Great to chat with @AndrewYNg at the Agentic AI Conference at Berkeley. Our full talk is now available on YouTube. We discuss if we should expect an AI jobpocalypse, the importance of open models, and whether AGI is still 50 years out or happened 30 years ago.

Great insights from my partner @DavidCahn6 on the latest @BigTechPod with @Kantrowitz. Worth a listen if you care about the state of AI and how these big bets will pay off. A few good ones: On Jensen: "He was one of the first people to make a bet on AI. The way that everybody else is fighting over share and this and that, I think Jensen just wants the pie to be really big. His fundamental world model is you win, I win." On resource allocation: The winners will be decided by what resources each player has (cash, talent, chips, distribution) and how coherently they allocate them. Founder-led companies play coherent games; committees don't. On market reactions: The lab leaders have told us what they believe and they're all playing for AGI. Back-test their decisions against that world model and the moves make sense. Markets get confused because they underprice both AGI and a correction while overpricing the status quo.

In re-sharing this post, I went down a rabbit hole on the Empire State Building. It was built in 1930 in 410 days and landed under budget, despite having none of our modern-day advantages or technologies. As a corollary, the World Trade Center broke ground in 1966 and took 7 years to build. The ESB is ~1,250ft to roofline, the WTC was ~1,370ft to roofline. The ESB did it by subordinating every design decision to speed: - No exotic materials or systems were used; just known steel, floor, and window systems - Window placement, stone thickness, and cladding attachment were all chosen to minimize on-site cutting and hand-fitting - Engineers explicitly designed systems so trades could work independently and in parallel without waiting on each other, reducing the risk of delays - Demolition started before design was finished, foundations were poured while upper floors were still being designed, steel was ordered a month ahead of need - The owner, architect, engineer, and contractor sat together through construction, resolving details jointly instead of each one handing off their piece independently - On-site narrow-gauge railways, dedicated hoists for brick and stone, and on-floor cafeterias removed friction from moving material and people Interestingly, the profit-maximizing design was determined to be 63 stories, but they built it to 85 stories for prestige and to beat Chrysler. Speed discipline saved the vanity height economics. For contrast, the WTC took over a decade because it was plagued by lawsuits, political fights, and novel/untested systems, the reverse of the ESB's playbook. Lots of learnings for startups, both inside the building (focus, parallelizing vs serializing, etc) and outside the building (regulatory, etc).

Kareem is a very thoughtful CEO and it's been a pleasure to work with him, Varun, and the Clay team. Great thoughts in here on his role as a "momentum detective" hopping into areas where energy is waning, Clay's just-in-time decision-making and preference for open communication, and their commitment to radical transparency with both their employees and customers.

"Underneath the statute is a belief, the one that strung the wire and funded the labs and put men on the moon: that building hard things at scale is worth doing, and that this is the country that does them. The American method was never the state alone or the market alone. The wire reached the farms because federal credit met local cooperatives that did the stringing. Apollo was a government program executed by four hundred thousand people who mostly worked for contractors. Warp Speed was public money and private molecules. Public purpose set the pace and wrote the check. Private ingenuity built the thing. Deep capital markets funded the improbable. A bankruptcy code cleared failure fast. Immigration imported a century of talent. The answer to a rival that builds by command isn’t to become one. It’s to run the harness again, by choice this time."

Our intern just built the first zero-person company. Listen's agent ran a loop: - Interview users - Build - Test with real people - Fix issues - Repeat 2,000 interviews and 100 concepts later: an app with 100s of paying customers. Here’s how it works:

There is a reason that advice is free. It worked for the person asked, but that advice is unlikely to work for us. Our situation is different. The world has moved forward. New solutions developed, and new challenges are present. What has worked in the past is unlikely to work precisely in the future.

