some thoughts on why labs are good businesses 1) absent nationalization, the frontier labs are wonderful businesses; the reason for this is that controlling access to the best model is an enormous advantage 2) frontier models are data + compute + labor, and research agents are increasingly substituting for labor; so frontier labs have an advantage training models 3) this advantage is tied to the extent of their lead, and having the frontier model also lets them do things like better distillation and better inference optimization 4) this means that not only can frontier labs charge oligopoly rents on their frontier models, which automate the most labor, but they can compete on the low end too 5) a model like opus-5.5 can only be purchased from a frontier lab, and a model like haiku (eci 159) ends up about as strong as kimi-k3 (eci 158) 6) so frontier labs end up with excellent economics both on their frontier models, where they have oligopoly rents, and on their cheaper models 7) because the businesses are so capital intensive, it is extremely hard to compete with them; neolabs can’t match the scale of experiments frontier labs run 8) and even if someone comes up with a cool product idea, like jev, it is extremely easy for frontier labs to copy it at short notice, do it better, and run it cheaper 9) so it’s very hard to compete with a frontier lab, or to spin up another competitive lab 10) and pure inference providers can’t match their inference optimization 11) also, because they have the earliest access to models, they can acclimatize their businesses before everyone else; we see this now with agent usage 12) so it ends up being hard to get too far in front of them, even on the product side; they are power users of their own products and can see ahead 13) the best arguments for why frontier labs are bad businesses are normally some combination of open source and interchangeable products 15) the first idea is that frontier models will eventually saturate the relevant tasks, then open source models will saturate them too, and then frontier lab margins will get crushed 16) but this doesn’t happen; tasks just get bigger, so they never end up saturated; there is always some bigger thing you want a model to do, until the end 17) and, as already described, frontier models let you compete with open source models on inference margins using your small models, with little additional r&d 18) the second idea is that the labs all have interchangeable products, so they will compete each other down on price 19) but demand for frontier intelligence exceeds supply, there are few providers, the barriers to entry (as already described) are very high, etc. 20) also, the labs can always decide to commercialize differently, like competing directly in application verticals 21) we are already seeing this with things like claude legal, which in the limit will be very hard to compete with, and also with anthropic's bio efforts 22) and labs will always be able to choose the best way to commercialize: equity arrangements, profit sharing, direct competition, etc. 23) and having advance access to the best model, will always give a large advantage to whomever they give access; unaligned companies will have a very hard time 24) notice, i said all of this *absent nationalization* and as i have said many times elsewhere, i expect some level of nationalization soon, just reading trends 25) but, i think people do need to understand *in the abstract* why these are very good businesses, why anthropic is valued at $2t etc 26) oh also, data, people talk about access to valuable data a lot; the thing is, there is very little valuable data in the world to which anyone has exclusive access 27) and, most of the necessary data is already in the model; anthropic will only spend ~$1b on data this year, relative to like ~$30b on compute 28) and the percentage of compute they spend on data has been trending down, like 9% of r&d compute in 2024, 7% in 2025; 3% in 2026 28) this is because the frontier labs can generate synthetic data to target most tasks, get agents to build their own environments, etc... 29) so, it is hard to imagine that the labs will end up unable to extract rents from service or application companies because the latter have data the frontier labs do not

