Anthropic’s $30B Ramp, Mythos Doomsday, OpenClaw Ankled, Iran War Ceasefire, Israel's Influence
In a Nutshell
Anthropic withholds its powerful Mythos model after it autonomously discovered thousands of critical vulnerabilities in OSes and browsers (e.g., 27-year-old OpenBSD bug), launching Project Glasswing—a 100-day AI cyber defense coalition with 40 tech giants to patch before release—signaling AGI-level risks and a pragmatic safety blueprint over regulation. The company aggressively ramps to $30B revenue run rate via Claude Code and enterprise adoption (1000+ firms paying >$1M/year), dominating 50%+ coding tokens amid flywheel effects, while cutting off OpenClaw access and copying its agent tech, sparking antitrust concerns. A US-Iran war ceasefire holds after Trump's threats, with Netanyahu's influence criticized; X's auto-translation boosts global discourse amid AI hype proving bubble doubters wrong.
These notes were generated by AI and may contain inaccuracies.
How many PRs do you think are going to get pushed to the core structural internet in 100 days? Over-under number: zero or 10,000, but even 10,000 would be meaningless. If it prevents browser history or dick pics from being released worldwide, worth letting 100 days pass.
Welcome to the number one podcast. David Friedberg out, replaced by Brian Gersten. Namaste reference. Accusations of being Ro Khanna apologist for noting shift from retard maxing to logical actions.
David Sacks back. Trimmed hot air and fluff. Promo for thesyndicate.com.
Jamal the Polymath. Retard maxing discussion. Ups and downs in the arena to be a winner.
J. Cal's interest in retard maxing. World caught up; been doing it without a name. Eli's videos good. Appeal: not ruminating, smoking cigars, living life. Works at every level. Detachment healthy: let go, live life, try stuff or don't. More you want something, less you get it.
Elisha Long (Eli), modern-day philosopher, new age sage, Lisan al-Gaib of modern internet. Marc Andreessen found him. YouTube channel. Roadmap for life. Build something, don't ruminate, stop blathering in head, do stuff.
Anthropic withholding newest model Mythos (Greek pronunciation), too dangerous. Autonomously found thousands of vulnerabilities: bugs in every major OS and web browser. Included 20-year-old exploits missed by security audits for decades.
Examples:
- 27-year-old vulnerability in OpenBSD (firewalls, critical infrastructure).
- 16-year-old bug in FFmpeg, missed by automated tools after 5 million scans.
- Linux kernel bugs.
Released hype video. Dario Amodei quote: > "Come on the program anytime, brother. But as a side effect of being good at code, it's also good at cyber. The model that we're experimenting with is by and large as good as a professional human at identifying bugs. It's good for us because we can find more vulnerabilities sooner and we can fix them. It has the ability to chain together vulnerabilities. So what this means is you find two vulnerabilities, either of which doesn't really get you very much independently. But this model is able to create exploits out of three, four, sometimes five vulnerabilities that in sequence give you some kind of very sophisticated end outcome."
Brad (investor in Anthropic): Deserves ton of credit. Could have released Mythos, broken core internet things (move fast and break things), but realized havoc: offensive hacking, expose browsers, browser history, credit cards. No government needed. Set up Project Glasswing: AI-driven cyber coalition with Apple, Microsoft, Google, Amazon, JP Morgan, 40 important companies. Goal: 100 days to find/fix/harden vulnerabilities before hackers exploit.
Threshold crossing: Mythos and OpenAI's Spud (first Blackwell-trained model, out soon) = beginning of AGI models with massive step-function intelligence improvements. Too smart for immediate release. Sandboxing, defensive alliances. Trust industry/market forces coordinating with government, not top-down regulation. Blueprint: pragmatic, aggressive to win AI race while protecting safety. OpenAI/Google will follow.
Sacks: Anthropic good at product releases and scaring people. Pattern: roll out model + study on worst implications. Examples: blackmail study (prompted 200+ times for result, reverse-engineered for headlines; no blackmail in wild despite open-source models at same capability a year later).
This cyber case more legitimate: coding models find bugs, string vulnerabilities into exploits. Next 6 months: AI-driven cyber detects dormant 20-year bugs. Pre-release period sensible: give capability to software companies for detection/patching before wide release. Chinese open-source like Gemma K2 6 months behind. Window of 6 months. OpenAI releasing similar soon.
