Anthropic's Generational Run, OpenAI Panics, AI Moats, Meta Loses Major Lawsuits
In a Nutshell
Anthropic surges ahead with Claude 3.5 Opus, agentic tools like Computer Use, and enterprise coding dominance, outpacing a refocusing OpenAI amid SaaS disruptions and shifting AI moats toward brands, networks, and hardware in an abundance era. Meta faces massive lawsuits over child safety and addiction on Instagram/Facebook, sparking debates on tort liability versus parental responsibility, while PCAST—co-chaired by David Sacks with builders like Andreessen, Brin, and Huang—gears up to counter China's AI/biotech lead through industrial policy. AI accelerates development cycles, erodes traditional valuations, and promises trillion-dollar consumer/enterprise plays, but demands focus amid market resets and regulatory battles.
These notes were generated by AI and may contain inaccuracies.
David Sacks is back in Texas after DC with little sleep. Discussion upcoming on PCAST and Sacks' role in Trump administration.
David Friedberg resembles Thelma and Louise; recent rant on California budget retweeted. Friedberg encouraged to rant on pod. Jokes about Friedberg running for California governor after Ohalo; Friedberg tempted to buy Ohalo for $5-6 billion to avoid politics. Group supports Friedberg for governor to save California's fourth largest economy. Oppo research jokes: Friedberg at rave in 1999 until 10 a.m., stole pistachios after poker tilt.
Anthropic on generational run; OpenAI crashing out.
January: Launched Claude for business users (cron jobs, connects to Gmail, Notion). Claude 3.5 Opus major step function; consensus top model, called inflection point and first agentic model by Jensen (November prediction). Dell: hit threshold for real productivity in teams.
February: Claw code plugins caused SaaS apocalypse ($6B annual run rate added). Computer Use: new agentic system for enterprise-grade open Claude functionality; control desktop from Claude app on phone.
Sacks: Admires Anthropic products; gave credit for Claude 3 last year. Big bet on coding as breakout use case (business or ideological). Most AGI pill of frontier labs; coding path to recursive self-improvement. Good business: code gateway to enterprise IT budgets, rapid revenue growth. Basis for extensions: Claude Code to Claude Co-work (generate PowerPoints/spreadsheets via code). Now agents: Computer Use like open Claude knockoff; Mac Mini run over. Firing on all cylinders.
Past issues: regulatory capture strategy – permissioning regime for chips/models (Washington approval for new models/GPUs worldwide). Heavy-handed; favors big companies, creates moats vs. entrants. Philosophical objection, not product detractor.
Pentagon kurluffle: Not involved, stays out of military procurement (party matters vs. policy). Emil Michael: If sell to Department of War, expect lawful uses.
Anthropic treated same as other LLMs despite criticizing administration, no donations. Conspiracy: Dario anti-Trump to attract PhDs (3-4k sought-after). Friedberg: Dario believes it; culture from beginning, now branding; represents employees' beliefs. Strategic advantage: 90% left-leaning PhDs.
Everything politicized: left/right media, food, AI tools; natural AI market fracture.
Chimath: Both incredible businesses; creating drama unnecessarily. Different GTM motions (may converge). Enterprise action through 80/90: all Anthropic. Philosophical issues with management ideology, capital use beyond tech/R&D. Technical team head and shoulders above; builds vibrant business. Issues: cost, token consumption (tactical, sortable).
Breathless pitting OpenAI vs. Anthropic for drama. Tweet/diagram: Apples-to-oranges; revenue recognition differences. OpenAI: 3/4 consumer subscriptions, 1/4 API. Anthropic: opposite (direct or via GitHub/Cursor). OpenAI conservative recognition; Anthropic gross tonnage. Headlines miss: different businesses guide growth. Normalize: OpenAI overwhelming revenue leader; Anthropic catching up. By IPO, clean normalized stories. Running with numbers for clicks, not business quality.
OpenAI owns consumer with ChatGPT (verb like Uber/Google). Others catching up.
Advice to startups: Focus focus focus; do 1-1.5 things incredibly well; avoid peanut butter smear.
Emil Michael smeared in Lever article: Investor in Perplexity, conflicted in Anthropic dispute. Perplexity LLM-agnostic, wrapper (multiple models, not Pentagon competitor). Emil's shares blessed by Office of Government Ethics. Suspicious timing like Sacks' NYT hit piece when opposing Anthropic. Anthropic hires seasoned political operatives (Biden admin members; Laura Loomer piece). Political operation, ruthless.
Perplexity features: Model council – queries three major models (incl. open source), shows differences/reasons. Great co-work competitor Computer.
OpenAI consumer market share: 100% (2023 category creator), 85% (2024), 75% (2025). Market growing tremendously (queries/searches), but share down. Competitors: Apple, Meta, Windows underrepresented (0.5 pt each → OpenAI under 50%). ChatGPT consumer challenges.
