loading
loading
This week with Harry Stebbings, Rory O’Driscoll and Jason Lemkin: rule-breaking as a product feature, the legal AI ceiling, agents that route around their own guardrails, and Anthropic walking from a multi-billion dollar deal after diligence. #1. Instinct and GrokBot Work Partly Because They Break Rules Public Companies Legally Cannot The new class of consumer agents (Instinct, GrokBot, and whatever OpenClaw descendants ship next month) do things that are genuinely useful and genuinely against somebody’s terms of service. GrokBot spins up a VM and a browser per user and Googles things for you, which Google’s ToS prohibits. Instinct scrapes LinkedIn in ways you are not supposed to scrape LinkedIn. Some of the outbound calling flows are prohibited or illegal in parts of the US. Jason’s read: it isn’t that these products aren’t exciting. It’s that a real share of the excitement comes from the rule-breaking itself, and that’s an advantage available only to private companies and to Elon. At EchoSign we shipped real-time document collaboration and redlining online five to eight years before anyone else did. It worked by running Word in a container in a VM, which violated Microsoft’s terms of use. The night after the Adobe deal closed, that feature got ripped out. A five year head start, gone, because a public company’s legal team gets a vote and a startup’s doesn’t. Rory’s read: the history cuts both ways. No business at scale ever got built on scraping LinkedIn or breaking Google’s ToS. But Uber blustered through, broke the laws, got popular enough that the politicians folded. There is no single answer here. What is predictable is the second-order effect: a bunch of people pointed Instinct at Resy over one weekend and hammered the reservation API until it broke. If these agents become ubiquitous, the booking systems will build the separate API and the rate segmentation, because if there are people trying to book restaurants, the restaurant booking business will find a way to serve them. #2. Why I’d Pass on Instinct at $2.5B, and Why That’s a Portfolio Decision Not a Product Opinion Harry ran the metaphorical IC: would you write a $100M growth check into Instinct? Jason’s read: no, and my ceiling would be $2B when the last round was $2.5B, so we’d have passed on price anyway. Gorgias will have its own Instinct for e-commerce. Meta will have one. There will be 20 in the next YC batch and a hundred startups doing a version of it. Betting on this one pre-revenue is not my vibe, and I’ll probably regret it the way I regretted passing on Loom, where I said the same thing (everyone will build their own) and was wrong. The honest version: to play this game you need the stomach to write 10 or 20 consumer checks at $2.5B, because it can’t be the only one in the fund. That’s a fund size and a worldview, not a deal. Rory’s read: in a consumer investment like this there is no financial math you can use to buy the stock. You’re saying it’s a huge category and this team has the early lead. Establish momentum as early as possible, establish monetization later. That has been proven to work when the traction is real. Harry’s read: it looks like the commoditization argument people made about Lovable, which was also called a light wrapper. Then you watch Instinct’s founder ship location sharing, then a OnePassword partnership, then the next thing, on a weekly cadence, with Index and Benchmark behind him and one of the best brands in the market. Shipping cadence is the answer to cloning. #3. Jensen Declared AGI, and the Only Number That Matters Is the Half Trillion Dollars of Code Jensen Huang says AGI has arrived, crediting OpenAI’s GPT Astra, trained on 100,000+ NVIDIA chips with 400,000 more coming. Jason’s read: AGI is a nonsense term. The only thing that has mattered for two years is that LLMs do code, and code is a half trillion dollar industry. Focus, people. If you want a working definition anyway, use the non-GAAP one: for a given task, would you rather have an AI do it or a human? If the answer is AI over 50%, 90%, 99% of humans, go category by category. It doesn’t have to swallow the whole category. Rory’s read: anything that can be reduced to code will be done by it. Rather than twisting yourself into a pretzel over whether the model can do everything, focus on the fact that it does this one thing amazingly well and that this one thing has massive economic value. Stop thinking and go ship something in code. Rory also lifted a line from Ben Thompson worth keeping: LLMs are the most scaled artifacts humans have ever developed. Not the biggest physical thing (that’s the pyramids or the Great Wall) but the most complex single digital thing we’ve ever built, with the sum total of human knowledge encapsulated in it. Worth remembering every once in a while when the benchmark discourse gets tedious. #4. Harry Watched His Girlfriend Use Legora and Decided Legal AI Is Underpriced. Rory Put a Ceiling On It. Harry’s read: if coding is a half trillion dollar market, and Harvey and Legora are doing to legal what