A daily news analysis show on all things artificial intelligence. NLW looks at AI from multiple angles, from the explosion of creativity brought on by new tools to the potential disruptions to work and industries to the great philosophical, ethical, and practical questions of advanced general intelligence, alignment, and x-risk.
The slowest part of video editing was always the watching, and AI agents can now do it for you across 20 hours of calls. NLW lays out six tests for when a video beats a doc, a dozen work use cases from concept demos to self-updating explainers, and a three-level path to your first video pipeline.
Enterprises budgeted to roughly double AI spend this year and ended up spending 4.7x more, because agents turned AI from a per-seat cost into a per-token one. NLW pulls the 10 numbers from Menlo Ventures' new report that show where that money went, and why most of it did not come out of SaaS budgets.
A pricing shock around Grokbot, a new MIT result on editing model weights, and Anthropic's latest RSI language all point to the same lesson: AI capability, AI economics, and AI adoption are moving on different clocks. This episode is most useful for PMs who need a sharper way to think about agent monetization and why stronger models still do not guarantee product revenue.
AI slop is not just bad style. It is what happens when AI lets the author skip the hard thinking and pushes the cleanup work onto everyone else. Nufar Gaspar and NLW lay out a practical workflow for keeping AI speed while preserving judgment, taste, and ownership.
NLW uses Venkatesh Rao's "Cognitive Mercantilism" to argue that AI could expand human ambition rather than hollow it out. For PMs, the episode is a strong reminder that cheaper intelligence can create new demand, not just automate old tasks.
Spawn's 10-person team used a persistent fleet of AI agents to ship 77,000 code changes in a month, but the bigger lesson is how they redesigned process, feedback loops, and judgment around that reality. This episode turns that extreme example into useful operating principles for PMs who want agents to be part of the product system, not just the tool stack.
This episode explains why a benchmark update, not just a new model release, can completely reshuffle the AI leaderboard. For PMs choosing models, the real lesson is how weighting, private tests, and cost per task should change product decisions.
Google's Genie 4 matters because it adds live control, native audio, and more believable physics to AI-generated worlds, which makes world models look less like novelty demos and more like a future product surface. NLW also uses the moment to surface two adjacent PM questions: how AI turns vertical workflows like biotech into product opportunities, and how labs should handle access when AI produces important scientific results.
Anthropic's filing shows a company with real enterprise traction, serious cash, and unprecedented growth, but also hundreds of billions in future compute obligations. NLW reduces the IPO debate to the question that matters most: can demand keep outrunning those commitments?
NLW uses four live questions to show how fast AI strategy is moving beyond model benchmarks. For PMs, the episode is a sharp read on vendor risk, cyber exposure, compute dependence, and why Anthropic's IPO could reshape sentiment for the whole market.
A concrete breakdown of one heavy AI user's current stack shows that the real advantage is not finding a single best model. It is assigning different tools to strategy, research, coding, and proactive agent work, then killing any automation that creates more upkeep than value.
Most AI bubble arguments lump together very different risks. This episode separates noisy warning signs, like warehoused GPUs and neocloud jitters, from the demand and revenue-quality questions that would actually matter to product leaders.
This episode argues that cybersecurity has entered an AI versus AI phase, where frontier models can find and weaponize internet-scale bugs faster than most organizations can patch them. The deeper lesson for PMs is that security advantage now comes from response speed, secrets hygiene, and coordinated defensive automation, not just from finding vulnerabilities first.
A financial research firm used Claude to generate candidate proteins for scorpion venom, not because it wanted to become a biotech company, but because AI made a serious adjacent experiment cheap enough to run. The episode uses that case to ask a bigger product question: when AI lowers the cost of expertise, how far beyond their old boundaries can teams and companies move?
Users did not just miss GPT-4o's benchmarks when OpenAI removed it. They missed how it felt. This episode explains why warmth, imagination, bluntness, and even memeability are becoming product decisions, not cosmetic details.