Every company/IT/CISO should take seriously, detect dormant bugs, roll patches. If all react, no doomsday. Fear drives correct behavior. Then arms race: AI cyber offense vs. defense.
Chamath: Mostly theater. Feb 2019, Dario at OpenAI did same with GPT-2 (1.5B parameters, end of days for spam/misinfo; methodical rollout over 6-9 months, nothingburger).
If Mythos capable: sophisticated hacker can do it now with Opus. Exploits easy = shut down internet 5 years to patch all, spend trillions (e.g., G-SIB bank). 100 days/2 months won't fix; even 6-9 months no. Capitalism/funding/adoption moves forward. Clever GTM: hyper attention/usage. Anthropic shooting lights out (Steph Curry/Clay Thompson).
Two takeaways: 1) Patching all needs years of internet shutdown. 2) Advanced hacker can do today with Opus.
Brad: Code red, DEFCON, deadly serious. Defensive/offensive opportunity. Coordinate with CIA (already happening; Dario informed them). Not nationalization. Frontier Model Forum (2023) cooperating on anti-adversarial distillation (Google, OpenAI, Anthropic).
No moratorium call (unlike Elon's 2023 6-month call). Instead: 40 companies, 100 days sandbox/harden, push forward. 10,000 PRs meaningless but prevents browser history/dick pics leaks. Compliment market over nanny state.
Sacks nuanced: No choice but treat seriously. Logical: code capability → cyber improvement. Pre-AI to post-AI patch period. Give benefit of doubt (past blackmail study hoax, but this legit).
Even if performative (dramatic video), Anthropic bet big on code, discovers quickest. 100 days advantage vs. hackers. Code pushed 10x-100x now; embed real-time security in tools.
Critique: Wrong companies (no energy/nuclear, airplanes). Expand circle. Those are customers anyway.
Dario ghosted invites despite intros. His Dwarkesh podcast excellent.
Peter Steinberger, founder of OpenClaw—the project that launched the agent era—tweeted that Anthropic was cutting off his access. OpenClaw users on Anthropic's $200/month subscription were using 100x more tokens than average due to its verbose nature. Anthropic required switching to the API, adding a zero or more to token costs. Shortly after, Anthropic announced their own agent technology as a safer, better version of OpenClaw.
Anthropic systematically copied OpenClaw features into Claude, then cut off access. Subscriptions blend usage across users: 9/10 use less than paid for, top 10% use more. OpenClaw became the #1 open source project on GitHub, with $200 subscribers using $2,000-$20,000 worth of tokens. Anthropic ended unlimited subscription use for OpenClaw, forcing API pay-per-use.
Anthropic's agent harness may be bundled in flat-rate plans. This could be price dumping (pricing under market) or bundling, especially if they subsidize their version while charging metered rates for third-party like OpenClaw. OpenClaw can still use Claude API; companies set their prices. They were selling dollars for 10 cents via power users, now pricing rationally.
Their closed beta agent isn't priced yet. If bundled flat rate for theirs but metered for others, it's anti-competitive, assuming dominant market share in coding. Enterprises prefer Anthropic's version for security; integrated with data warehouses/lakes. Letting OpenClaw loose on datasets isn't wise—it's a different product.
Anthropic has huge advantage cloning OpenClaw into Claude. Sam Altman hired Peter Steinberger (not buying OpenClaw) to subvert the open source project and get his ideas inside OpenAI. Competitors trying to kill/compete with OpenClaw:
- Anthropic's Claude managed agents (released Wednesday, April 8th)
- Perplexity's computer launch
- Hermes agent (open source, released February 25th)
- Ali Baba's agent on Quinn model
- Elon's Grok computer from xAI (play on Microsoft)
- Amazon/Apple new Alexa and Siri (less limited)
- Nothing yet from Satya Nadella/Microsoft
Everyone building multi-function agents (question-answering to action-taking) because that's what consumers/enterprises want. OpenClaw like open source Android—disruptive. Open source will win, take 90% token usage, undercut frontier models. Biggest threat: SLMs (smaller language models) verticalized for desktops/laptops.
Don't give your entire business/life to Anthropic or OpenAI.
Foolish to feed lifetime knowledge into LLMs unless boosting investments.
Does Anthropic have dominant market share in coding? Yes (Jacob), yes but not dominant/lead (Brad), >50% coding tokens (Sacks). 50%+ is dominant, even in trillion-dollar market with <10% penetration. AI coding is 5% of broad market—mostly white-sheet de novo code. Long-horizon enterprise-grade software building is still shit; doesn't work out-of-box for $500B banks/$100B insurers—hand-tuned.