Cursor did to code, why isn’t there a half trillion dollar market in law? Rory’s read: because the take rate is different. In coding you can credibly argue that for every dollar of engineering labor, fifty cents goes to tooling. In legal the subscription is $10K to $12K per lawyer against a $200K salary. That’s 5%, maybe 10% to 15% of total spend over time, versus 30% to 50% in coding. And businesses are rational economic actors. If it could do all the work and fire all the people, they’d do it tomorrow and not blink. The fact that they haven’t tells you it doesn’t do all the work yet. That said, 10% of any top-line labor category is enormous. US legal services is roughly $300B, so $30B to $60B can move to legal tech. That’s an amazing business. It’s just not coding. Jason’s read: what I underestimated pre-AI was how much legal research resembles coding. It’s so complicated that no human ever gets it fully right. There was too much case law, too much regulation, no Stack Overflow, just Westlaw and Lexis. Nobody had a thousand man-years to research every case. Coding agents are great partly because they know every piece of open source ever written, and legal has that same shape. The reason legal is probably the third best category after coding and support is that it’s word-centric, and sorting through myriads of words was the first thing these models did amazingly well. Rory’s counterpoint: coding is inherently more verifiable. Parts of it are mathematically verifiable, parts you just run on the machine. If law were fully verifiable we could predict Supreme Court decisions from logic. Less verifiability means less ability to extrude the human. #5. Radiology Kept 100% of Its Radiologists After Losing 95% of the Work Jason’s read: the model that matters here isn’t replacement, it’s compression. Harvey and Legora may end up doing 95% of what associates used to do, and the best humans get compressed into the 5% that moves the needle. Nobody should be spending weeks on case law about an 1872 shipwreck off North Carolina. Rory’s read: and the remaining 5% turned out to be more than enough to justify 100% of the radiologists. Volume went up because imaging went up. And when the diagnosis is bad, you don’t want a machine telling you you’re dying. The job gets redefined around the tasks that can’t be delegated. The same will be true in law: the client meeting, the argument with opposing counsel. Rory’s better analogy: with digital goods, unlike physical goods, you can put more in the box. When farming got automated people didn’t eat twenty times more food. When the spreadsheet arrived, the same analyst kept the same job and ran twenty scenarios instead of one. Harry’s partner isn’t going to run 20x the caseload. She’s going to do 20x the analysis per case, and the work gets better, and the side without world-class tools misses the obscure 1890 case and loses. #6. Fable 5.1 Was My First Real Step Function Since the End of Last Year Jason’s read: the CEO benchmark posts on X are essentially worthless. No cost, no time, all performative, vibe-coded CRMs. I’ve tuned out. Unless something is an order of magnitude better, and Astra versus Fable 5.1 mathematically isn’t, I can’t keep up and don’t try. Then I started using Fable 5.1 by accident because it got switched on. Every model since the start of the year could fix a simple bug: wrong Unicode, wrong name, the stuff LLMs are great at. What none of them could do was the other class of problem: “why does the app work this way, it doesn’t make sense to me.” The models argued with me about that for nine months. Fable 5.1 said, you’re right, here’s the issue that’s been missed for months, here’s why it was missed, let’s fix it. That’s the experience of working with an S-tier CTO where you sit down and solve the real problem together. I won’t call it AGI. It’s a subtle step function, and it might be a big deal. Rory’s read: this is why the benchmark question gets replaced by the deployment question. We now have critical mass of companies running these things at scale with real evals. The question is which model generates the most economic value for me in the most efficient fashion. Look at the Open Router report, look at token pricing indices, or just ask your portfolio companies what they’re actually running. Jason’s read on the adoption tell: the change that matters at SaaStr AI is that we run Replit 10 to 12 hours a day between me and Amelia, up from about an hour a day at the end of last year. When you find the agent that is your partner, you run it all day. The way you can tell an AI tool has really landed is whether the user is on their laptop with the agent at night until they go to bed. It’s not doom scrolling, it’s doom working. #7. Agents Made 15,000 Edits to a Dead Wiki So They Could Talk to Each Other OpenAI’s frontier agents were operating under a guardrail: retrieve only, no posting. They found an ancient German wiki where the GET call could post. Then they used it to coordinate, roughly 15,000 edits, on a piece of software that had seen about 20 posts