Cutting side projects: Sora video app shut down (Disney $1B investment, licensing for Disney Plus cancelled).
New focus: Chase Anthropic in enterprise. Offered PE investors 17.5% guaranteed minimum return in JV to deploy AI, ease upfront costs (rollups in services/accounting/legal; Josh Kushner effort; 8090 software factory).
Chimath: OpenAI incredible consumer mindshare (kids sticky; hard switch even to Gemini unless stumbled into). Healthy market segregation. Postulate: OpenAI consumer win → multi-trillion company. Decide: strongest traction? Enterprise bleed to consumer? Win enterprise if yes. Enterprise vs. consumer: different features/expectations. Pick one and nail: OpenAI pick consumer (juggernaut, brand).
Friedberg: Consumer queries not all free. Spotify 290M subscribers (~$20/mo global avg); Netflix 325M. AI (book travel, questions, calendar, email) most valuable meta service. More subscribers than cable TV. Cell phone $50-60/mo (pandemic: kept cell/electricity). ChatGPT like that. Ultra-valuable; layer services (video, finances). Ecosystem like iPhone app economy: connectors/embedded tools; advertisers pay to embed (consumer pay or ad).
Sacks: Possible few hundred million premium subscribers. Most take free ad-supported (successful). ChatGPT displacing Google search (not 10 blue links; compelling AI chat). Ad-supported + premium both work. Investor pref: B2B sticky, upsell, >100% NDR vs. consumer low WTP/high churn. But top companies consumer-first (Meta/Google/Apple). Both models work.
Determinant: Google/Facebook bridge ad networks to AI. Facebook MIA; Google vigorous (existential).
Google will compete vigorously for consumers because it's existential to them. Search and AI chat are merging, so ad links will merge into chat advertising. They must adapt and compete for the consumer. Google is in an outstanding position for the agent play because they already have access to your calendar, documents, and email. The agent doesn't have to earn trust because users already trust Google with their data. Expect the Google version of OpenAI's Operator soon, as people don't want to share documents with new services. Google has so much free cash flow they can view it as two separate companies: GCP for enterprise and Google consumer for chatbots, keeping them segregated. Startups struggle with organization and financing without a profit engine. Valuations reflect belief in Google's durability more than others.
Google announced Google Workspace Studio for AI automation, already online and playable. They joined the Operator party. The PE story reflects the broader market: what are these companies worth? Key fork: path to superintelligence with infinite abundance where describing things makes them appear, or just good next-gen software? Financing assumes the former. GC is buying accounting firms, healthcare firms, hospitals. Future VC looks like PE buying hospitals, accounting firms, BPO in India, then running with AI.
PE bets on owning change management around AI. Assumption is businesses automatically use AI for efficiency, but McKinsey studies show 95% of enterprise pilots fail. Latent value exists but deployment is hard. PE owns the business and change management to create value. Their model makes solving this existential.
Market debates PE ratios based on cash flow durability: Facebook 30x, Nvidia 40x, Tesla 200x asymmetric, Caterpillar/John Deere 15x. Public markets guess when cash flows end. With superintelligence, what is anything worth in year 10-20? Infinite abundance disrupts all companies with constant churn. Canary in coal mine: SaaS talks, SaaS apocalypse. Capital markets view company health differently with superintelligence making everything fragile. Market rerates SaaS down, shifts from P/E to cash on hand multiple. Implications for Silicon Valley and employees: equity dream assumes 15-20 year gigantic value, but 5-6 year disruption cycles mean just cash. Employees rationally demand more salary over equity, changing valuation multiples.
Chart of SaaS companies (Mag 6) and Mag 7: market cap / annual free cash flow shows payback years. Snowflake 2023: ~100 years, now halved. ServiceNow, Workday similar cuts. Public markets rerationalizing: if superintelligence coming, careful on multiples. Mag 7: Apple, Microsoft, Meta, Alphabet multiples ripped up, believing monopolistic durable cash flows forever. Nvidia accretive, well-run, $200B revenue, high margins, but treated like ServiceNow/Snowflake.
Reset is flight to quality free cash flow in Mag 6. Big tech (Google, Apple, Microsoft, Facebook) not disrupted despite predictions; they've gotten nimble at copying/incorporating features.
Toma Bravo LP slides highlight companies integrating AI, with high-quality people, beachhead, customer access, enterprise users. Selective dispersion: winners integrate AI for 10x output with same capital/labor. In superintelligence world, what do you pay for business durability?