50 years tech debt: hundreds of trillions lines mediocre/bad code, legacy languages. Customer gets 60-year-old pensioners for COBOL/FORTRAN interpretation ($100B/year revenue company). 99% tech debt poorly addressed; untangling Gordian knot takes decades.
BitTensor Tau project Ridges AI (subnet 62): open source, anyone contributes. Spent $1M in Tau rewards, hit 80% of Claude 4 in 45 days via anonymous contributions improving coding (like Codex/Claude Code). Flywheel racing like Bitcoin.
Zero probability important companies outsource production code to open source. But open source training/orchestration disruptive: Venice project. Orthogonal to raising tens of billions—capital markets may run out of $10-20B checks.
Capital markets may run out of $10-20 billion checks, so the solution is total distribution. There will be a very successful open source project for pre-training. No real company with skin in the game will open source their codebase for re-engineering. Coding tools will succeed; history shows open source projects like Linux, Kubernetes, Apache, Postgres, Terraform deeply embedded in enterprises. 15-20 years ago, similar arguments were made against enterprise adoption, favoring Oracle, but open source is taking hold in startups—the tip of the spear—with subscriptions at $29 vs. $200. Invest in things with 90% chance of going to zero; do your own research, no crying in the casino.
Code market (AI-generated vs. human) is 5% today, heading to 95% over next few years. Early leader with 50-60% market share gets most developers, codebases, and training tokens, creating a flywheel and potential data scale effects. Market for coding tokens could be monopolized. Anthropic's revenue run rate is fastest growing at scale ever seen.
Coding advances agents, as agents need to write code for tasks. Companies should play fair: no anti-competitive tactics, discrimination, unfair pricing—keep nose clean for government scrutiny with 20/20 hindsight.
Anthropic revenue: early 2023 turned on revenue via API; end 2024 $1B run rate; Feb 2025 launched Claude Code ($4B mid-2025, $9B end-2025); April 2025 $30B run rate (tripled). Over 1000 enterprises paying >$1M annually—coveted big fish like Slack, Salesforce dream of. Dario Amodei (ex-OpenAI co-founder, left due to disgust with Sam Altman per New Yorker/Ronan Farrow story) focused on enterprise.
Added $4B revenue Jan, $7B Feb, $10-11B March—Databricks + Palantir combined in one month. Started year with bubble fears; now largest revenue explosion in tech history. Hit $30B run rate by March (planned for year-end). Model/product capability hit near-AGI threshold: labor augmentation/replacement, not IT budget. Co-work growing faster than Claude at same stage. Near-infinite TAM for intelligence; millions of self-interested parties (1000+ enterprises >$1M) demanding product despite throttling—makes them better at business. Doing this with 1.5-2GW compute (adding 3GW this year). 65-70% token consumption in big companies is open source (cheap Chinese/etc.). Jevons paradox: unit cost of intelligence plummeting as capabilities improve. Major company: $100M token run rate on $5B OPEX, nearing peak employment but intelligence consumption rising.
Low penetration: global 2000, use cases, within use cases; models improving. End-year: $80-100B run rate possible. OpenAI imminent incredible model will inflect their revenue too.
Utility proven (!). Revenue ramp proven (!). Can they be profitable? $400-500B min investment into data centers to climb out. Investor list: can't talk profit → EBITDA → margin → revenue → gross revenue (where we are). Confusion on Anthropic gross vs. OpenAI net, take rates, run rate vs. recognized revenue. Years from steady-state free cash flow/ EBIT. Big cost: compute—fixed at 1.5GW regardless of $1B vs. $80B revenue, so gross margins exploding (from negative 18mo ago to 50-60% rumored; not negative). Accidental profitability: can't spend revenue fast enough on compute (2500 people vs. Google's 120k at same revenue). Inference cost down 90% YoY. Gross vs. net: single-digit % via hyperscalers (like Snowflake/Databricks commissions); adjust 5-10% conservatively—distraction.
Incumbents (Meta, Google, SpaceX) want compute problem (control field; Meta performant due to infra). Anthropic/OpenAI want efficiency. Early innings.
