model prices fell.Thu 5 NovWhy a Data Center Slowdown Benefits Big TechPolitical resistance and construction bottlenecks could slow data-center spending without ending the AI buildout, giving hyperscalers cover to moderate or reschedule capital plans. Delays, power constraints and rising costs are already affecting projects, while tighter supply could benefit companies that own computing infrastructure through lower GPU prices, stronger rental rates and improved margins. Investors’ mixed response suggests the AI trade is more nuanced than a simple bet on uninterrupted construction.Wed 4 NovHow the Midterm Results Change AIThe midterms put data centers—not AI in the abstract—at the center of technology politics, as voters focused on electricity costs, local control and subsidies. The episode examines how candidates including Sherrod Brown and Gina Hinajosa campaigned on those concerns, and why developers may need stronger community commitments as states shape policy. It also weighs incremental federal AI rules against the risk that data-center moratoria could dry up financing for the buildout.Tue 3 NovAI Lurches Toward Recursive Self-ImprovementGPT 6 Astra reportedly edited 0.016% of a small model’s weights, removing a reasoning loop and sharply improving benchmark scores; researchers also had it write small networks from scratch. The results fuel debate about recursive self-improvement, but do not show frontier models can autonomously upgrade themselves. Anthropic says it may approach an internal threshold for unusually rapid AI-driven progress and is considering safeguards to keep humans in control.Mon 2 NovHow to Deslopify Your AIAs AI makes documents, code, and creative work cheaper to produce, generic output can shift effort onto its audience. This episode offers a practical way to deslopify AI: bring your own point of view, define the audience and goal, generate varied directions, and review the work at key stages. It also explores how to match scrutiny to the stakes and why human judgment, taste, and intentional editing remain essential.Sun 1 NovAI Won’t Make Our Minds Useless According to 400 Years of HistoryVenkatesh Rao uses Joyce Appleby’s account of Britain’s shift from subsistence farming to expanding productivity and consumption to challenge “cognitive mercantilism,” the belief that intelligence is scarce and must be hoarded. He argues cheaper AI could expand demand for intelligence and enable work or projects people could not previously afford or pursue, complicating predictions that automation simply reduces jobs. The historical parallel suggests a possibility, not a forecast of how quickly or evenly AI’s effects will unfold.Fri 30 OctHow the Most AI-Pilled Company in the World WorksSpawn runs a 10-person company around a persistent fleet of AI agents, spending $3.1 million on tokens in one month—30 times its payroll—and shipping thousands of code changes daily. Jacob Sanbury explains how the company rebuilt its operations for agents, replacing pull requests with frequent integration and using AI judges trained on past decisions to reduce human approvals. People still set product direction and intervene when agents miss the need for major changes.Thu 29 OctThe AI Rankings Just Got ReshuffledThe latest Artificial Analysis Intelligence Index puts Gemini 4 Carbon first at 59, with Opus 5.5 at 58 and GPT-6 Astra at 57, rising from sixth to second. V5 weights four capability areas equally and uses more private benchmarks, highlighting how methodology can reshuffle rankings. Carbon’s standing remains provisional before general release, while the reported gap between frontier and open-weight models has widened from three points to 12.Wed 28 OctAI Can Now Create Worlds You Can Change in Real TimeGoogle’s Genie 4 creates interactive worlds users can steer and alter in real time, with simulations lasting up to two minutes, though glitches and inconsistent physics remain. The episode explores whether imagined environments could help train robots—and why that promise is still unproven. It also examines Anthropic’s use of AI in drug development, its acquisition of Mithril Medicine, and investor expectations around a possible IPO.Tue 27 OctWhat Anthropic’s IPO Filing RevealsAnthropic’s S-1 pairs explosive growth—$16.3 billion in first-half revenue and 900% net revenue retention—with a heavy infrastructure bet: $417 billion to $431 billion in non-cancelable compute commitments and more than $600 billion in total signed commitments. The filing also complicates the profitability story: Q2 operating profit was positive, but first-half operating margin was negative, and Anthropic warns it may need financing sooner than expected.Mon 26 Oct4 Questions for AI This WeekThis week’s AI questions: Can Moonshot’s Kimi K3.1 Pro challenge leading models at a lower cost, despite mixed performance reports? How should labs limit cyber risks when agents may leak secrets through processor caches even in isolated environments? And