When we started Applied Compute this was our thesis in a nutshell. "Companies need to turn their workflows, domain knowledge, and accumulated judgment into AI systems that improve with each use. Private evals should capture whether a model is actually improving against outcomes that matter to the business (not just external benchmarks!). Private reinforcement learning environments should let models grow stronger on real traces from inside the organization. Its knowledge base makes institutional memory queryable and use of tokens more efficient. This loop becomes the new IP of the firm. I think of it as a hill climbing machine. And unlike most assets, it compounds. Every improved workflow generates better training signal, which accelerates the accumulation of tacit knowledge unique to the firm. The companies that build this early will have an advantage that is hard to replicate, regardless of any new individual model capability."

Lux family co @appliedcompute brilliant founder @ypatil125

Microsoft could have easily chosen to define the frontier as dedicated access to GPT and Claude on Azure. AI Foundry had a durable business serving these models within the Microsoft ecosystem. Two models could have dominated market share of all tokens. In this world, we would be on three-month release cycles hoping that GPT/Claude-Next's new SOTA on public benchmarks would translate into wins on our private evals. The only way to compound on data would be through prompting, muddling more and more context into the first user message. An ecosystem of exclusively frontier models no longer makes sense where the following trends trends have taken foot: (1) to improve capabilities across the board (FrontierCode, GDPVal, etc.), general intelligence requires a scale that is extremely expensive to serve; (2) there's no free lunch in upgrading to the newest model as scarce GPU compute has driven costs up (see the recent Anthropic and Google deals to serve on Colossus); (3) training a state-of-the-art model on just your own tasks is possible as frontier training infrastructure is now available to the public. The new architecture will combine "generalist" models with "company veteran" models that improve the same way that star human performers do: through learning from experience operating inside of your institution. The technical stack looks something like the following: (1) You'll need to automate how you transform production data into private RL environments. This means transforming unstructured data into a curriculum a model can learn from that looks like what happened in prod: e.g. replicating a SEV by mocking the state of a production database when it happened, with un-hackable graders that are aligned with what you care about in production. (2) Private RL environments need a post training stack to be useful. Model weights/checkpoints trained on these environments will participate in the cadence of traditional software release cycles. (3) Inference endpoints will serving production traffic become "alive" as they become attached to a training runtime. Each new batch of data produces environments that are inputs for the next training step. Each step produces a new release candidate for production; if it passes the A/B test, you'll do a rolling weight update to models that serve higher quality tokens for your customers.

Your data is your edge, but only if your AI is built on it. Rent a generic model and so can your competitor. The companies with an edge are deploying custom models that they own and improve over time. Our co-founder @rhythmrg recently stopped by @southpkcommons to share how companies are owning their intelligence with Applied Compute.

People are starting to realize the value of post-training and RL

One of the most brilliant teams Lux has continued to invest more to fuel them in every round possible! Thank you @ScottWu46 for your talent magnetism as a founder leader and for the bar of brilliance and momentum you’ve set for @cognition !

1/ We’ve raised over $1B at a $26B valuation, led by @Lux_Capital, @generalcatalyst, and @8vc. Our enterprise usage has grown >10x since the start of this year, and our run-rate revenue grew to $492 M. We launched Devin two years ago as the first AI software engineer. Since then, cloud agents have gone from niche to mainstream, and today they are the fastest growing way to create software.

Huge congrats to the team! Relentless execution and focus ftw

I wanted my first public video to be learnings from the past 7.5 years here at @AppliedInt. Hopefully it’s useful to the next set of builders 🐻

behind the scenes at AC, featuring what we now call the pinapple shirt excited to share some of the work @raymondmfeng and I have done, RMSD has proved super valuable in our customer engagements!

Today we're announcing our Series C funding: $355M at a $4.65B valuation, led by some great investors @generalcatalyst and @Redpoint. We've had insane growth in the last year, but we're still very early. So proud of the team and what we have built so far!

Huge congrats to @bernhardsson @akshat_b and the whole @modal team! In an age where so many teams are moving to own their own models, Modal provides powerful primitives we use for training and serving models at scale.

Frontier models set the floor. Specialized models raise the ceiling. With Modal, @AppliedCompute is training custom agent workforces for companies like DoorDash, Mercor, and Cognition.

Exactly! The winning strategy is not betting on who has the best model this month. It is building the deployment layer where intelligence actually compounds. That means serving the best possible agent tokens on durable infrastructure: route to any model, train your own when it makes sense, and own the context, harness, environment and interfaces around the agent. Applied Compute is building this customer-first deployment layer. We help customers build intelligent systems where the value compounds on their side.

The real power of forward deployed engineering has always been putting strong technical people directly alongside the operators who own the outcome. That proximity forces the work to solve the actual problem instead of some sanitized version of it. In the AI era this principle has become even more valuable. Agents can now sit inside real workflows and improve from actual decisions, which means the highest-leverage work is extracting the tacit knowledge that lives with subject matter experts, building evaluations that reflect how things actually break, and closing the production feedback loop so agents get better from real outcomes.

For the better part of 4 years, I’ve considered myself reasonably early to emerging sub-themes in AI and Robotics. Every time I begin researching a new one, I encounter promising private companies and, without fail, @GavinSBaker or @wolfejosh are already investors in them.

Harvey is a great example of a company carving out a strong competitive position by building proprietary intelligence We had a great experience teaming up with them to support their new Legal Agent Benchmark with post-training and eval methodology Thanks @gabepereyra for visiting during our team all-hands today to break down what proprietary intelligence looks like in law!

seeing all the old house photos really brings me back. lots of fun memories and late night gym sessions... wish I took more pictures but all I have is us trying to recreate the cognition logo with dumbbells

We started the company knowing that, despite remarkable progress on public frontier models, there was a frontier that had not yet been explored. The destination was clear (finding ways to leverage data, internal processes, and knowledge built up over many decades to produce systems that get better over time), but we didn't have the infrastructure to get there. The "private frontier" belief has played out more now, as the winners of this era will get there by honing their internal intelligence every day.