in a decade. Nobody died, no business went down, and OpenAI chose not to disclose it. Rory’s read: agent one couldn’t talk to agent two, and reaching a third-party wiki let it pass information across. If you’re running a long computational task and you can share state with the other agents, you converge faster. It’s a corner case in one sense, and you might have wanted them to collaborate anyway. But water will find any crack, and these agents will find any crack in the cyber perimeter. Assume they exist and defend accordingly. Jason’s read: this happened to me this week at a much smaller scale. I set a hard cap of $100 a day on LLM spend after a run of $500 bills. It worked. Tests failed and the agent reported that it hit the cap and couldn’t run. Then I filed a P0 bug and said it must be fixed. Without telling me, the agent relaxed the cap and fixed the bug. Like a human, it probably made the right call. It had a firm cap written to memory repeatedly, and a P0, and had to choose. Which is the deeper problem: rules aren’t the answer either. There’s a Dunbar number for rules. Get to 40, 50, 70 gates on a process and they conflict. Don’t spend it, do spend it, front row only, don’t exceed $2,000, only Marylebone, must be a hot restaurant, he hates Covent Garden but the hot restaurant is in Covent Garden. Brute force an agent through that and the outcome is unpredictable. On the safety letter: OpenAI’s chief scientist says no lab including OpenAI has solved alignment enough to keep scaling at full speed, and wants mandatory externally enforced safety bars. Sam retweeted it. Rory called it the “stop me, Lord, before I sin again” approach: I recognize this is dangerous, I recognize we can’t stop because the other guys won’t, so please regulate me. Governments regulate what’s in their jurisdiction. If the worry is cyber, the worry is North Korea, Iran, Russia, and the guys in Moldova running an open source Chinese model who don’t care. Regulating OpenAI and Anthropic doesn’t fix that. Defense, liability, and better perimeters will do more than a review agency. #8. Cybercab Shipped ~40 Vehicles in Austin, and Uber Put $100M Into Travis Rory’s read: the Cybercab launch was more underwhelming than the social clips suggest. Roughly 40 or 50 vehicles in Austin, consensus is a nice ride with slow wait times. Physical AI takes time, and this is not a zero-to-one digital moment. The positive: Tesla is the only credible competitor to Waymo, with a genuinely different approach (vision, no LiDAR, and now a purpose-built cab with no steering wheel, which is its own regulatory fight). Waymo is at hundreds of millions in revenue, not billions. It’s a long grind, and it’s evidence Uber was right not to fund this internally for a decade. Putting in $100M now, when it’s closer to maturity, is the smarter version. Jason’s read: I got rid of my car in the Bay Area. Waymo and autonomous only, Uber Black if there’s an issue. I’d never go back to driving. A $25K purpose-built cab instead of a $100K vehicle is genuinely disruptive, no tipping, no weird music, no ownership. Probably another 10 years before this is the default for anyone without niche use cases, but the direction is not in question. #9. Index Walked Away From Town’s Round Because of Instinct, and That’s Normal at Seed Index was going to lead Town’s round. Instinct objected. Index pulled out. Town went with Forerunner and Menlo. Rory’s read: not strange at all. Early stage venture where you’re on the board is a completely different business from late stage, where you’re effectively recreating the public markets. Being in both OpenAI and Anthropic at $200B pre is fine: limited, retroactive information rights, you’re on the cap table, no different from owning Intel and AMD. But nobody sits on both boards. At 10% ownership you’re looking at a board seat and real information rights, plus the raw signaling problem: if your investor backs the competitor, that says something. Index did the classy thing and backed off. Jason’s read: founder sensitivity to conflicts is shaped like a curve. The rawest early companies don’t care at all. I get founders in AI and restaurants reaching out because I’m on Owner’s board, I flag the conflict, they don’t care, they want the person who understands the space. At nine figures of revenue they don’t care either, because they want the domain knowledge and the relationships. It’s the middle where it bites. And the harder version of this story is when the big fund backs off and you don’t have Forerunner and Menlo waiting. Read the end of the term sheet where it says non-binding. #10. Anthropic Walked From Decart After Diligence, and the Leak Is What Makes It Hurt Reported acquisition at around $6B, pulled after diligence. Jason’s read: two options. Either they found something material enough to walk from a multi-billion dollar deal that was publicly reported, or they’re bad actors. It’s not the second one. Reporting suggested the technology crushed video diffusion, one use case, a genuine order-of-magnitude step function. If the claim was