McKinsey & Company
06:11 PT
What can leaders learn from Asia’s founder-CEOs?
So what
Nothing actionable; it's a book trailer. File Shapers and Founders if you want Asia founder-operator case studies for board conversations.
A book trailer dressed as research: McKinsey is promoting Shapers and Founders, its new study of Asian owner-CEOs, with a video of the authors walking through the traits they found.
- Authors Gautam Kumra, Cheryl Lim and Joydeep Sengupta draw on interviews with Asia's founder-CEOs
- The claimed traits: mission and resilience, long-term vision held alongside present-tense focus, discipline across both transformative and incremental innovation
- Audience pitched as CEOs, aspiring entrepreneurs, investors and board directors in any market
- No data, no cases, no numbers in the email itself — the content is behind a video link
Lenny's Newsletter
05:31 PT
Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)
So what
Stop briefing your team that models are next-token predictors; Anthropic's own researchers say planning and concept representation are the better working model.
Emmanuel Ameisen, an Anthropic interpretability researcher, argues the next-token-predictor framing is actively misleading about what models do inside, and that anthropomorphizing them is closer to correct than the mechanical description.
- Models form general representations of concepts and plan many tokens ahead without stating those plans
- Ameisen describes something structurally similar to the conscious/unconscious split in human processing; the "J space" work probes model self-awareness
- Models are grown rather than built, which is why apparent capability jumps sit on a smooth scaling curve
- Prior career: applied ML at Stripe and Insight; author of Building Machine Learning Powered Applications
- References Anthropic's own public artifacts — Tracing the thoughts of a large language model, Claude's Constitution, Golden Gate Claude
Every
04:58 PT
Codex’s Comeback Up Close
So what
Codex and ChatGPT for Work at 40 million weekly users kills your assumption that Claude Code owns agentic coding by default.
Dan Shipper's reported account of how Codex went from a flopped Cloud launch to the app that ate ChatGPT from inside, plus a Gemini 4 Argon review that puts Google back in the frontier pack.
- Codex and ChatGPT for Work now total 40 million weekly users, built by a small team in OpenAI's "backwaters" after Codex Cloud flopped and the Windsurf deal collapsed
- Shipper interviewed Greg Brockman and Thibault Sottiaux over months; paid subscribers get full transcripts
- Gemini 4 Argon is strongest as a collaborator — Shipper uses it to structure essays — but overbuilds unsupervised, and its Antigravity desktop app trails Claude Code and Codex by six to twelve months
- Mike Taylor keeps memory off in Claude and ChatGPT and opens a fresh thread per task, naming seven context failure modes including distraction, neglect and poisoning
- Matt Hartman of Factorial Capital backs technical founders on the assumption every early-stage startup pivots, and screens for an "earned insight"
The Sunday Synthesis
04:22 PT
The Sunday Synthesis: October 18, 2026
So what
Karpathy's capability gap explains your portfolio's uneven AI returns — access and ambition separate winners, not model quality.
A link roundup whose two strongest items are tweets: Karpathy on why AI capability looks completely different depending on who you are, and Collison on agents rewriting consumer purchasing.
- Karpathy splits the field into casual chatbot users, professionals on frontier models, and people with internal access — ambition, task type and access explain the gap
- Collison argues personal agents shift leverage from distribution to product quality, since agents can actually evaluate options on a consumer's behalf
- Elena Verna on AI pricing: credits, tokens and usage multiples obscure what a subscription buys; legible pricing becomes the competitive advantage at mass market
- Casey Winters' PM interviews find AI raising output faster than business impact, making prioritization and evaluation the binding constraint
- David Haber on long-running agents owning whole customer journeys and charging for outcomes; Noah Smith compares the debt-financed AI build-out to the railroads
Noahpinion
23:34 PT
Why are we ruled by barbarians?
So what
Trump's exit hands the agenda to an ideology with no charisma and no corruption drag — price that into post-2028 US policy assumptions.
Noah Smith uses the cancelled plan to televise Nidal Malik Hasan's execution by firing squad as evidence that the New Right holds premodern values the rest of America does not share.
