@DevinAI and I fixed this for good today. ☺️ Introducing logostoslides.com - the best way to find and download multiple logos at once with exclusively transparent backgrounds, high-res, etc. Bankers, consultants, and other deck-builders of the world, salvation is here!

Astra getting worse at chess at it “learns” is the most relatable thing I’ve seen an AI model do 😔

This, but Civilization :)

So many venture investors are used to the last decade of underwriting where the main risk was the ability to capture demand - find pmf, scale sales etc… For many of the new businesses being funded today demand is no longer the issue - if you can make low cost reliable ballistic missile interceptors or high bandwidth low cost memory then ya obviously you will have insane demand. The risk moves to supply - ‘can you actually make the thing’ - a question literally no one asked in the era of HRISs & Neobanks - feels like a lot of faces will get ripped off people pricing ‘demand’ for magic beans that will never be delivered.

🥳 no more need to use @DevinAI from a mobile browser on my walk home from work

Pictured: toy trains, a sword fight, and my talk at the world's largest transport tech conference. 🚂🤺🚀 Did that really just happen?! InnoTrans books up two years in advance, and we somehow snagged a cancellation with just three weeks to go. THREE WEEKS to design a booth, write a talk, rebrand EVERYTHING, make swag, create a promotional video (thanks @HyperFrames_), and ship a ton of new code, all as a startup less than a year old (!!) We were the youngest exhibiting company by far, next to giants like Siemens, Stadler, Huawei and DB. Worth every late night. Came home with more follow-ups than we can handle, and now the fun part begins. Lucky to be building alongside @yonizim1, @RobertIshaq and Avia 💪

Introducing Exponential.NYC. Regular, thematic meetups of folks who are hitting token limits. The first event is soon- if you're interested, apply at the link! We have folks from Anthropic, Ramp, Cog so far but also a cross section of NYC: hedge funds, biotechs, etc.

Rebranding AI to SI

So pumped for this launch and perhaps even more pumped that @marwanmattar has returned to X! :)

This is what “dark inference” looks like! $1.5m in annual spend would be a top ten customer for many name brand SaaS companies

This, but life is and was always one turn. You either play it with honor or without, and if you pretend to act honorably only because it might be multi-turn, well, that’s without.

The fastest way to achieve career growth is to join a company that's growing so quickly that promoting you into an impossible scope is a better option than hiring from the outside. Then, all you have to do is the impossible.

No matter what your priors were, I think you have to accept that we're in a version of AI takeoff right now. Astra crushing ARC-AGI-3 is one thing; a model almost as intelligent for 20% of the cost three weeks (!!) later is something else entirely. Recursive self-improvement is starting to show up in two big ways: 1) Models that are dramatically cheaper than anything before for the capabilities (Jev, H3 Max, Sol 6.1, Opus 5.5, etc.)- pretty clear that the models had a hand in training/releasing these 2) Frontier intelligence too dangerous to release (Astra 6.1, etc.) And beyond that, the pace of model/product releases is now literally impossible to keep up with. To the point where it feels like something fundamental is about to break, at least for those trying to keep up.

What if everything goes right? Every day, I’ll take one startup from @lennysan’s top 100 list and sketch out the most ambitious case I can. How it started, what it's building, how big it could get. Day four: Applied Compute @appliedcompute Founding story: @ypatil125, @rhythmrg and @lindensli all worked at OpenAI on different teams. Yash, who is 23, dropped out of Stanford before working on post-training and Codex. Watching OpenAI train models using evals/hill-climbing, he had the insight that each company on earth has specific tasks/capabilities they need- and so there was a need for a company to help them train more specific models for themselves using their proprietary processes and data. In October 2025 (a year ago!) the company launched publicly with $80m in funding and Cognition, Doordash and Mercor as early customers. Product vision: Applied Compute has a Palantir-ish spin- it supplies both applied AI expertise and infrastructure to train models. This means it employs applied AI research engineers who are deeply embedded with customers to help them train models for their use cases. Like Palantir, its aspirations extend beyond that. It built an internal training and inference platform to help it serve customers, and recently announced that it would make it available to customers under the name AC2, which it called an "applied agent cloud". It has also expanded into helping customers run models on its infrastructure, and its stated aim is to monetize more on usage than on the training step. The more I read, the clearer it is that Applied Compute is organized around the core operating principle that companies will train/host their own dedicated models, and the company that helps them do that has a claim to managing the inference and getting paid well if the model proves effective. There don't seem to be sacred cows when it comes to how Applied Compute helps them do that- which makes sense given how the world is evolving, especially the tightening of loops between models in deployment and training. Stats: > 46 employees on LinkedIn, up 4x Y/y > In talks to raise $350m at a $3.25b valuation, per Forbes. > Raised $80m at $1.3b in April > Has raised from Kleiner, Lux, Greenoaks, Elad Gil, Benchmark and Sequoia Bull case: Beyond assembling a stellar team that is over half former founders (!!), the Applied Compute team has aligned itself with a bet that seems to be paying off, as companies ranging from Harvey to Cognition enjoy huge success with specialized models that they've trained themselves (both are Applied Compute customers). While there will no doubt be enormous demand for general-purpose, "do anything," frontier-intelligence models, evidence is mounting that specialized models can be better, faster, and cheaper for specific use cases. Applied Compute is perhaps the company making the purest bet on that, and while other vendors will no doubt enter the race, that clarity of purpose seems likely to pay dividends, as does keeping a deliberately small/intense/elite team. If we enter a world where a meaningful share, if not a majority, of token spend occurs on these sorts of models (and that is a very plausible world), the only bottleneck to Applied Compute's growth may be executing against the pace and breadth of that transition.