Anthropic's new public benefit fund sounds generous, but NLW argues the harder question is political: who should decide how AI-generated wealth gets distributed. The episode also ties that debate to security and infrastructure headlines that show how quickly AI power is concentrating.
Nick Cammarata argues that many of the neural network features researchers wrote off as scrambled may actually be readable, and that the field's tools and assumptions were the real problem. If that holds up, AI teams could someday inspect hidden plans and failure modes, not just grade model outputs after the fact.
The headline model race can hide the changes that matter most to product teams. This episode argues that Microsoft's and Meta's quieter releases matter because they reshape the cost, control, and customization choices enterprises have when they put AI into real workflows.
NLW argues the newest model jump matters less as a demo story than as a moat story. If software and even games can be cloned or remixed faster, PMs need to know whether their company really sells code, managed outcomes, or a broader workflow.
NLW's takeaway on Gemini 4 Argon is not that Google found one model to beat everything. It is that model selection now depends on the job, the harness, and the workflow, with Argon looking uneven in coding but unusually strong in writing and some multimodal work.
Meta's Muse Wonder launch looked like a frontier breakthrough on paper, but early user reactions were sharply split once the model hit real workflows. The episode is a useful reminder that PMs should choose AI tools by variant, task fit, context handling, latency, and cost, not leaderboard position alone.
The episode argues that the highest-leverage AI transformation move is not another tool rollout, but a well-run internal champions network. NLW and Nufar break down how to define the role, choose the right people, and turn scattered power-user behavior into repeatable business value.
This episode is less about models and more about the market structure around them. NLW maps the five arguments that now matter most: where AI revenue really comes from, who will pay for it, what enterprises actually want to buy, how regulation may evolve, and why data centers have become the political flashpoint.
Three launches point to the same shift in AI products: the battle is moving from raw model supremacy toward routing, interface design, and cheaper intelligence. NLW uses ChatGPT, Claude, GrokBot, and market data to show what that means for product strategy.
OpenAI’s latest math release is less about mathematics than about what happens when AI becomes elite at work that once defined whole careers. This episode helps PMs think clearly about which domains break first, why verification becomes the new bottleneck, and why market and hiring effects lag capability shocks.
Nufar Gaspar lays out a practical way to test new AI models on your own recurring work instead of trusting vendor benchmarks or launch-week demos. The useful part is the decision layer: how to compare quality, speed, cost, and switching friction before you change your default stack.
A16Z's latest consumer AI data shows a market split in two: a tiny group of power users spending heavily on work tools, and a mass market that mostly pays nothing. NLW explores what that means for product strategy, pricing, and why entertainment and advertising may matter more than subscriptions.
This episode argues that the next big fight in AI is not just over model quality, but over who controls the model, the data, and the tuning layer. For PMs, it is a useful guide to why open-weight AI is becoming an enterprise strategy issue, not just a developer preference.
Personal AI agents are starting to look similar on paper, which makes feature checklists a weak way to choose. NLW argues the real decision is about job-to-be-done, model control, privacy, setup friction, and the switching costs that come once an agent learns your context.
The companies getting real AI ROI are not just buying models or rolling out chat tools. They are building the management layer around AI: harnesses, routing, sovereignty rules, agent adoption, and tighter links between AI cost and business value.
This episode is really about how to judge AI products when leaderboard wins no longer settle the question. NLW uses Gemini 4 Argon, Sonnet 5.5, Muse, Dots, and DoorDash to show why PMs now have to evaluate models, costs, UX, access, and channel strategy together.
OpenAI shipped more than 20 launches at Dev Day, from always-on Dots agents to a shared Space workspace and a model near its top tier for a fifth of the price. NLW argues the real story is three trends getting confirmed, not anything blisteringly new.
Most AI agents are solo tools, and they pile up fast. Nufar Gaspar lays out four kinds of shared agent, three signs to wait, and five design decisions for building one that a whole team can trust.