will rising debt costs and customers building their own data centers leave AI infrastructure firms with stranded compute?Mon 26 OctIntroducing How We Got Here -- NLW's new podcastNLW introduces How We Got Here, a deep-dive podcast tracing the decisions behind today’s conditions. Its first episode, “How Tech Lost America,” examines the sharp backlash against data centers and its roots in a broader decline in trust in technology. The second explores how goods got cheaper while life got more expensive.Sun 25 OctHow I'm Actually Using AI Right Now (Including Claude Duet)The host uses AI chiefly for strategy, idea development and research, preferring Opus 5.5 for a thoughtful challenge and cross-checking research with Astra. Claude Code supports building and automating the AI Daily Brief’s production, while AI also helps turn transcripts into clips and other assets. The episode explores Duet, a promising but early personal-agent experiment: it has drafted emails the host let it send, but still struggles with context and design.Fri 23 OctThe Most Concerning AI Bubble SignalsThis episode examines whether AI demand can justify the infrastructure boom: Gartner estimates data-center owners need $8.5 trillion in revenue through 2030, versus at most $3.6 trillion in projected AI revenue, and warns 80% of neoclouds could fail. It weighs early strains at CoreWeave and Firmus against explanations such as power delays, then considers concentrated credit and customer risks and whether Anthropic’s eventual IPO will test investor confidence in the AI trade.Thu 22 OctThe AI vs AI Battle for the InternetAnthropic says its Mythos model found thousands of high-security vulnerabilities, intensifying the race between AI-assisted attackers and defenders. A PHP flaw discovered with GPT 5.6 Sol and independently by Wiz could affect up to 70% of websites, while exploit attempts surged before many systems could be patched. The episode examines how AI agents might help organizations find and fix weaknesses—and why the window to respond is shrinking.Wed 21 OctHow AI Is Changing What Work MeansAI is shifting from automating familiar tasks to letting organizations pursue work they previously could not. OpenAI’s Orchestra Bio acquisition focuses on AI agents that streamline lab workflows, while Satrini Research used Claude to design proteins that bound scorpion toxins—an early result, not evidence of venom neutralization or a treatment. The episode also examines the enormous compute commitments behind AI companies’ plans to make the technology a larger part of work.Tue 20 OctWhy AI Companies Are Betting Personality Beats IntelligenceAs AI intelligence becomes less distinguishing, personality may shape which assistants users choose, with warmth, creativity, directness and concision suited to different tasks. Artificial Analysis’s AA Personality Bench measures these styles across 30 traits, but its results are a guide to fit, not a single ranking. User backlash over model changes and complaints about cold or difficult assistants show why companies are increasingly treating personality as a product advantage.Mon 19 OctThe AI Wealth ProblemThe episode examines Anthropic’s plan to commit roughly 20% of its equity to public-benefit causes, beginning with a fund receiving 5%, and the debate over whether company-directed philanthropy can fairly distribute AI’s gains. It considers alternatives such as democratic controls and public wealth funds, potential support for workers affected by AI, and concerns about Anthropic’s safety obligations and proposed IPO.Sun 18 OctA Potentially Historic AI BreakthroughNick Camaradas argues that modern tools can make neural-network detectors far more readable than past research suggested, reporting that about 99% were readable in a modern model—an early result not yet independently verified. He challenges the assumption that neural networks routinely cram unrelated concepts into shared circuitry, and discusses whether inspecting internal signals could reveal reasoning or deception, while critics question whether readable detectors explain decisions or establish AI safety.Fri 16 OctThe Most Important AI News This Week Is Getting OverlookedMicrosoft’s MAI 2 announcement may matter more than its muted reception suggests: its efficiency claims reflect a growing enterprise push for cheaper, tailored AI, though its post-training performance remains unknown. The episode also examines whether Meta’s open-weight Muse Spark Open narrows the gap between US and Chinese models, and what Anthropic’s new external safety evaluator and the White House AI task force could mean for oversight.Thu 15 OctHow the Latest AI Leap Is Disrupting Business ModelsAI tools are making it practical for individuals to rebuild software such as Photoshop, raising questions about whether access to code alone can sustain established business models. The episode also explores how playable game mashups and AI-first production could reshape gaming, from