Discount rates rising, valuations collapsing due to uncertainty on multiples, discount rates, terminal growth. Counter-AI: physical experiences, high asset low obsolescence (HALO). Examples: LNG (Cheniere), real life mining, space industry (15-30T/year moon opportunity, SpaceX IPO insane multiple). High discount rates steamroll some, but AI unlocks others or leaves unaffected, flowing capital there.
Great moats: brands, network effects, management teams. Management not a true moat (Buffett: businesses strong enough for monkeys). Elon/Tesla relentlessly innovates, especially agentic AI age. Moats subtle: network effects, physical production difficulty. Key in digital abundance. Apple: ecosystem network, hardware, custom silicon, brand. Tesla: hardware, Elon innovation. Meta/Google: brands, management, constant innovation.
Bet brands go to zero: abundance cheaper/faster/better trumps affiliation. Tesla disrupted BMW, Mercedes, Porsche; BYD/Geely in China. Model Y outsells due to better price/operations. Premium luxury eroding (LVMH, Ferrari stock charts show pricing power loss). Value brands like JetBlue, Tesla Model Y, Apple MacBook Neo ($600-700). Brands bringing abundance at same/less cost capture share.
AI disruption lowers costs: Model Y cheaper, Cybercab, BYD $15k everywhere, Apple Neo $600 laptop. Hard to predict disruption. Year ago, Google toast from ChatGPT taking search/AdWords share. Now bullish on Gemini/agents; stock doubled. Counterfactual: agents so good no phone check needed, no app wall (e.g., voice Uber). Enterprise software clients want strangulation as a service: hide UI/products, natural language shim for complexity.
Tracks OpenAI Operator, Perplexity, Claude Co-work: query schedule without Notion/Gmail/Calendar. Dumb $100 device chat terminal suffices. Counter: still need dashboard/UI for readouts, maps (e.g., Uber location). Even Siri++ needs Apple UI. Apple debate: brilliant missing AI wave/no capex, or missing capabilities? Hard to know.
Liquidity May 31-June 3. Chimath took control of All-In Corporation LLC speakers. Announced: Dan Loeb, Sarah Frier, Bill Ambron, Andrej Karpathy. Four heavy hitters, six more, surprise guests. Andrej: 5-10 min slides on AI future + fireside. Bill Aman points of view. Freeberg shoutout: coded HR replacement, new website, doubled CTR via autoresearch (formerly many man-months). OpenAI growth recipes/playbooks incredible. Every day new era, disorienting: pull 1-2 year projects into 3 days, no hiring. Sequencing compressed, ideas rush in. Bought annotated.com 15 years ago ($4k) for bookmark service: highlight NYT paragraph, add comments, save.
A developer vibecoded a Chrome extension in a weekend for a service sitting on a domain for 15 years. This manifests ideas instantly, like Star Trek replicators: "CRM system, build annotated.com, build this new website for $80-90. Boop."
Speakers share experiences of tech waves:
- First felt watching dot-com boom as derivatives trader in Toronto, joined Winamp (original iTunes), missed financially but was in the right place: "You got to the beach with a surfboard."
- Crushed mobile and social wave.
- Current AI wave feels 100x bigger, a tsunami like Nazaré's 100-foot waves: "Glory or goodbye."
Personal moments: first PC, internet and Mosaic in early 90s, iPhone.
Rough week for Meta: two verdicts in two days.
- New Mexico jury found Meta liable for allowing child predators access to minors on Facebook/Instagram. Ordered $375 million damages. AG's office created fake child profiles; predators contacted them, showed up. Whistleblower, former Meta engineer, testified his 14-year-old daughter received sexual solicitations on Instagram.
- LA jury found Meta and YouTube negligent for designing addictive platforms harming youth mental health. Plaintiff (Saxs) started YouTube at 6, Instagram at 9; notifications, algorithms caused compulsive use leading to depression, anxiety.
Jonathan Haidt (interviewed jointly) pushes keeping kids off cell phones/social media until 16, cheerleading verdict.
Contrarian view on tort litigation: costs US economy $900 billion/year (3% GDP), growing 10%/year. Hits big companies, food, restaurants—tort tax. Companies pass costs: less R&D, product development, fewer launches, restrictions.
Agree social media harms kids immensely; ban until 16 (maybe adults too). But question responsibility:
- Sugarcane farmers for diabetes? Soda companies/retailers/FDA?
- Parents absent, letting kids play with guns, go to sketchy areas, 100 hours video games, eat soda/chips, use screens/social media.
Personal responsibility key, especially with AI flooding options. Human agency distinguishes success/failure.
Counter: Companies knowingly damage (RJR Nabisco cigarettes addictive, hid risks; auto seatbelts; asbestos, lead paint). Facebook knew addiction damaging to young girls, leaned in without safeguards?