Early 2025: AI bubble claims (hundreds $B capex, no ROI). Doomers contradictory: AI too powerful (job loss) or not (no payoff). Now proof in coding vertical alone—fastest revenue growth ever. Agents next, massive enterprise. Silicon Valley bets intuitively on 'build it they will come' ($350B top-4 hyperscalers capex this year; Jensen $1T by 2030). Pays off, keeps US dynamic/leading.
OpenAI: hand-wringing, employees leaving, consumer-first strategy question; shut Sora, unwound Disney; secondary trading below last valuation. Anthropic >$380B. Many great OpenAI founders/contributors now at Anthropic/others.
Anthropic executed: focused coding/Co-work path to AGI/ASI (no multimodal/video/hardware/chips/data centers); 2500 people aligned; kicked OpenAI's ass last 90 days.
Don't count out OpenAI: peak FUD; great researchers/models. SPUD (or Codex 5.4/5.5) imminent—excellent, on par with Mythos, usable; Codex fastest ramp. Not zero-sum: massive TAM. Both on wave; buyer of OpenAI shares. OpenAI awakening to regain mojo.
Don't count out Meta, Google, Elon (impressive drops soon). Team America: 5 frontier models competing (incl. Mythos self-imposed safe harbor, no moratorium/gov't); accelerates lead vs. world. Stop adversarial distillation, distribute products globally. Polymarket IPOs before 2027: SpaceX 95%, Cerebras 94%, Anthropic 51% EOY, 44% chance.
Polymarket closing market cap for Anthropic shows $158,000 in volume. Chamath's $400K bet would tilt the market. 78% chance above $600 billion, 19% chance it doesn't resolve. Brad Pacheco first invested in Anthropic at $130 or $150 billion round, making it a 7X or 5X return for Altimeter. Many invested earlier and will do better. Altimeter has billions in both Anthropic and another company.
Existential concerns in venture: IGV stock index down 30% YTD, down 5% today, software stocks plummeting. VCs terrified to invest outside frontier models, SpaceX, military modernization—seeking AI-safe terminal value. Crowding into these areas. Barbell approach: heavy in frontier companies and small defensible teams in AGI world. Enterprise software fate: PE takeouts, consolidation, or AI adoption?
Market too pessimistic on some software companies. Software cheaper to generate, but moats differ. Baby thrown out with bathwater; value buys available.
AI value capture across stack layers: started with chips (NVIDIA first to trillions). Hyperscalers benefiting. Now model layer (Anthropic, OpenAI huge). Next: application layer—does it go to models or turbocharge apps like Palantir? Hope for beneficiaries at all layers. What about Salesforce, HubSpot, Oracle?
Too early to tell. Energy/data centers clear winners.
Liquidity event sold out, waitlist of hundreds. Top-tier speakers. Brad joining. Producer Nick sold out Tiffany spots.
Chamath in Wikipedia peanut links article, large category top 5% (joke from Nat/Claude).
Chamath not afraid of hacks—nothing more embarrassing than his own pod comments.
Liquidity agenda incredible: NAP poker, dining, five-star/six-star (Chamath demanded six-star, no mids, elite only).
Liquidity 2026 programming: going well. Next speakers: Brad Gerstner, Thomas Lafont of Kotu. Three confirmed for best ideas pitch, each running $1-2, 6-7 billion (superstar compounders, new stone). Lineup: Bill Ackman, Andrej Karpathy, Dan Loeb, Thomas Lafont, Brad Gerstner, Sarah Fryer, more to come, surprises. J-Cal has surprises.
Fifth annual All-In Summit in Los Angeles, September 13-15, allin.com. Sacks can attend for business reasons.
Two weeks into ceasefire (started two days ago). VP J.D. Vance, Wyckoff, Jared Kushner headed to Islamabad, Pakistan, for talks this weekend on peace deal.
Easter Sunday Trump Truth post: "open the strait and crazy bastards or you're going to be living in hell. Just watch. Praise be to Allah."
Tuesday: "a whole civilization will die tonight, never to be brought back again. I don't want that to happen, but it probably will."
8 p.m. deadline; at 6:30 p.m., POTUS announced two-week ceasefire if Iran opens strait. Quote: "We received a 10-point proposal from Iran, and we believe it's a workable basis on which to negotiate." Maybe tollbooth on strait, US takes majority, split with Iran.
Netanyahu interpreted ceasefire as level Lebanon, dropped 160 bombs in 10 minutes yesterday.
Sacks: Not on White House foreign policy team; past comments made headlines as "Trump advisor." Personal opinion: terrific ceasefire, great Islamabad meeting. Wars escalate out of control; credit to president for ceasefire and team to work it out.