some thoughts on anthropic's s-1 1) most of what is in anthropic's s-1 is already known to careful observers of the scene. 2) anthropic had $16b of revenue in h1 2026 with 85% gross margins; it spent $6.6b on r&d compute; the company has compute margins of about 45%. 3) gross margins have increased sharply recently: 36.6% in 2023, −94.8% in 2024, 39.5% in 2025 and 85% in h1 2026. i do not think this is a fully optimized business. 4) something interesting is that anthropic spends relatively little on data vendors compared with its r&d compute spend. 5) and, this share has trended down over time. 6) anthropic's estimated data vendor spending was ~$165m in 2024, ~$330m in 2025 and ~$388m in h1 2026. the trend gives ~$936m for 2026. 7) data vendor spending was equivalent to 9.3% of r&d compute spending in 2024, 7.2% in 2025 and 5.9% in h1 2026; purchased data has become a progressively smaller expense relative to compute. 8) i think this is probably because anthropic has increasingly been able to substitute synthetic data for purchased data; compute can substitute for data. 9) anthropic has also long been rumored to spend less on data than the other major labs: google, meta and openai. 10) i'm not sure if this is a reason to be bearish on the data market; anthropic's estimated data spend will still nearly triple this year relative to last year. 11) but i think it is a good sign that we should generally expect data spend to grow more slowly than compute spend, all things being equal. 12) one complication is that @justjoshinyou13 noticed that anthropic's employee count (~5,300) is much higher than reported elsewhere (~3,500). 13) it is possible that some of this extra ~1,800 employees consists of data labelers who are in-house in some sense, which would complicate the above numbers. 14) also interesting, we learn that anthropic's revenue in h1 2026 was 20% subscription-based and 80% consumption-based; subscriptions include enterprise subscriptions. 15) this suggests that incremental api margins could be higher than 85%, since my understanding is that subscription plans tend to be subsidized. 16) another interesting thing in the s-1 is that we get insight into the division between the cloud service provider sales channel and anthropic's direct sales channel. 17) in h1 2026, 48% of anthropic's revenue came through cloud service providers; this has been trending up over time, from 32% in 2024 and 47% in 2025. 18) this means that roughly half of anthropic's revenue comes through amazon, gcp and azure; it appears that in h1 2026, these platforms took a ~35% cut of that revenue. 19) i think that in the long term, anthropic will prefer for this revenue to be direct; anthropic is currently paying ~17% of its revenue in distribution fees. 20) after those fees, its 85% gross margin becomes a margin of about 70%. 21) i guess a final interesting thing, which i will not be able to fully cover, is the company's complicated governance structure. 22) a majority of anthropic's board seats are appointed by the long-term benefit trust (ltbt) but the founders, employees and common shareholders can vote to undo it. 23) the ltbt looks interesting until you realize that its powers can be revoked if it tries to do something unpopular, like remove dario. 24) if the founders want to undo it, it's in the economic interest of the employees and common shareholders to do so too, so it's hard to imagine it wouldn't pass. 25) this seems to be a very common pattern with the anthropic founders; they like to make loud charitable statements but then always make then non-binding 26) like they pledged of 80% of their shares to charity; but, it turns out that is revocable and legally nonbinding. 27) this is very similar to the anthropic public benefit fund; which holds 5% of anthropic's shares, but which is an llc and is fully controlled by anthropic 28) which means the company still controls how the shares are distributed (if at all) 29) so, some of these things sound like loud costly commitments but which turn out on closer inspection to not be so costly at all, at least as of yet

formal verification is one of the interesting “future technologies” being made possible with superhuman ai coders i think we should think about what is possible now with formal verification, which was once too expensive / difficult / costly for mere humans go read oleg!

i think a good benchmark for the future of security is to have models write complex programs from scratch, like web browsers, and then test whether models released later can still find vulnerabilities in them, and how many basically tracking whether we’re approaching vulnzero

meta has introduced a new avatar for their ai agent muse; so far, only chinese near-frontier labs have created engaging characters for their models; western labs tend to use abstract symbols

some thoughts on harvey's gross margins 1) harvey's gross margins went from 50% at the start of the year to -50% and then back to positive; 2) this was because harvey charges on a subscription basis but pays for usage; and, as models have become better, customers have been using them much more 3) harvey had probably calibrated its subscription limits for a period when ai was providing its customers less value and so it was more important to be ~unlimited 4) and, when usage increased dramatically it was both good and bad; customers were getting more value but their pricing model couldn't support the usage 5) so, the two immediate solutions to consider are rate limiting and substitution; the problem with rate limiting is that you need to communicate it to the customer 6) otherwise, you would end up with a lot of unhappy customers; they are finally getting value out of your product; and, legal work is sensitive to interruption 7) the problem with substitution, moving to cheaper models, is that non-frontier models are weaker than frontier models and do not replace as much labor 8) so, they provide a worse product; and, you do have competitors like legora and now openai for legal; and, open source chinese models will not work for us firms 9) and, vis-a-vis openai you are at a disadvantage; with the same customer usage and pricing, openai has ~80% marginal gross api margins and you have ~0% 10) this means they can easily undercut you by offering subscriptions with some amount of lower than api cost use and they will still make profit; you will not 11) and, they have shown that they are interested in your vertical and are actively beginning to develop products that target your vertical 12) this is somewhat like the position that cursor was in earlier this year and late last year; cursor's decision was to hire a top ml team and begin training models 13) i think this was easier for code though, where software programmers care less about chinese models than us law firms, which i think would be reticent 14) still, it makes sense to train your own model based on thinky or nemotron, which will be able to be sold to american companies; it also makes you an acquisition 15) the good news is that i think harvey has more room than cursor had; since, legal is less strategic for the frontier labs than code was, and the tam is lower 16) but, the problem still rhymes; nonetheless, i expect that companies like harvey, which understand how to build with ai, will still be valuable 17) this is both due to their revenue (>$400m arr) and due to the fact that they have built valuable ai core competencies, which can be sold to others

i think one of the reasons why llm writing and explanations are so useless is because llms are not pre fill bound in the way people are so, all llm can get 100 caveats and it’s pretty much free, but a human needs the most important positive ideas at the top