Liftoff of Starship's first orbital flight

(turbopuffer co-founder) yeah, Cursor called us in July and told us their new model code evals didn't show the same gain from semsearch as previous models. selecting the right piece of context in a single turn used to be huge, but modern code-local agents can answer complex questions with simple search tools like grep(1) in local code bases, with a low-scale amount of data to search, and when you need a full copy of the data on your machine anyway (to edit and run), offloading to search infrastructure is overkill. grep works fine, and is how it's been done by expert programmers for decades search engines are necessary if you have lots of data to search over, and want to select the right context for the task rather than downloading all of it to your agent's disk. search engines also let you spend compute up front to save turns at search time. large datasets or high qps are where it shines we are close to the Cursor team, and will work with them on more large-scale search in the future :) we'd have done the same thing

Daily spend on 1-800-Flowers in our credit card panel. 😂

What if everything goes right? Every day, I’ll take one startup from @lennysan’s top 100 list and sketch out the most ambitious case I can. How it started, what it's building, how big it could get. Day three: Anduril @anduriltech It strains the definition of startup, but it felt interesting to write up. Founding story: Anduril was founded in 2017 by @PalmerLuckey, @traestephens and a cohort of Oculus/Palantir employees. Brian Schimpf serves as CEO, though Palmer serves as the public face of the company. Anduril was build from the beginning to be a next-gen defense contractor focused on AI-powered autonomous systems- sort of a hard complement to Palantir's success selling AI-powered software to the government and three letter agencies. Product vision: Anduril's products are built around Lattice, a software suite that combines command/control, mesh networking, autonomy and deep integrations with other software platforms to do sensor profusion at scale and connect Anduril's devices with each other as well as other sensors. Though its sentry towers (for managing boundary surveillance) and drone munitions are most mature (including deployment to Ukraine), Anduril has developed a full suite of next-gen military solutions, including counter-drone systems, an uncrewed fighter jet wingman (Fury), an autonomous underwater vehicle (Ghost Shark) and a family of cruise missiles (Barracuda). On top of all of this, Anduril has always emphasized mass-manufacturability and has developed Arsenal-1, a mass-manufacturing facility in Ohio designed to produce its systems at wartime scale. Beyond intelligence and Silicon-Valley style innovation, Anduril's core product vision involves cutting through the mess of subcontractors and component-makers that has historically bottlenecked mass production of complex systems (the F-35 program has 1800 supplier spread across every state, in part for political reasons). Tesla used vertical integration and iteration to build superior cars vs. automakers with their complex supply chains- Anduril is bringing the same approach to defense. Stats: > 9,944 employess on LinkedIn, up 57% Y/y > $2.2b revenue in 2025 > $11.4b raised from Founders Fund, A16Z, Thrive, Lux, Greycroft et al at > $61b valuation as of May Bull case: The nature of war is changing and Ukraine shows us that intelligent, autonomous and expendable systems are a requirement to be competitive in peer/near-pear conflicts. This will lead to both vendor turnover (as innovation is critical) and a sea-change in the nature of defense production, as stockpiles of munitions (and manufacturing capabilities) matter more than high-end systems to deliver them (i.e. fighter jets, aircraft carriers). On the surface, the existing big five defense primes (Lockheed, RTX, Northrup, Boeing, GD), seem to cap Anduril's size. They have $350b of combined revenue (including non-defense rev, i.e. Boeing's commercial business) against $689b of market cap. However, they effectively share their economics with a massive ecosystem of supplier/subcontractors and seem unlikely to reap the gains from labor displacement as autonomous systems become a larger share of defense budgets. In autos, Tesla has $1.4T of market cap on ~$100b of annual revenue while the automakers (Toyota/Volkswagen/GM/Ford/Mercedes/BMW/Honda/Hyundai/Stellantis) have $600b of combined market cap against $1.8T of revenue (!!). The bet on Anduril is that a similar dynamic will play out in defense, with legacy primes struggling to adapt. Anduril's Fury win is a powerful proof point- it beat Boeing, Lockheed and Northrup and shares the program only with another smaller player (General Atomics). This is the first time a major fighter aircraft program hasn't gone to a prime since the 1970s. The bull case is that Anduril is able to build a $100b+ revenue business selling hardware to the US and allies, and, as the key technology partner to the Department of War, is able to earn a call option on future weapons development as autonomy continues to gain share and keep troops out of harm's way.

I really think RRGP (run-rate gross profit) should be the new ARR. Infinitely better way to compare companies, especially in today's environment where revenue is consumption-based and often includes lots of pass-through and/or structurally lower GMs.

I probably watch this at least once/week. “Ain’t nobody gonna give you nothin. Everybody wanna be a bodybuilder, but don’t nobody wanna lift no heavy-ass weight.” - Ronnie Coleman piped.video/4UlgXIL0-3g