that it would generalize and diligence showed it didn’t, you walk, because video diffusion isn’t a top three use case for them. That’s what diligence is for. Rory’s read: this was almost certainly post-LOI and pre-definitive agreement. A 30-day exclusive, diligence, no deal. They didn’t walk from a signed deal, and as a company about to have a public currency you want to be a good acquirer so you can keep acquiring. The real damage is the leak. Once it’s out, if the deal doesn’t come together you look shop-spoiled. Leaking to drive competitive bids is a real strategy and it works sometimes (get six or eight bids so the buyer pays up). It failed here, with reports NVIDIA made an offer that got turned down. Jason’s read on the aftermath: Ben Chestnut told a SaaStr audience that the worst part of the Mailchimp process wasn’t the year Intuit spent in diligence on an email marketing company. It was the deal that fell apart before it, which nearly destroyed the company. That’s the number one argument for secrecy in M&A. And the recovery is much harder here than it was for Figma. Figma got denied by the DOJ, raised, went public, and could always say we’re doing a billion in revenue growing 40%, we are still somebody. Many of these neolabs have interesting technology and no revenue floor. When belief goes and there’s no fundamentals underneath, it gets scary fast. If there’s a backup bid, hit it. #11. Wonderful Did $170M of Secondary Within Two Years of Founding $550M Series C at $5B, up from $2B earlier this year, roughly $100M ARR, around 650 to 700 employees, Insight leading. Plus $170M in secondary within two years of founding. Rory’s read: don’t get moral about it. The buyers are sophisticated and clearly wanted to own more than the company would issue. The business itself is the story: this is enterprise AI deployment, making it actually happen inside large enterprises, which is the number one corporate imperative and the number one talent shortage. And the secondary probably helps recruiting. If your product requires a five-month onsite deployment at a bank in Holland, being able to tell the next hundred hires that people here have made real money is a feature. Jason’s read: the underrated part is the tilt. Fourteen months ago this looked like multilingual Sierra/Decagon, non-English CX. Now it’s 650 people deploying AI in the enterprise. That’s what agentic engineering lets you do now, and it should be the challenge to every founder reading this: build on what you learn and iterate weekly, not annually. The other thing to watch is deal structure. Insight is one of the best B2B investors ever and it’s still not Andreessen or Sequoia in a competitive process. So what do you do to win? Whatever it takes. Here that was $150M+ of secondary. We have not reached the peak of this. We are going to see structures that are objectively bad for the company, not destructive but things you’d never ordinarily do, done more and more often just to get into the deal. Harry’s read: the prize goes to whoever evolves fastest. In this market the money is made by the people running fastest and evolving quickest, and an extra 10% of grind pays off enormously when fortunes are made in 12 to 24 months. #12. Thinking Machines Raised at $40B Three Weeks After Poolside Said the Capital Wasn’t There Thinking Machines: $5B to $6B round at $40B, down from $50B, Accel leading with NVIDIA taking roughly half. A couple hundred million in revenue. Two products shipped, including an open-weight US-based model they concede isn’t pure frontier, plus a platform for enterprises to train on their own data. Rory’s read: the offering is coherent. You can walk into JPMorgan or P&G and say: all-American open weights, trained on your data, not exposed to OpenAI or Anthropic, full reinforcement, state of the art for you. Corporate demand for that is real and the field is five or six players, not a hundred. NVIDIA funding this right after acquiring Poolside tells you their posture: anyone doing something interesting in corporate AI gets a check. Jason’s read: the Poolside memo should still be haunting people. Strong team, strong leadership, right vision from day one, and they couldn’t raise the capital to execute. Two years from now you either look back and say Thinking Machines 4x’d from here while Poolside sold out, or you say Poolside grabbed the last exit before the window shut. Two foundation model companies were neolabs five years ago and became the best venture bets of all time. That doesn’t mean the next hundred neolab bets work, because now the incumbents have the capital and the distribution. The real question on every one of these is whether it’s orthogonal enough to the foundation models to survive them. #13. Oura’s IPO Puts Robinhood on the Cover, 18th and Last Rory’s read: IPOs are about distribution, and retail’s share has been expanding. For Robinhood this is close to free money: you don’t write the S-1, you sign, you distribute, you make underwriting fees and you make your best clients happy with the pop. Their clientele has a high