Some thoughts on Dario’s post:
1. Dario does not actually address Gavin Baker’s account of what he said – something he could easily deny if it were inaccurate.
2. Dario claims his critics live in a “bubble” where all regulation equals regulatory capture. He calls this an overly simplified view and notes that “Many people outside this bubble think of regulation as something that constrains corporate power and benefits ordinary people.” This argument is a straw man. Of course treating all regulation as capture would be overly simplified – but almost no one holds that view. I have repeatedly argued for strong antitrust enforcement to keep industries competitive, especially Big Tech. If Anthropic continues toward monopoly or duopoly status, I would be among the first to demand those rules apply.
3. Regulatory capture is not vague or in the eye of the beholder. Nobel laureate George Stigler defined it as regulation acquired by an industry and designed and operated primarily for its benefit. Stigler challenged the traditional view that government regulation arises from a benevolent state protecting the public from market failures. Rather, industry groups have concentrated stakes and pour resources into influencing regulators, whereas the public’s stake is diffuse and unorganized. The revolving door between companies and the agencies that regulate them compounds the problem. Anthropic understands these dynamics: it has hired multiple senior Biden AI-policy officials and built a substantial government-affairs operation plus a network of aligned organizations to push its preferred frameworks at state and federal levels.
4. Dario has consistently pushed for a new federal agency to review and approve frontier models prior to release – a proposal framed variously as an “FDA for AI,” an “FAA for AI,” and most recently a “FINRA for AI.” I call it a “DMV for AI” because a review process modeled on the FAA or FDA (which takes years) or FINRA (which issues rules for a staid industry widely seen as protecting incumbents) will create long queues as AI models wait for testing and approval. This process will only become more labyrinthine as rules accumulate to prevent theoretical harms. This would handicap the U.S. relative to China, which will not adopt the same constraints. It would also undermine Anthropic’s own business model, whose pricing power depends on remaining ahead of open models. Whatever Dario states today, it is difficult to believe the company would simply accept outcomes that erase that advantage.
5. Anthropic is on track to become one of the most valuable companies in history, with the resources to navigate any approval process and shape the rules while competitors wait. Dario wants open models under heavier scrutiny – he has called them dangerous in Senate testimony, criticized them for not being centrally monitored or withdrawn, and linked them to IP theft. He says he has never sought a ban, but he could achieve a similar result by insisting that identical rules apply to both open and closed models. The U.S. risks becoming an island of costly closed models while the rest of the world races ahead with broader choice.
6. Dario acknowledges that AI is structurally centralizing but attributes this mainly to chips and scaling laws. Access to compute matters, but the deeper risk is who decides which capabilities are available to whom. His preferred pre-deployment testing and FAA/FINRA-style oversight would place that gatekeeping power in a federal bureaucracy working hand-in-glove with a small number of frontier labs – reinforcing centralization rather than countering it.
7. The second part of Dario’s post assumes we have amnesia about Anthropic’s well-orchestrated campaigns hyping AI fears. His May 2025 claim that AI would wipe out 50 percent of entry-level knowledge jobs within five years still lacks supporting evidence fifteen months later. Similarly Anthropic breathlessly promoted its heavily contrived “blackmail” study on 60 Minutes. Yet Dario blames public negativity on a long-standing loss of trust in institutions rather than his own messaging.
8. These narratives have done more than anything to shape public fear. People are left asking the same question Mark Zuckerberg posed: why race to build a future you describe in such negative terms? Thomas Sowell’s "The Vision of the Anointed" captures the mindset – elite intellectuals convinced that only they are enlightened enough to control the outcome. As Zuckerberg notes, concentrating power in the hands of an enlightened few has rarely produced the promised results; the practitioners turn out to be less enlightened in practice than in self-conception.
9. Gavin Baker summarized the disagreement cleanly on our pod: Dario believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize. Dario appears to believe, sincerely, that safety and progress are best served by centralizing authority in a marriage of corporate and state power. The weight of human history gives us reason to fear that outcome.
Dario Amodei: 1/2 Thanks Gavin for an especially thoughtful exchange. I don't usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important conversation.
First, on regulation, I think that “either concentrate it in the hands of a

Completely nuts that PBM revenue is comparable to pharma revenue in the status quo.
Pharma companies take on enormous scientific and execution risk to discover and advance medicines. PBMs are unnecessary middlemen that do neither of those things.
Mark Cuban: Tweet 4: Legacy Model vs. Open Net Model
The macro shift:
• Gross Invoiced Spend: $750B ➔ $487.8B (−$262.2B / −35%)
• PBM Retained Spread & Fees: $360B ➔ $0 (Eliminated)
• Pharma Revenue: $390B ➔ $390B (Stays whole)
• Wholesaler Margin: $15B ➔ $23.4B (+$8.4B)
•