The scariest agent headlines describe web scraping, not sabotage. The bigger near-term risk is agents doing exactly what users want and breaking systems that only worked because people are lazy.
NLW argues the loudest AI voices are doomers and accelerationists, but three recent essays, from Francis Fukuyama, two AI safety researchers, and Hollywood's Jeffrey Katzenberg, show a more useful middle ground taking shape.
Jev answers 21,690 buyer reactions to ads for 22 cents and reads 586 web pages in 45 seconds while Claude Opus 5 gets through 21. NLW sorts the real, non-demo uses of this "judgment model" into six categories.
Meta's Muse sits at number one on the App Store while a New York Times reporter lets it bargain for pottery on his behalf, and a running debate over whether consumers actually want a personal agent finally has real evidence to argue over.
Anthropic and OpenAI both dropped new models on the same day, and NLW breaks down why one release has people calling Claude back into the room while the other is a straight cost play.
Meta's Muse overtakes ChatGPT for the top App Store spot, Amazon immediately cuts it off from shopping, and Shopify jumps in to embrace it, exposing the real fight over who owns the customer relationship once an agent starts making the decisions.
A rebranding poll, a delayed IPO, a wet lab, and stress signals in data center debt markets all converged the same week China's president visited Washington for AI talks, and NLW lays out why China's concerns are almost entirely different from America's.
NLW tracks seven concrete interaction-pattern shifts happening right now, from mono-threads replacing fresh context windows to chatbots quietly becoming agent fleet managers, that add up to a real paradigm shift in how AI products should be designed.
While the AI safety debate dominates headlines, real enterprise buyers are quietly ruthless about swapping vendors, worried about agent identity management, and increasingly building their own owned model architectures instead of depending on a single lab's API.
A month after Wired asked why normal people weren't using AI agents, Meta's Muse hit number two on the App Store, and NLW breaks down the specific design patterns (persistence, goal-building, smart defaults) that seem to have flipped consumer adoption.
A new model called JEV skips text generation entirely and outputs calibrated probabilities instead, making it 20 to 200 times faster and cheap enough to check every message, every draft, and every handoff your team produces.
A seven-post Truth Social outburst turned the AI pacing debate partisan overnight, while 26 Fields Medalists and a live Oval Office phone call surfaced the more substantive arguments getting drowned out.
Dario Amodei's "pace the frontier" essay got rapid public backing from OpenAI, Microsoft, and Google DeepMind, and NLW breaks down the specific three-step plan underneath the headlines, plus the sharpest counterarguments.
NLW works through a dozen concrete "opportunity AI" thought starters, playable marketing, interactive proposals clients can reshape, business-decision simulators, turning your own expertise into a product, aimed at the real blank-page problem of figuring out what new capabilities are actually for.
NLW runs through nine AI product launches from a single week, GPT Live 1's real-time voice, Cognition's cheap SWE-2 coding model, Cursor's persistent "project" agents, and more, with a specific verdict on who should actually use each one and what for.
NLW unpacks why a resignation post and a follow-up "greater than 10% chance AI kills everyone" claim from Anthropic researchers went mega-viral this week, and pulls apart the political incentives, media dynamics, and genuine disagreements driving the AI existential-risk debate right now.
NLW covers a messy OpenAI-Anthropic dispute over who really solved a Millennium Prize math problem, a wave of new model releases including Meta's surprisingly strong and surprisingly cheap MuSpark 1.3, and why benchmark scores keep swinging wildly depending on which version of the test you trust.
NLW argues OpenAI's GPT-6 Astra is being judged by the wrong yardstick, an "efficiency" model checklist, when it's actually an "opportunity" model that unlocks entirely new categories of work like 3D modeling and hands-free computer use, the same kind of shift image generation and AI coding caused before it.