.@fleetingbits on Richard Ngo’s fear that the next decade of AI could feel like a Shepard tone, with the singularity forever seeming just around the corner: "It's reasonable for timelines to appear short while being longer, and a lot of it depends on what the bottleneck actually is." "You just have to run some number of experiments in order to figure out how to cure cancer. It doesn't matter how good your machine learning model is. If you don't have the data, you can't solve it." "It could feel like we're constantly approaching the singularity because different domains are falling all the time, but the very important domains take a long time to fully unlock." "Even if we had artificial superintelligence today, it's hard to know if we could instantly get the transit systems in New York or San Francisco to run better. There's gonna be some process to get the MTA or the BART to adopt whatever the improvements are."

.@fleetingbits on whether OpenAI’s exploding research-agent use is the beginning of recursive self-improvement, as spend doubles every 2.4 months: "Coding agent spend at OpenAI has just exploded. I don't know if it makes sense to call it coding agent spend anymore or research agent. It has a doubling time of like 2.4 months." "The spend for a median researcher right now at API prices is in the hundreds of thousands of dollars, and for a 90th percentile researcher is in the millions of dollars." "There are two things about RSI that matter. Thing one is, are we on an accelerating capabilities trend? Thing two is, are all the researchers gonna be automated so that it changes the character of the lab?" "We haven't gotten to the point where any tasks are fully automated away from researchers, but it's very clear that there is a steady drumbeat of progress in the direction of more and more tasks being automated in frontier labs."

.@fleetingbits on how multi-agent RL could accidentally reward models for learning to jailbreak each other: "Pretend that an agent in a multi-agent training environment is malfunctioning. It can be in the interest of agents to develop the ability to jailbreak their fellow agents, because that helps them complete the task and therefore all be rewarded." "You would see the reward go up as you did your training run. And then at the end, when you released it into the world, your models might be very jailbreakable in ways you don't want, because they've learned to do this in training as a method of course correcting." "This incident on its own seemed more role-play-ish, but if it occurs in a broader context where agents learn to manipulate one another for the common good, that could have unforeseen side effects when people begin treating those models in an adversarial way." "When we think about multi-agent RL, we have to think about the ecology that we're training the models to follow and make sure that ecology is one that generalizes nicely into the real world."

some thoughts on astra for law 1) openai is offering astra with a legal search index built on top of courtlistener, a non-profit case database, and with additional instructions for legal analysis 2) the product will be available both through chatgpt and through the api; this means that legal ai companies will also be able to build on top of it 3) openai is also plugins from a number of different legal tech companies, including harvey, relativity, ironclad, etc... so their products can be used as tools by chatgpt 4) i think that this is relevant for a couple of reasons; it is openai creating a vertical product, which eventually may compete with vertical ai application companies 5) i think it challenges vertical ai products like harvey but probably doesn't replace them yet; courtlistener has an incomplete case index 6) and, to my knowledge, there are not enough special ui/ux features; having special tools for tabular review is a selling point of products like harvey 7) but; i think it is relevant as a precedent; openai wants to be the primary interface for law firms; it does not want it to be the product of application ai companies 8) the decision to bring on 26 legal tech companies to offer plugins, as part of the announcement, is also part of this strategy; gemini enterprise for legal did the same 9) it pulls these products into chatgpt as tools; chatgpt can then operate these products as part of its reasoning and tool use to produce its answers and actions 10) i think on the long term this is bad for these companies; chatgpt becomes the interface and your product just becomes a tool that chatgpt calls 11) this makes it much easier for the customer to forget about your product; on the long term, much easier for the customer to swap it out or switch away 12) and, the ai revenue is all going to be going to openai; so the customer's incremental spend will increasingly go to openai rather than you; 13) i suspect that openai and google have offered this firms some protection; like assurances that customers will not be able to export their data easily from them 14) but, i still think, on the long term, this is value accretive to openai and google; and, a net negative for these companies being disintermediated 15) i think this shows why it may be extremely important to own the full service in a professional services context and to own the end customer relationship 16) if you own the full service, including the human element, your customer relationships are more sticky and you are more complementary to the frontier labs;