People keep asking: do Flock cameras really help fight crime?
So I built a database of over 700 times Flock license plate readers have helped solve a crime, apprehend an offender, or recover a stolen car or abducted child.
The full database is here: https://www.flockstopscrime.com
Charles Fain Lehman: Fun little Flock-related project dropping tomorrow, watch this space 👀

Also in Flock news: new working paper shows Flock reduces auto theft, boosts clearance rates.
@smourtgos and @ian_t_adams merged data on 216 Flock deployments with PD crime counts.
They find an 11 percent reduction in motor vehicle thefts, and a 16 percent increase in motor vehicle clearances. Among recorded recoveries, cars were recovered about 0.28 days faster.
Full paper here: https://www.crimrxiv.com/pub/zleg04q3/release/1
Charles Fain Lehman: People keep asking: do Flock cameras really help fight crime?
So I built a database of over 700 times Flock license plate readers have helped solve a crime, apprehend an offender, or recover a stolen car or abducted child.
The full database is here: https://www.flockstopscrime.com

My 6 biggest takeaways from @rileybrown on how he uses Codex to run his 1.5M+ follower content business:
1. Remix proven thumbnails with Codex and Paper
Riley asks Codex to collect thumbnail formats from top YouTube channels in his niche and place them in Paper. He then uses Paper’s image-generation tools to mix his face with proven formats, giving him several directions to compare before choosing what to go with.
2. Riley's AI stack for end to end video creation
Riley uses Codex and Supadata to research topics and pull transcripts from top-performing videos. He dictates ideas with @wisprflow, turns them into diagrams in @excalidraw, pulls logos with SerpAPI, and creates animations and B-roll with @Remotion.
3. Reverse-engineer a winning YouTube hook, then remix it
Riley’s research AI skill finds and pulls written hooks from successful videos in his niche. He then works with Codex to adapt the structure to the video he wants to make. He scripts the hook because it matters most, then talks naturally for the rest.
4. Make the video you want first, then package it
Riley starts with what he wants to learn or explain and saves the title, thumbnail, and hook for the final 10% of the process. “I would rather make what I want and then package it after than to only chase viral ideas.” That runs against the usual YouTube advice to do the packaging first.
5. “I’ve never looked at an AI skill file once.”
Riley works with Codex until a workflow produces the result he wants, then asks it to turn that workflow into a skill. He improves the skill by judging its output instead of reading the file.
6. Quality matters more than efficiency
Riley resists using AI to create many videos in batches because every video still needs a strong human element. AI helps with research, hooks, diagrams, animation, and thumbnails, but Riley stays in the driver’s seat and spends time to make each video better.
Overall, it was fascinating to see Riley's content creation workflow and compare it with my own.
📌 Watch the full episode now: https://youtu.be/N34zz1-RSGw
Peter Yang: “The moat is quality over a long period of time."
Here's my new episode with @rileybrown, where he showed me how he uses Codex to run his entire content business (1.5M+ followers), including how to:
→ Create thumbnails, animations, and B-roll
→ Research and craft winning

Last week, hundreds of founders and builders gathered for AGENT NIGHT.
If you weren't there, here's what you missed including the preview launch of WorkOS Airlock:
01:25 Why agents need a new access control system
18:00 Introducing Airlock: Intent-Based Access Control
21:11 WorkOS Airlock live demo
34:06 The Age of Personalized Software - @davidcrawshaw (@ssh_exe_dev)
41:06 Get your product into agent retrieval and model training data - @inazarova (@evilmartians)
47:26 Make moves with Mastra Factory - @abhiaiyer (@mastra)
51:31 How to be the very best Self Healing Pokémon Master - @bdougieYO (@papercompute)
57:33 Panel: The State of Agents with @JayaGup10, @Altimor, and @swyx

Progressive policy took @UCBerkeley from Manhattan Project to “can’t do algebra” in two generations.
Steve Guest: Cal Berkeley is cooked: I teach calculus at Berkeley. Some of my students can’t do middle school math
“Since the University of California abandoned the SAT and ACT, thousands of faculty say it’s admitting students who aren’t prepared for the coursework.”
“I teach mathematics at

New enterprise governance tools are coming to Replit: comprehensive audit logs, more control over enterprise workplace settings, and admin API to manage accounts at scale. https://replit.com/blog/new-enterprise-governance-tools

It's time to rethink RL.
Translating real world use into model improvements requires redesigning post-training algorithms for non-verifiable, per token rewards.
At @aiDotEngineer 's World Fair, we share our insights into scaling algorithms like SDPO for continual learning.