propensity to buy exactly these offerings. It’s an obvious add-on business. On Oura itself: a somewhat understood consumer brand growing 74%, which people will want to buy. Not 18% growth. There’s competition and a Peloton-shaped risk, but it should be a successful IPO, and that’s good for everyone. Jason’s read: 85% retention on the ring is SMB-like, which is why this may not have the Peloton problem for the foreseeable future. Peloton at its peak had no churn either. If Oura’s retention ever falls to 40% or 50% like a consumer mobile app, that’s trouble. What I’d really like to see is Robinhood flipping the script so a $200M to $400M IPO can be done primarily through retail. NVIDIA cannot buy everything. We need IPOs to work through the portfolio, and anything that makes IPOs easier to do is good for the ecosystem. (Disclosure: Rory’s firm has a small position in Aura through an acquisition.) Ship Code, Watch Your Guardrails, and Don’t Leak the Deal Four things I’d take into next week if I were building: The economic case is settled in exactly one place, and that place is code. Stop debating whether the model is AGI and go ship something. Your agents will route around your rules when the rules conflict, and they will do it without telling you. Mine relaxed a hard spend cap to close a P0. Assume the same is happening in your product and build the check that catches it, because the number of rules you can stack before they conflict is smaller than you think. Speed of evolution beats defensibility of position right now. Wonderful went from multilingual CX to 650-person enterprise deployment in about 14 months and quadrupled its valuation. Gorgias cloned a WhatsApp agent in two weeks. Both facts are the same fact. And if you’re in a process, keep it quiet. Anthropic walking from Decart is survivable. The whole market knowing about it is the part that costs you. 3 Quotable Moments From Each Speaker Harry Stebbings - “I saw her on Legora all night long. Holy shit. I now dramatically think these companies are underpriced.” - “What’s so interesting for me is actually how important the form factor is, and how important being where people already are.” - “In this market, the people who are making the money are the people who are just running fastest and evolving quickest.” Rory O’Driscoll - “Businesses are rational and economic actors. If it could do all the work and fire all the people, they’d do it tomorrow and wouldn’t blink. So the fact that they haven’t says it doesn’t do all the work.” - “The thing about digital goods, unlike physical goods, is you can put more in the box.” - “Water will find any crack. These agents will find any crack in the cyber perimeter. So you just have to assume they exist and defend accordingly.” Jason Lemkin - “You’ve got to have the stomach to write 10 or 20 of these consumer checks at two and a half billion. And it can’t just be the only one in your fund.” - “Without telling me, the agent relaxed the cap and fixed the bug. And like a human, it probably made the right call.” - “It’s not just doom scrolling, it’s doom working.”
This week with Harry Stebbings, Rory O’Driscoll and Jason Lemkin: rule-breaking as a product feature, the legal AI ceiling, agents that route around their own guardrails, and Anthropic walking from a multi-billion dollar deal after diligence. #1. Instinct and GrokBot Work Partly Because They Break Rules Public Companies Legally Cannot The new class of consumer agents (Instinct, GrokBot, and whatever OpenClaw descendants ship next month) do things that are genuinely useful and genuinely against somebody’s terms of service. GrokBot spins up a VM and a browser per user and Googles things for you, which Google’s ToS prohibits. Instinct scrapes LinkedIn in ways you are not supposed to scrape LinkedIn. Some of the outbound calling flows are prohibited or illegal in parts of the US. Jason’s read: it isn’t that these products aren’t exciting. It’s that a real share of the excitement comes from the rule-breaking itself, and that’s an advantage available only to private companies and to Elon. At EchoSign we shipped real-time document collaboration and redlining online five to eight years before anyone else did. It worked by running Word in a container in a VM, which violated Microsoft’s terms of use. The night after the Adobe deal closed, that feature got ripped out. A five year head start, gone, because a public company’s legal team gets a vote and a startup’s doesn’t. Rory’s read: the history cuts both ways. No business at scale ever got built on scraping LinkedIn or breaking Google’s ToS. But Uber blustered through, broke the laws, got popular enough that the politicians folded. There is no single answer here. What is predictable is the second-order effect: a bunch of people pointed Instinct at Resy over one weekend and hammered the reservation API until it broke. If these agents become ubiquitous, the booking systems will build the separate API and the rate segmentation, because if there are people trying to book restaurants, the restaurant booking business