Brad: Week six into week seven. March 4 tweet on Trump doctrine in Iran: massively destroy military capabilities, kill lethal weapons builders, get out, reserve right to repeat, zero democracy building—Iran builds what comes next. Markets: last year tariffs drew Nasdaq 15%, intraday 20%; Iran drawdown only 5-7% on S&P/Nasdaq—trust Trump to get out. Terrifies with tweets but resolves he'll exit. Markets bounced back pre-Iran levels. If JD lands plane, and Netanyahu commits to Israel-Lebanon government talks, markets off to races.
NYT inside-room piece: Trump followed Netanyahu into war February 11. Netanyahu's four-part pitch to attack Iran at White House. JD Vance warned of regional chaos, breaking MAGA 2.0 big tent (proven true: hand-wringing from Megyn Kelly, Tucker Carlson). Rubio anti-regime change, ambivalent on bombing. Susie Wiles concerned gas prices pre-midterms. Gen. Dan Cain: "sir, this is standard operating procedure for Israelis. They oversell, plans not well-developed. They need us, hard selling."
Rubio early war: "we knew Israeli action coming, knew it would precipitate attack on US forces, that's why we did it."
US foreign policy driven by Netanyahu? Jewish-Americans feel he's not helping diaspora. Chamath: President decides, not foreign leader. Markets price as small blip—best prediction market. End game likely. Israel concerned about losing US as steadfast ally—problematic for them more than US. Off-ramp makes economic/geopolitical sense; Israel wants to preserve relationship.
Americans don't want war, believe dragged in by Israel/Netanyahu too influential. Netanyahu's Gaza/Lebanon/Iran actions causing anti-Semitism.
Brad: Judged by outcomes. Optimist: verge of Gulf states transformation (Saudi, Qataris, Kuwaitis, Emiratis hopeful). Bring Iran into fold. Doing nothing in Iran/Venezuela had risks; calculated. Let cards play, history judges.
Sacks: In Gaza, wide berth for Israel/Netanyahu criticism. Naftali Bennett (former PM) tweeted polling: Israel unpopular in US, needs fixing—good sign. X auto-translates Hebrew.
Grok auto-translates best content from Japan, Israel, France, surfaces it. Replies auto-translated, enabling cross-language engagement—profound truth mechanism. Elon/X team deserves Nobel Peace Prize.
Chamath: English only in feed. Brad: sees Middle East, Chinese, Russian, Japanese. Base Japanese extreme (anti-immigrant, not Japanese culture).
Journalists not translating/covering; real-time velocity. Imagine Russia-Ukraine.
X better than ever, 70% fewer employees than Elon arrival day. Peak employment hit; do more with less. Nikita Bier: going ham on bought auto-reply accounts ruining feeds—subscriber replies mode works.
Sacks: 50/50 open/chaos vs. subscribers ($30K/year to charity, knows 2,000 of 1M followers).
Followers: Chamath 2M (engages haters, lights them up, funded kid's college after casino bet loss joke). Sacks alt cap 1.4M, Jason 1.1M.
Jacob Martin, would you fight the black women too much when you're more happy at how much it'll anger all these other goofball dorks living in their mom's basement? Yes, who literally have no take, they take no responsibility for their lives. They should enjoy those Hot Pockets. For those folks in their mom's basements, the Hot Pockets and the fish sticks are ready and you get one more hour of Xbox from mom.
We missed you Friedberg, but this is the best episode in two years. We will see you all at the Liquidity Summit except for the 400 people on the waitlist who aren't going to get in. We got an email from the guys at Athena because we were just oh my god, they're going to hire like 500 new Athena assistants. They had a thousand people after last week when we mentioned how much we love Athena. Go to athena.com. Those are like 500 hard-working men and women working in the Philippines, great jobs. Sax, I'm going to get you a couple Athena assistants as a birthday present. That's something to get, you're going to love this. Sax, oh Athena assistants are the best. Congratulations to my friends over there.
All right everybody, we'll see you next time. Love you boys. On Brad, see you tonight. Favorite podcast says we open source it to the fans and they've just gone crazy with it. Love you.
Nice queen of quinoa, these are gone. We should all just get a room and just have one big huge orgy because they're all just useless. It's like this sexual tension but they just need to release somehow. What you're about to be, what you're your feet. We need to get merch. Are you?
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