“I’ve never looked at a skill file once.”
From @rileybrown:
“Ask AI to do a thing using a skill. If it doesn’t do a good job, tell it to change the skill so it doesn’t make that mistake again. Then test the agent in a new chat with clean context.”
“Hold your AI to a standard, then improve the skill by measuring the outcomes.”
📌 Watch the full episode here: https://youtu.be/N34zz1-RSGw
Peter Yang: “The moat is quality over a long period of time."
Here's my new episode with @rileybrown, where he showed me how he uses Codex to run his entire content business (1.5M+ followers), including how to:
→ Create thumbnails, animations, and B-roll
→ Research and craft winning

Made a website where you control a worm just by scrolling 🐛
No code. No canvas library. Just prompts.
All built in @Replit Design.
Replit ⠕: Introducing Replit Design.
The next era of design, for everyone.

🆕 Context Engineering in 2026: Compaction, Memory & Cost
https://www.youtube.com/watch?v=WP3hjUXd918
@Whats_AI, @samridhivaid and @omar_solano1 return!
This workshop is about engineering the context window so rot stops happening, shown with @towards_AI's open-source AI tutor, which answers questions for students of our AI-engineering courses.
Context engineering is deciding what the model sees on every single call — instructions, history, retrieved course content, memory, and tool outputs — and it's the line between a tutor that holds a coherent session and one that forgets the student's setup halfway through.
We'll move in three stages, mirroring how the project actually went. The concepts:
- the two root problems (a finite window, a stateless model),
- the full compaction toolkit (truncation, trimming, tool-result clearing, summarization, and offloading to files — and when each actually helps),
- memory that survives across sessions, skills loaded on demand, and
- production-grade retrieval (chunking, metadata, course scoping, hybrid search, reranking, and evaluating).
We'll cover the tutor's architecture, and the evaluation harness we used to measure every run on Gemini — tokens, cost, latency, and memory probes instead of vibe-checks. At real volume, even Gemini Flash got expensive, so we tested whether open and local models could match the quality for a fraction of the cost and match result quality.
Everything is open-source and will be shared during the workshop.

I think Grok @Bot is a glimpse into the future of personal AI agents.
Here's my new tutorial where I show you how to set up 5 useful bots:
1. An advisor to create and manage your bots
2. A YouTube researcher to find outlier videos
3. An X scout to find viral and funny tweets
4. A digital Marie Kondo to clean up your inbox and save money on paid subscriptions
5. A personal concierge to save money on trips
I also tested a Gamer bot to see if Grok Bot can install and play classic games like Doom, Red Alert, and Commander Keen.
Plus, I discuss the biggest barrier to Grok Bot adoption and whether it can replace ChatGPT as my daily driver.
📌 Watch now: https://youtu.be/MkVcHbviYOw

You can now run black-box pen tests for your Replit apps.
Security scans that test your Replit apps the way external attackers do.
Replit Agent can fix what it finds in a single click.

Last week, Exa took over SF's Pioneer Building, previously home to Stripe, OpenAI, and most recently xAI.
Elon assembled the Colossus cluster here (his mattress was still lying there during our tour).
Alec Radford trained GPT-3 here.
The Collison brothers built the web's financial backbone here.
They were all infra companies:
Stripe -> Financial infra
OpenAI -> Intelligence infra
xAI -> Compute infra
Each entered Pioneer as a small startup with enormous ambitions and left as a foundational layer for the world.
Exa -> Information infra
We entered Pioneer as web search for agents, but by the time we leave we'll have organized all the world's data into a foundational information layer.
This will require things like deploying satellites, new misinformation detection algorithms, and embedding retrieval on an outrageous scale. Whatever is necessary to keep billions of people and trillions of AIs as informed as possible.
A trusted information ecosystem is necessary to safely navigate this chaotic AI era, an era that was set in motion by our previous occupants.