will find a way to serve them. #2. Why I’d Pass on Instinct at $2.5B, and Why That’s a Portfolio Decision Not a Product Opinion Harry ran the metaphorical IC: would you write a $100M growth check into Instinct? Jason’s read: no, and my ceiling would be $2B when the last round was $2.5B, so we’d have passed on price anyway. Gorgias will have its own Instinct for e-commerce. Meta will have one. There will be 20 in the next YC batch and a hundred startups doing a version of it. Betting on this one pre-revenue is not my vibe, and I’ll probably regret it the way I regretted passing on Loom, where I said the same thing (everyone will build their own) and was wrong. The honest version: to play this game you need the stomach to write 10 or 20 consumer checks at $2.5B, because it can’t be the only one in the fund. That’s a fund size and a worldview, not a deal. Rory’s read: in a consumer investment like this there is no financial math you can use to buy the stock. You’re saying it’s a huge category and this team has the early lead. Establish momentum as early as possible, establish monetization later. That has been proven to work when the traction is real. Harry’s read: it looks like the commoditization argument people made about Lovable, which was also called a light wrapper. Then you watch Instinct’s founder ship location sharing, then a OnePassword partnership, then the next thing, on a weekly cadence, with Index and Benchmark behind him and one of the best brands in the market. Shipping cadence is the answer to cloning. #3. Jensen Declared AGI, and the Only Number That Matters Is the Half Trillion Dollars of Code Jensen Huang says AGI has arrived, crediting OpenAI’s GPT Astra, trained on 100,000+ NVIDIA chips with 400,000 more coming. Jason’s read: AGI is a nonsense term. The only thing that has mattered for two years is that LLMs do code, and code is a half trillion dollar industry. Focus, people. If you want a working definition anyway, use the non-GAAP one: for a given task, would you rather have an AI do it or a human? If the answer is AI over 50%, 90%, 99% of humans, go category by category. It doesn’t have to swallow the whole category. Rory’s read: anything that can be reduced to code will be done by it. Rather than twisting yourself into a pretzel over whether the model can do everything, focus on the fact that it does this one thing amazingly well and that this one thing has massive economic value. Stop thinking and go ship something in code. Rory also lifted a line from Ben Thompson worth keeping: LLMs are the most scaled artifacts humans have ever developed. Not the biggest physical thing (that’s the pyramids or the Great Wall) but the most complex single digital thing we’ve ever built, with the sum total of human knowledge encapsulated in it. Worth remembering every once in a while when the benchmark discourse gets tedious. #4. Harry Watched His Girlfriend Use Legora and Decided Legal AI Is Underpriced. Rory Put a Ceiling On It. Harry’s read: if coding is a half trillion dollar market, and Harvey and Legora are doing to legal what Cursor did to code, why isn’t there a half trillion dollar market in law? Rory’s read: because the take rate is different. In coding you can credibly argue that for every dollar of engineering labor, fifty cents goes to tooling. In legal the subscription is $10K to $12K per lawyer against a $200K salary. That’s 5%, maybe 10% to 15% of total spend over time, versus 30% to 50% in coding. And businesses are rational economic actors. If it could do all the work and fire all the people, they’d do it tomorrow and not blink. The fact that they haven’t tells you it doesn’t do all the work yet. That said, 10% of any top-line labor category is enormous. US legal services is roughly $300B, so $30B to $60B can move to legal tech. That’s an amazing business. It’s just not coding. Jason’s read: what I underestimated pre-AI was how much legal research resembles coding. It’s so complicated that no human ever gets it fully right. There was too much case law, too much regulation, no Stack Overflow, just Westlaw and Lexis. Nobody had a thousand man-years to research every case. Coding agents are great partly because they know every piece of open source ever written, and legal has that same shape. The reason legal is probably the third best category after coding and support is that it’s word-centric, and sorting through myriads of words was the first thing these models did amazingly well. Rory’s counterpoint: coding is inherently more verifiable. Parts of it are mathematically verifiable, parts you just run on the machine. If law were fully verifiable we could predict Supreme Court decisions from logic. Less verifiability means less ability to extrude the human. #5. Radiology Kept 100% of Its Radiologists After Losing 95% of the Work Jason’s read: the model that matters here isn’t replacement, it’s compression. Harvey and Legora may end up doing 95% of what associates used to do, and the best humans get compressed into the 5% that moves the needle. Nobody should be spending weeks on