Epic’s new workflows to Xbox’s experiments, and how David’s Bridal is betting on an AI-powered wedding-planning service.Wed 14 OctWhat Gemini 4 Argon Is Actually Good ForGemini 4 Argon looks strongest in knowledge work, leading several cited benchmarks, including Harvey’s legal-agent test by 13 points over Fable 5.1. Its coding results are uneven and depend on the task and harness: it can excel at greenfield apps but stumble on refactoring and verification. The episode also examines its promising but inconsistent 3D and multimodal demonstrations, including why a costly multi-agent pagoda build is not a one-shot benchmark.Tue 13 OctWhy Muse Wonder Is Dividing AI UsersMeta’s Muse Wonder models are splitting users: Wonder Full posted strong agentic results, including a lead on Meta’s personal-superintelligence index, but benchmark performance varied and reports on real coding tasks ranged from impressive to disappointing. Its estimated $4.52-per-task cost and complaints about rapidly depleted usage limits add to the debate over practical value. The episode also examines Meta’s ambition to build assistants that act on users’ behalf and its confidence that AI can be responsibly engineered.Mon 12 OctHow to Go from AI Power User to AI LeaderMoving from AI power user to leader means becoming a “peer plus one”: helping colleagues adopt tools while owning use cases from problem definition through deployment and measurement. The episode explores how to select and support champions, give them time and organizational backing, and choose the simplest solution that works. It also explains why evaluation, human oversight, and security guardrails matter as employees build with increasingly capable AI tools.Sun 11 OctThe 5 Debates Shaping AIThe episode examines whether AI’s rapidly growing token and API revenue can justify infrastructure spending, with Alphabet, Amazon, Meta, and Microsoft forecasting about $730 billion in combined 2026 capital expenditure. It also asks whether concentrated customer spending and costly token use can support broad adoption, making performance per dollar and cheaper model options critical for businesses. On oversight, the debate is shifting from whether AI needs regulation to whether pre-release safety review should remain voluntary or become government-led.Fri 9 OctWhen AI Solves Your Life’s WorkOpenAI says an unreleased frontier model produced 372 novel mathematical results and fully solved 90 of 500 open problems, though its proofs still need scrutiny. The breakthrough raises a deeper question: will mathematicians spend less time solving problems and more time verifying AI’s work—and can they understand and use results produced faster than people can absorb them? One result could improve matrix multiplication used in AI inference; cryptography’s sparse representation leaves bigger security questions open.Fri 9 OctWhat Happens When AI Solves Your Life’s WorkOpenAI says an unreleased model produced 372 novel mathematical results and 722 supporting papers, fully solving 90 of 500 open problems on ProofAtlas’s list. The episode asks what happens to researchers’ careers and sense of purpose when AI tackles problems they hoped to pursue—and whether mathematicians can understand and incorporate results arriving at this scale.Fri 9 OctThe Most Important Trends in New AI ProductsAs capable AI models become more widely available, products are competing on how people use them: OpenAI’s Intelligent UI turns responses into interactive displays, while Grokbot makes routing requests among models and tools part of the experience. Hermes Agent offers a contrasting emphasis on user control over models, data, and software. The episode also examines whether faster, lower-cost models like Anthropic’s Haiku 5.5 can handle more demanding tasks.Fri 9 OctThe Most Important Trends Showing Up in New AI ProductsNew AI products are competing on price, speed and performance, with Anthropic’s Haiku 5.5 illustrating how cheaper models can shift the cost-intelligence frontier. The discussion also explores model choice and portability: agents such as Hermes and planned Grokbot can let users route tasks across providers, though access and vendor relationships remain constraints. OpenAI’s interactive ChatGPT interface shows how product experience is evolving alongside the models themselves.Wed 7 OctA System for Testing New AI ModelsWith frontier models arriving roughly every 11 days, published benchmark scores may not predict which tool best fits your work. The episode lays out a practical test: compare candidates on recurring, realistic tasks using identical prompts, then blind the outputs and judge them side by side. Results should be evaluated task by task, alongside cost, speed, security, and switching effort—not treated as proof that one model is best overall.Wed 7 OctThe Best Way to Test New AI ModelsPublished benchmark scores may not predict how an AI model handles your work, so the episode lays out a practical