The original lore was when @realGeorgeHotz dropped on an early @latentspacepod pod with @swyx that GPT-4 was a 220B x 8 MoE model.
June 2023

The River API was tested in this blog post and outperformed Tinker on reinforcement learning runs with identical training code. We spent a lot of effort to get details like routing replay right so you get the best possible results with the API.
Ashwinee Panda: We can now RL large MoEs with 0 train-infer mismatch! And doing so can improve performance (pictured task: teach Qwen3.6-35B-A3B to play Wordle). Everything is open-source and we did a bunch of ablations. 🧵

It’s not enough to scan your code for vulnerabilities; it’s important to try to break them with pen testing.
Replit ⠕: You can now run black-box pen tests for your Replit apps.
Security scans that test your Replit apps the way external attackers do.
Replit Agent can fix what it finds in a single click.

Sometimes I wonder why random people with no connection to SF hate it so much.
Other times I know exactly why.
Imagine a guy pledging to invest $100 million into a forgotten retail strip - in the neighborhood he grew up in - through a privately-funded nonprofit and being accused of a “hostile takeover” by arguably the most influential local politician.
Now imagine the guy wants to do it in literally any city besides San Francisco. They’d throw him a parade - and rightly so.
You hear a lot that San Francisco hasn’t lived up to its potential because of anti-growth politics.
I disagree.
It’s a miracle San Francisco isn’t worse off given the entrenched anti-growth politics of the past 50 years, with periodic respites like the present moment.
Tyler | Kenji Capital: Remember the tech billionaire buying up a huge chunk of Fillmore Street in San Francisco?
Well, Neil Mehta and his partners just acquired ANOTHER property over there for ~$8.6M.
It's part of an interesting experiment:
Can one investor help revitalize an entire neighborhood

Dial website UI updated to match the iOS version in less than an hour 🙌🏽 thanks @Replit
https://dialmoments.com/
Elias Reyes: The iOS app UI vs the current web UI for https://dialmoments.com/ 😅 time to update the web one. I love the mobile UI version.

We didn't get into YC until we stopped trying to get in.
After the 4th rejection for InsForge, Tony and I decided to generationally lock in and focus completely on building product and acquiring users.
We looked at how the best devtool companies (Supabase, Resend, WorkOS) got developers to care. We realized we should run our own Launch Week!
After 30 days of insane heads down building, we did it.
InsForge Launch Week 1 went completely viral. We hit #1 on Product Hunt, #1 on GitHub Trending, 1M+ views on X and LinkedIn.
We generated enough motion that YC couldn’t ignore us. They ended up reaching out to us instead of us applying. That's what got us into YC.
Paul Graham: If you want to get into YC, don't focus on getting into YC. Focus on building stuff and understanding your users. That's what gets you into YC.

My comparison of Grok Bot vs. Hermes vs. ChatGPT Work:
Hermes:
- Pros: Open source and very customizable
- Cons: Requires DIY setup on Mac Mini or virtual private server
ChatGPT Work
- Pros: Browser use and voice features are best-in-class, great for power users
- Cons: UX is confusing across Chat, Work, and Codex. Browser cannot auth to your favorite apps (yet)
Grok Bot
- Pros: Has a built in persistent cloud computer. UX is simple, focused, and delightful
- Cons: Less flexible than the other options (I typically want multiple threads with the same bot or workflow). Starts at $200/mo (you can get decent usage out of ChatGPT Work for $20/mo)
Overall, I'm still using ChatGPT as my daily driver but my bet is that OpenAI, Anthropic, and others will soon follow Grok Bot's steps with the persistent cloud computer.
I also think that Grok Bot will improve rapidly - the team knows how to ship fast and iterate.
📌 Watch my full tutorial here: https://youtu.be/MkVcHbviYOw
Peter Yang: I think Grok @Bot is a glimpse into the future of personal AI agents.
Here's my new tutorial where I show you how to set up 5 useful bots:
1. An advisor to create and manage your bots
2. A YouTube researcher to find outlier videos
3. An X scout to find viral and funny tweets
4.

one fun thing about my last three jobs (FigJam, Cowork, now Grokbot) has been having @petergyang as a critic and collaborator throughout.
Peter Yang: I think Grok @Bot is a glimpse into the future of personal AI agents.
Here's my new tutorial where I show you how to set up 5 useful bots:
1. An advisor to create and manage your bots
2. A YouTube researcher to find outlier videos
3. An X scout to find viral and funny tweets
4.

95% of Mission high school students did not meet CA math standards, yet they are taking top spots at all the UCs ie Berkeley, UCLA, UCSD, etc., while Lowell students (71% met math standards) are only competitive w/ Mission students at UCSB, Merced, Riverside or Santa Cruz. UCs punish Lowell students for working hard bc of their race, while setting up Mission kids up for failure & depression. Race-based admissions fails all kids & was deemed unconstitutional by Scotus - it must end.
So where do the students actually end up going... 🧵