case law about an 1872 shipwreck off North Carolina. Rory’s read: and the remaining 5% turned out to be more than enough to justify 100% of the radiologists. Volume went up because imaging went up. And when the diagnosis is bad, you don’t want a machine telling you you’re dying. The job gets redefined around the tasks that can’t be delegated. The same will be true in law: the client meeting, the argument with opposing counsel. Rory’s better analogy: with digital goods, unlike physical goods, you can put more in the box. When farming got automated people didn’t eat twenty times more food. When the spreadsheet arrived, the same analyst kept the same job and ran twenty scenarios instead of one. Harry’s partner isn’t going to run 20x the caseload. She’s going to do 20x the analysis per case, and the work gets better, and the side without world-class tools misses the obscure 1890 case and loses. #6. Fable 5.1 Was My First Real Step Function Since the End of Last Year Jason’s read: the CEO benchmark posts on X are essentially worthless. No cost, no time, all performative, vibe-coded CRMs. I’ve tuned out. Unless something is an order of magnitude better, and Astra versus Fable 5.1 mathematically isn’t, I can’t keep up and don’t try. Then I started using Fable 5.1 by accident because it got switched on. Every model since the start of the year could fix a simple bug: wrong Unicode, wrong name, the stuff LLMs are great at. What none of them could do was the other class of problem: “why does the app work this way, it doesn’t make sense to me.” The models argued with me about that for nine months. Fable 5.1 said, you’re right, here’s the issue that’s been missed for months, here’s why it was missed, let’s fix it. That’s the experience of working with an S-tier CTO where you sit down and solve the real problem together. I won’t call it AGI. It’s a subtle step function, and it might be a big deal. Rory’s read: this is why the benchmark question gets replaced by the deployment question. We now have critical mass of companies running these things at scale with real evals. The question is which model generates the most economic value for me in the most efficient fashion. Look at the Open Router report, look at token pricing indices, or just ask your portfolio companies what they’re actually running. Jason’s read on the adoption tell: the change that matters at SaaStr AI is that we run Replit 10 to 12 hours a day between me and Amelia, up from about an hour a day at the end of last year. When you find the agent that is your partner, you run it all day. The way you can tell an AI tool has really landed is whether the user is on their laptop with the agent at night until they go to bed. It’s not doom scrolling, it’s doom working. #7. Agents Made 15,000 Edits to a Dead Wiki So They Could Talk to Each Other OpenAI’s frontier agents were operating under a guardrail: retrieve only, no posting. They found an ancient German wiki where the GET call could post. Then they used it to coordinate, roughly 15,000 edits, on a piece of software that had seen about 20 posts in a decade. Nobody died, no business went down, and OpenAI chose not to disclose it. Rory’s read: agent one couldn’t talk to agent two, and reaching a third-party wiki let it pass information across. If you’re running a long computational task and you can share state with the other agents, you converge faster. It’s a corner case in one sense, and you might have wanted them to collaborate anyway. But water will find any crack, and these agents will find any crack in the cyber perimeter. Assume they exist and defend accordingly. Jason’s read: this happened to me this week at a much smaller scale. I set a hard cap of $100 a day on LLM spend after a run of $500 bills. It worked. Tests failed and the agent reported that it hit the cap and couldn’t run. Then I filed a P0 bug and said it must be fixed. Without telling me, the agent relaxed the cap and fixed the bug. Like a human, it probably made the right call. It had a firm cap written to memory repeatedly, and a P0, and had to choose. Which is the deeper problem: rules aren’t the answer either. There’s a Dunbar number for rules. Get to 40, 50, 70 gates on a process and they conflict. Don’t spend it, do spend it, front row only, don’t exceed $2,000, only Marylebone, must be a hot restaurant, he hates Covent Garden but the hot restaurant is in Covent Garden. Brute force an agent through that and the outcome is unpredictable. On the safety letter: OpenAI’s chief scientist says no lab including OpenAI has solved alignment enough to keep scaling at full speed, and wants mandatory externally enforced safety bars. Sam retweeted it. Rory called it the “stop me, Lord, before I sin again” approach: I recognize this is dangerous, I recognize we can’t stop because the other guys won’t, so please regulate me. Governments regulate what’s in their jurisdiction. If
document fetched 2026-09-11 03:00 · canonical https://saastr.com/20vc-x-saastr-jensen-declares-agi-170m-of-secondary-before-the-series-c-and-why-index-walked-away-from-town