Q2 recap for @harvey - +$100M NNARR - 53% DAU/MAU Key hires (including Q1) - Anique (CPO) - prev VP of Product at Rippling - Rachel (CMO) - prev CMO at Notion - Brooks (CISO) - prev CISO at Roblox - Keith (CSO) - prev CPO at Google Product - Agent unification - cloud agents can use all Harvey product surfaces - Command center (EA) - monitor adoption and ROI by use case - Contract intelligence (EA) - agentic contracting platform for enterprises Eng - Migration to cloud agent infrastructure - Integrating open source inference providers - Scaling document processing (54TB / week) AI - Legal Agent Bench - Open source post training - Published multiple research directions with partners We invested heavily in cloud agent infrastructure at the end of last year and in Q1. In Q2 we also unified many of our product surfaces (collapsed as @winstonweinberg says) by making them all tools accessible by our cloud agents. Prior to this, there were a lot of capabilities in Harvey that were often only discovered by power users. As cloud agents get better and our product becomes more connected we are seeing users discover more of the product by learning from their agents (see plot of product surfaces per user).

What we learned from the DeepSeek R1 moment, that also applies to GLM-5.2, and will apply to others in the future: Limits push you to get creative - Not having enough, whether it's chips, money, or time, forces clever solutions that you'd never find when you have plenty. So treat limits as a reason to invent, not a problem to fix. When you're stuck, try the opposite of what feels natural: cut the budget, shorten the deadline, or raise the bar instead of adding more. The big surprises are usually predictable - When a model is improving fast, that speed is the clue. Strong results aren't a shock; they're just where the trend was already heading. People come around to new ideas slowly, so if you pay attention and form a view early, you can get ahead before everyone else catches on. Don't write off huge leaps as cheating or fake - When a competitor jumps way ahead, the easy reaction is to assume they cut corners or made up the numbers. Maybe some of that is true, but there's almost always real innovation behind it. Better to assume they genuinely beat you at something and figure out what you can learn.

Very excited to see this new effort from Stripe, Visa, Coinbase, Mastercard, Amex, Blackrock, and many others to build a new open stablecoin that shares economics back to users and distributors. OpenUSD will be natively issued on Tempo on day 1!

Palmer gets it: If you're going to say something everyone agrees with, you might as well have said nothing at all. You're not going to build a following of people who say, "I just love his right-down-the-middle, very hedged takes that everyone agrees with." If some people love what you're saying and some people hate what you're saying, that's a lot better than having everybody lukewarm agree with you. Don't waste time communicating about what everyone already agrees on. Focus on the things where you need to change their mind.

the destruction of American education over the past decade is an incredible self-own competence is objective. a child can either do the math or they can’t. but in the u.s., a lot of people have reasons not to say that! parents don’t want to hear their kid is struggling. teachers don’t want scores used to manage them. districts don’t want embarrassment. progressives worry accountability will create inequity. conservatives don’t want federal authorities. so we end up with process, weak standards, and excuses to explain away bad outcomes. people object that it’s phones, covid, demographic change. ok! but we fail globally when others have phones, covid too — vietnam is much poorer than the u.s., yet performs well in international math comparisons. some countries treat math as a basic skill everyone needs to master. here, it is part of a fight about fairness, autonomy, and feelings in the age of AI — if people can’t do basic math, read closely, or think through problems, ai won’t make them more capable. it will become something they rely on without understanding. the countries that come out ahead in the global race won’t just have better technology. they’ll have people who know how to use it, question it, build on it. we need the national ability to decide something is worth doing coherently (teaching math!) the US has the money to teach math well but it has not shown the will. we are failing the next generation

How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching. Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting to open weight models like GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task. 91% of our employees were never hitting their usage caps, so instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models, so they can check each other's work. Better Routing – In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing. For instance, you may want a frontier model for planning, but not for execution where they can be overkill. Ultimately, humans shouldn't be choosing models - AI can automate this task. Better Caching – Cache misses are the easiest way to drive your cost up. All of our requests are cache aware, so we’re reusing a warm cache wherever possible. For example, our cache hit rate went from 5% → 60% in LibreChat once properly implemented. Keep Context Lean – Start fresh sessions when switching tasks. Scope file context narrowly. Disconnect unused tools. Don't just compact. The goal isn't fewer tokens used, it's fewer tokens wasted. Better Visibility – Our engineers can use as many tokens as they want, from whatever model they want, but we’ve made usage visible – and the more you spend on AI, the more impact we expect. The goal isn't to suppress usage. It's to build the infrastructure that makes exponential growth sustainable. Putting this into practice has cut our AI spend nearly in half, while our token usage continues to grow.