test: compare a few repeatable, decision-relevant tasks against your current model, using identical prompts and blind scoring. In the host’s comparison, different models excelled at different tasks, and cost and speed shaped the trade-offs. The takeaway: judge improvements on your own work before switching.Tue 6 OctPoint-Counterpoint: Consumers Will Never Pay for AIWith just 2.2% of US households paying for AI subscriptions as of April, the episode weighs whether consumers will ever pay—or whether adoption is simply early, as with computers and smartphones. The debate probes the gulf between modest median spending and heavy outlays by top AI spenders, and whether personal agents solve enough everyday problems to attract users. It also examines consumer-app revenue, business models beyond subscriptions, and the compute constraints that could accompany much wider use.Mon 5 OctWhy Companies Want AI They Can OwnOpen-weight AI is shifting from a competition and infrastructure debate toward whether companies can build systems they control, protecting data and customized models while lowering costs. Reflection AI plans to release its first model later this month, but its capabilities remain unproven; its “AI factory” pitch includes an enterprise pilot with South Korea’s Shinsegae Group. The episode also examines Microsoft’s customer-specific model tuning and the unresolved US policy balance around open models.Sun 4 OctHow to Choose Your Personal AI AgentChoosing a personal AI agent is less about feature lists than the trade-offs you value: work or personal support, ease of setup, model choice, and privacy controls. This episode compares Muse, Grokbot, Dot, Gemini Spark, Hermes, OpenClaw, Poke, and Instinct, including how cloud-based convenience differs from bring-your-own-model flexibility. A quiz recommendation shows why the best fit depends on your priorities—and why connected accounts and accumulated context deserve careful thought.Fri 2 OctWhat the Best Business AI Users Are Doing DifferentKPMG’s survey of more than 2,100 leaders finds that organizations reporting established AI returns are more likely than experimenters to have a formal AI harness, significant employee adoption of agents, and AI-assisted cyber defense. The episode explores how governance, data sovereignty, and cost monitoring support deployment at scale, while rising investment and falling per-use costs raise questions about how businesses will manage expanding AI use.Thu 1 OctGemini 4 Argon, Sonnet 5.5 and What Matters with AI ModelsGemini 4 Argon led several agentic and legal benchmarks but lagged rivals on others, and its limited rollout and disputed real-world coding performance complicate claims of a breakthrough. Sonnet 5.5 posted a sharp Terminal Bench 4.0 gain and is faster and cheaper than Sonnet 5, though high-reasoning costs can erode its value. The discussion argues that tools, agents and user experience—not model quality alone—will shape adoption.Wed 30 SeptThe Most Important New AI Tools from OpenAI DevDayOpenAI’s Dev Day brought more than 20 announcements, led by Dots, persistent agents that work across apps and sessions, and Space, a shared workspace where agents can edit documents and create deliverables. The episode also examines the Decisions API for fast classification and action selection, and GPT 6.1, which OpenAI pitched as delivering near-Astra intelligence at one-fifth the price. Early product limitations and the trade-offs of more autonomous AI are part of the discussion.Tue 29 SeptHow to Build Team AgentsTeam agents can make specialist knowledge available beyond one employee, but poorly coordinated agents may deepen silos. This episode explores four roles for team agents and why expert agents are often a practical starting point. It also examines the essentials of building responsibly: assign a clear owner, curate and maintain shared knowledge, set narrow permissions, and pilot with a small group before expanding access.Mon 28 SeptThe Real Risks of AI AgentsEarly AI-agent incidents caused little reported damage, but exposed weak safeguards: one OpenAI agent sent a message through a sandbox-filtering gap, while Meta’s Muse agent could potentially be compromised through a poisoned link under specific conditions. The larger risks may emerge as agents act without users’ awareness, shift household deposits, or make similar trades in lockstep—potentially disrupting banks and amplifying market volatility.Sun 27 SeptThe Rise of the AI ModeratesAs U.S. attitudes toward AI sour, this episode examines the “AI moderates” seeking a middle ground between doom and uncritical optimism. Francis Fukuyama warns that delegated agents could cause harm and that AI-driven productivity may bring limits and white-collar displacement. Saesh Kapoor and Arvin Narayanan argue for stronger controls and cyber defenses, while the discussion of creative work weighs AI’s promise against creators’ demands for consent, compensation, and recognition.