- Every demographic segment except MAGA Republican men opposed the livestream; Trump dropped it after the polling
- Smith's frame: the New Right defines Western Civilization as European descent, Christian heritage, and veneration of both — Christendom rather than Christ
- Rubio at Munich and Athens, Hegseth at Normandy, and Vance in 2025 all cast mass migration as an invasion of a single transatlantic civilization
- Musk backs the AfD, Vance met them and campaigned for Orban, Trump supports Le Pen — Smith calls it a Western Civilization Internationale
- The electoral read: pro-immigrant sentiment has risen in America since Trump returned, and Trump is underwater on the issue; Smith argues the ideology, not another charismatic figure, fills the vacuum when Trump goes
AI with Allie
11:07 PT
TUTORIAL: go from ChatGPT.com to building your first agent
So what
Your non-technical staff can run Codex agents today; the gate is a desktop download and a dropdown, not a training budget.
Allie Miller's pitch is that Codex is not a coding tool but OpenAI's agent workspace, and that the only barrier for non-technical professionals is a dropdown they have never clicked.
- At a workshop of daily ChatGPT users, the room audibly reacted when she switched the desktop app from ChatGPT to Codex; an OpenAI employee watching her video did not know one of the features existed
- Her five-step walkthrough: ramble the problem by dictation, then force the model to interview you to 95% understanding before it proposes anything
- Goal mode wants a measurable finish line — she gives "under 500 words, preserve every link, check the final count" as the shape
- Scheduled tasks plus connected CRM, email and calendar can plan a month of in-person client meetings, draft outreach and book the restaurants
- Multi-agent splits and live voice mode with computer use: Codex navigating Chrome while you talk to it in another window
- The whole thing funnels into her AI Agent Mastermind cohort, October 20 to November 12, enrollment closing Monday
Superhuman AI
09:07 PT
🚶🏻 Robot helps paralyzed woman walk again
So what
Rivian routing assistants to Anthropic makes the dashboard a distribution channel — assume cars become a surface you can sell into.
A robotics-only edition: cheap hardware and assistive devices are reaching consumers while the car becomes the next surface for frontier model assistants.
- UCLA's PhoneBot is a 19-inch bipedal robot using an Android phone as torso, camera and processor, with roughly $400 of other parts
- Rivian CEO RJ Scaringe says future vehicles could run third-party assistants from companies like Anthropic, routing simple commands locally and complex ones to the cloud
- DEEP Robotics has DR02 humanoids directing traffic at Chinese intersections; Johns Hopkins has rice-grain magnetic robots that dock in pairs to deliver antibiotics to heart infections
- Volvo expanded its NVIDIA partnership for hands-free driving; Uber and Pony.ai will launch robotaxis in London; Tesla filed a patent for touch-sensing robot skin to fix Optimus hand reliability
- Tau Robotics opened bookings for humanoid home cleaning in Los Angeles; Agile Robots is training on human task video to get past the real-world data shortage
The Information
08:01 PT
Your Office Isn’t Spy Proof
So what
Run a counterintelligence pass on your portfolio's hiring — a $100 socket recorder plus a headcount binge is the entire attack surface.
Abram Brown turns the FBI's arrest of a former OpenAI engineer into a warning that fast-hiring AI companies are structurally easy to bug, and the security consultants he called all agreed.
- Abdulrahman Alfozan allegedly hid a recording device wrapped in toilet paper in an OpenAI conference room; he planted a second but mislaid it and had to retrieve it from lost and found
- Bill Priestap, former FBI counterintelligence head, says recruiting pressure means companies often onboard people they barely understand
- PI David Abbott: a usable device costs under $100 and looks like a phone charger, coffee cup or smoke alarm — and finding one means there is a second
- Strider Technologies has raised $110 million selling AI that monitors internal comms for espionage signals; Greg Levesque says the threat is coerced scientists, not trained operatives, and they leave digital exhaust
- Must-reads: Cursor staff investigated whether confidential data reached SpaceX for AI training; Anthropic's rapid but slowing growth could weigh on IPO valuation; a16z and General Catalyst are racing to secure chips for their startups