Response: If legal, who's responsible for use? Whiskey to alcoholic? Potato chips to obese? Age gating, labeling (e.g., alcohol in pregnancy covered up, led to 70s/80s bar signs). Tobacco/alcohol bad but choice, publicize risks.
Product liability wrong; emphasize human/parental responsibility. Tort lawyers biggest political donors (Dems local, GOP national, judges); litigation business.
Don't want government/companies dictating life.
Trial lawyers want Meta as next tobacco, but social networking ≠ smoking (harms subjective, benefits too; assumption of risk). Not all harms immense/known—up to parents.
Parental empowerment: screen time limits, app controls.
AI apps: Parents want bans after self-harm cases. But good for 10-year-old using ChatGPT for research, AI-native skills (China K-12 AI education). Monitor, but don't ban.
LA case: 20-year-old won millions for depression/body image from social media. Causation issues: abusive home, father abandoned, mother body-shamed her.
Dangerous precedent: Sue Spotify for sad playlists? Free services—choice to stop if harmful. Lawyers create next tobacco.
Consensus: 2-3 hours/day correlates with depression/anxiety/eating disorders in young girls (many studies). Chamath (ex-Facebook, Stanford 2017/2018): Doesn't let kids use; not constructive for developing child.
AI chat vs. social: Different guardrails, cul-de-sacs for dark topics. Social: fast-twitch algorithm; outcomes compound until incentives change.
New Mexico $375M child exploitation; LA individual $3-6M. Navigated Section 230 via product liability—map drawn, floodgates open, death by 1000 cuts.
Want kill switch under 16/18, better age validation. COPPA useless—6-year-old vibes around it. Harder for porn sites than unverified Facebook.
White House AI framework: online child safety (social + AI). Age assurance, parental controls. Social not pure ills—let parents decide.
Dopamine addiction. Apple's Family Sharing hard: Schools give Chromebooks for LMS/homework with YouTube backdoor (shorts).
Industry won't self-police (addicts adults too). Australia/Malaysia enforce 16; Spain/Germany/UK following. US: Handset makers (Android/Apple) default age verify on family plans, opt-in access.
Jonathan Haidt: Phone lockers in schools—kids complain then love.
School phone bags ineffective. High school social pressure huge: No Instagram/TikTok until 16, but Snapchat needed or "no friends, sit in dark room."
Other parents won't agree uniformly.
Compromise: Snapchat in high school, Instagram/TikTok at 16. iMessage indispensable (location tracking), but kids want Snap: "Dad, don't be a loser."
Slippery slope banning as parent.
Tip: Replace kids' headphones with over-ear brand like Shocks (goes over
Discussion on over-the-ear headphones allowing listening to audiobooks or music while remaining aware for conversations, reducing drama around usage. Debate on lawsuits for emotional distress when kids turn 20 over childhood headphone use from ages 6-10. Consensus that such claims are crazy and parents bear responsibility.
President Donald Trump, the 47th president, announced his Council of Advisors on Science and Technology (PCAST). Saxs appointed as member and co-chair alongside Michael Kratsios, director of OSTP. Saxs continues as AI adviser on behalf of PCAST, expanding from previous SGE role (130 days used) limited to AI and crypto to broader issues including nuclear power, quantum computing, advanced semiconductors, and biotech.
PCAST Members
Members include Mark Andreessen, Sergey Brin, Michael Dell, Larry Ellison, David Freeberg, Jensen Huang, Lisa Su, Mark Zuckerberg, plus Nobel Prize winners in physics and experts in quantum computing, fusion, nuclear, biotech.
Emphasis on doers and builders who created products/companies like Mark Andreessen (first internet browser), Jensen Huang (GPU), valuable for recommendations on advanced semiconductors.
Role Division
Saxs focuses on tech; Freeberg on science.
Freeberg honored by appointment, notes PCAST roots in FDR era amid nuclear physics and quantum mechanics breakthroughs. Current era mirrors with AI reinventing possibilities and acute race against China.
- 10 years ago, China published 50% of US scientific research papers in peer-reviewed journals.
- Last year, China published 50% more than US across physics, material science, chemistry, biochemistry, broad life sciences.
China shifting from copycat in biotechnology to leader in subdomains, threatening to engulf pharmaceutical industry and lead in medicine and foundational AI. Focus in DC and Silicon Valley on China's advances in research discovery and industrialization.
Defends industrial science/technology leaders on PCAST for this critical industrial race, not just discovery race.
Chamath joining president's advisory council on bomb pots; Wagering joining. Joke about President's Council on Podcasting with Lex Fridman.
Saxs thanks president for co-chair role; PCAST modern incarnation from 1990 by George H.W. Bush, over 30 years. New take emphasizes builders/doers for industrial policy catch-up. 15 members named, 9 more slots (up to 24 total); possible second round for missing expertise, president's decision.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