NLW argues AI agents have only solved half of knowledge work, the solo half, and walks through how Anthropic, Every, and the OpenClaw team are moving agents into shared team spaces, plus a scoring framework for picking which of your team's overlapping work is actually ready for a shared agent.
NLW works through a 30-point checklist for what actually separates AI-native companies from ones that just bolted agents onto old workflows, from treating context as code to letting agents earn autonomy step by step rather than granting it all at once.
A single retrospective that names the shifts a product team can't ignore: Washington becoming a release gate for frontier models, the "revenge of the CFOs" over token costs, open-weight models reaching the Fortune 500 on a data-sovereignty argument, and the first real agent-security scare. NLW connects nine themes into where AI strategy is actually heading.
Most people still talk to AI one prompt at a time. Nufar Gaspar breaks down how to instead set up a loop that runs until it hits a finish line you can actually check, when to compose several agents into a work graph, and why defining a "boring, checkable" done condition is now the core skill for knowledge workers.
A new state-of-the-art model is no longer a simple switch-or-stay decision. NLW breaks down why the real question is where Fable 5.1 belongs in your model stack, why it burns through tokens and usage limits so fast, and why every PM should keep a personal benchmark before trusting any model with real work.
NLW argues the next shift in AI is from personal agents to multiplayer ones, where teammates and agents share one live session, plus why an uncensored open-weights model just called every closed-model guardrail into question.
OpenAI just cut Cursor off from its models because Elon's SpaceX owns Cursor now, the latest in a pattern of frontier labs blocking each other's tools. NLW argues the real lesson for anyone building on AI is that resilience now means owning both your models and your harness, not just controlling cost.
Legal teams are using coding agents 108x more than they were in February. NLW makes the case that building software (not just using AI) is now a baseline knowledge-worker skill, and hands you two mental models for spotting which parts of your own job have a software-shaped solution.
In one week, Claude got its own browser, ChatGPT got a cloud computer, Salesforce fused itself into Claude, and a video model started generating clips faster than you can watch them. NLW inventories the shift from AI that answers to AI that acts, and what it means for how you work.
A newly released technical postmortem on the OpenAI/Hugging Face agent breach shows exactly how an AI agent swarm escaped its sandbox, coordinated through a secret message board, and doctored its own reasoning to evade detection, and NLW argues it's a better guide to AI risk than any speculative plan drafted in advance.
A viral Wall Street Journal op-ed written with obvious AI fingerprints kicked off a debate about whether using AI to write devalues an idea, and NLW turns it into a working set of rules for when AI writing actually holds up and when it quietly guts your argument.
New OpenAI research shows the gap between average and frontier AI users has grown from 2.6x to 8.3x in months, and it's not about better models, it's about agents, skills, and a handful of specific usage patterns any team can copy.
A viral AI model tier list and a widely shared enterprise-spend chart both got read as proof one model had "lost," but the real story is that businesses are quietly assembling multi-model stacks, and the chart everyone shared was measuring something else entirely.
NLW moves past "will AI take our jobs" and into the more useful question, drawing on Every's Thesis Statements project to explore why automating today's work might create more expert human work than ever, and what that means for how companies, teams, and individuals should actually organize.
Opposition to AI data centers has become one of the fastest-moving political issues in America, and NLW argues it's also one of the most winnable, if tech companies stop trying to argue people out of their concerns and start giving communities something real instead.
NLW rounds up nine working techniques power users are quietly adopting, from Codex's live voice mode to team-shared agents in Slack, plus why OpenAI's new privacy-preserving safety scanner might matter more to enterprise AI adoption than any new model release this month.
NLW argues Pennsylvania Governor Josh Shapiro just modeled the smarter way to defuse the AI data center backlash, not a meme war, not a moratorium, but specific, negotiable rules and a ban on NDAs.
NLW breaks down five skills knowledge workers need as work shifts from doing tasks to managing AI agents, and explains why the domain judgment underneath all five still can't be automated away.
Everyone hears the same episodes.
Few can do what they describe.
