Lenny's Newsletter
05:32 PT
90 minutes of unfiltered product advice from Snap and Discord’s product chief | Peter Sellis
So what
Use the core-product-drives-growth claim as a blunt instrument on any portfolio company staffing a separate growth org before its retention curve flattens.
Peter Sellis — Snapchat's first PM, later Discord's head of product — talks with no employer to protect, and spends most of the episode on why product organizations and most PMs are bad.
- Designs teams like cell-structured organizations: small, autonomous, resilient to the loss of any one node
- Argues growth almost always comes from the core product, not growth teams bolted onto the side
- Post-mortems why Snap's ads business never reached its potential, with Eric Seufert and VCG auction mechanics in the background
- Offers three "oxymorons" of great product management and a candid account of managing Nikita Bier
- Systems thinking runs through it: Thinking in Systems, Forrester, five whys, the Peter principle
Every
05:01 PT
Copy Our Homework
So what
Jev-style sub-cent instant classifiers make per-row LLM judgment economically viable in pipelines where you currently sample or skip entirely.
A Sunday roundup format — "Steal these" — collecting the week's nine concrete agent workflows, plus vibe checks on GPT-6 Astra for VFX and a new instant-classifier model called Jev.
- Randy Counsman's unbounded agent swarm burned 4.5 billion tokens over a day and a half on one Blender model; the fix caps the orchestrator at five subagents plus a judge agent with a reference image
- Marcus Moretti routes by cost: Sonnet for basic checks, Opus for tightly scoped changes, Fable 5.1 to plan large features and delegate coding
- TypeSafe's Jev returns a probability for a plain-English yes/no question in under a second; Mike Taylor scored the entire Every archive for a fraction of a cent
- Afterimage tested Astra on VFX: convincing on rigid objects like glassware and buildings, still fails on living things or anything contacting the scene
- Anthropic merged Chat and Cowork into one Claude that selects its own tools — Every reads this as convergence with OpenAI on a single desktop app for knowledge work
Noahpinion
02:51 PT
How we can beat China
So what
Stop underwriting allied-bloc industrial policy as a durable tailwind; price China exposure against regional self-defense demand instead of US-coordinated containment.
Noah Smith publicly abandons the strategy he spent the Biden years advocating: a US-led allied economic bloc to overmatch China is no longer available, because America is too internally divided to execute it.
- The Doshi/Campbell allied-integration approach requires a unified, reliable US; Smith thinks reconstituting that takes decades, exactly the window of China's peak relative power
- Argues Trump has forfeited the moral case — invasion threats, detention infrastructure, civilizational rather than rights-based foreign policy — so more American power no longer means a more liberal world
- Redefines the goal from beating China to preventing Chinese domination of smaller states: Japan, India, the Philippines, Korea, Ukraine, the Baltics
- Expects no GOP reversion to Reaganism and only marginal Democratic reversal, making the shift structural rather than cyclical
Superhuman AI
09:09 PT
🚑 Robots now have their own 'ambulance' services
So what
Field service and motion-data capture are the picks-and-shovels layer here; Mecka's $500M mark is your comp for the next data-collection deal you see.
A robotics special arguing the sector has crossed from demo to deployment infrastructure — generalization, safety certification, factories, and now repair fleets.
- Figure's Helix 2.5 did zero-shot chores in 30 unseen homes: tidying, folding and storing towels, making beds
- Agility's Digit 5 detects humans and stops or powers down on approach, removing safety barriers; $300M in multiyear orders, North America launch early 2027
- UBTECH's Liuzhou plant is the first built for 10,000-unit humanoid scale, using robots to assemble robots, four hours of testing per unit
- GMO unveiled a humanoid ambulance — a repair van with spares and a backup robot to swap in mid-shift
- Goldman projects 6.5M humanoids shipping annually by 2035; Mecka AI nearing a Sequoia-led round at $500M for human motion data; Boston Dynamics unlikely to IPO in 2027, still unprofitable
Interconnects
08:42 PT
Where I stand on RSI
So what
Model frontier-lab economics as falling cost per unit intelligence rather than a capability discontinuity — that changes which inference-layer businesses survive price compression.
Nathan Lambert plants a flag against the intelligence-explosion consensus forming inside OpenAI and Anthropic, arguing what the labs are actually experiencing is thousands of agents working well — not recursive self-improvement.
- His baseline stays "lossy self-improvement": automatable research is too narrow, parallel agents show diminishing returns, and resource and political bottlenecks resist AI acceleration
- Cites Anthropic's Fable 5.1 and Mythos 5.1 system card: internal model use is key to maintaining the current rate of progress, with no clear signs of dramatic acceleration beyond it
- Dwarkesh timelines he summarizes: Schulman says ~2 years to 10x researcher uplift and 3-4 years to ASI; O'Neill says 5-10 for both; Beren ~5 years
- Predicts RSI lands on efficiency, not peak intelligence — LLMs get much cheaper at fixed capability, Jevons prevails, margins improve
- Quotes Schulman that post-training resists automation because someone must decide how the model should behave, and it's easy to break in ways benchmarks miss
- Calls the escalating extinction-risk discourse very misplaced and notes 2023-24 risk timelines did not arrive
McKinsey & Company
08:06 PT
The quarter’s top themes
So what
Your enterprise buyers are still reading books and human-skills essays over deployment guides — sell transformation narrative, not capability, at the CEO level.
A quarterly popularity ranking of McKinsey's own posts, which is mostly useful as a read on what enterprise buyers are clicking.
- The No. 1 item is a CEO book recommendation list, ahead of every AI piece
- Rewired 2.0 claims successful AI transformation lifts EBITDA roughly 20 percent, drawn from 30-plus case studies
- Human skills in the AI era ranks second, framed around AI agents joining the workforce
- Indosat Ooredoo Hutchison is the featured AI-native transformation case, emphasizing operating model and partnerships over technology
- The latest McKinsey Global Survey on AI shows deployment spreading across more functions than in 2024
The Information
08:05 PT
Dario Amodei's Tabloid Era
So what
Political risk around your AI-safety-aligned holdings is now media risk, not just regulatory — reputational attacks will precede any legislation.
Abram Brown argues the New York Post's dozen-plus stories on Dario Amodei and effective altruism in one week mark a turn: AI safety advocates are now populist villains, not obscure technologists.
- Post headlines called Anthropic staff "cult-like" and METR "super-woke globalists"; Trump called Amodei's slowdown proposal a "SICK conspiracy"
- Brown notes the contradiction: a Politico survey found 63% of US adults think AI could eventually destroy the world, yet the people saying so draw scorn
- He disputes the cabal framing — the coverage names only Dario, Daniela Amodei, and Dustin Moskovitz on repeat
- The Information's debut 50 Most Influential People in Tech includes the Amodei siblings at the top of Anthropic and no other EA-affiliated figure
- Also reviews Sharon Weinberger's Valley of Death, where a VC tells her defense-tech returns depend on lobbying and government relationships, not the product
Every
08:02 PT
Every NYC Meetup, September 24: We let the Every agent plan it
So what
Treat this as the cheapest available demo of agent-run operations: event logistics is a real end-to-end workflow, not a benchmark.
A short invitation to Every's September 24 Brooklyn meetup, notable only because the agent @Every planned the entire evening.
- The agent set the run of show, the food, the bar, and wrote the invitation itself
- It read the company Slack and invented a drink from what it found there
- Free for subscribers, 6-8 p.m. ET at Every's Brooklyn brownstone