Max

@appliedcompute, @Stanford

[Dan DiPietro @danyulol]

https://t.co/Q5HIFZB4GF

[Yash Patil @ypatil125]

Most companies aren't leveraging their most valuable asset - their traces.

There’s a ton of software value to build around raw inference and models like Jev open up a whole new set of possibilities because of their architecture and how cheap they are to run.

When you’re generating billions of tokens across training and production, you need to understand which failures keep happening and how often.

In this example, we use a frontier model on sampled traces to build a failure taxonomy. Then, we freeze it for an annotation pass and use Jev to classify the full corpus. This way, the expensive work of figuring out what to look for doesn’t need to happen on every trace!

For one annotation across 10k traces, our benchmark estimates came out to about $11 with Jev versus $479 with Haiku 4.5. The implication here is that it is a lot more practical to build scalable systems around model observability.

We’re building a bunch of stuff like this in AC2 because we want customers to get more out of their inference.

We are in a world where the number of tokens being produced is increasing exponentially. This only highlights the need for observability infrastructure.

[Bryan Lee @_brylee10]

I implemented a system in @appliedcompute’s platform for automated failure mode clustering with Jev to surface errors at an even larger scale than before.

RL training produces billions of tokens in traces. I always manually read many traces to understand model behavior, but finding agent failures (like reward hacking / hallucinations) at scale is easy to miss without automation. Here’s how it works:

[Yash Patil @ypatil125]

Ever since we started serving inference for customers a few months ago, our business has taken off. We are clearly seeing that training wins (and keeps) inference workloads.

Inference isn’t a commodity when you can help customers improve the model - not just serve it.

Models should get better the more you use them. When you have evals or metrics that you want to optimize for, you can use online training techniques like On-Policy Self Distillation or frontier grade RL post-training to systematically target and improve specific behaviors in your model for your use case.

Towards inference that enables continuously improving models! More to come soon!

[Yash Patil @ypatil125]

The hyperscalers didn’t just build businesses renting out CPU compute. They built extremely valuable products around data, networking, security, storage, etc. We’re now building around AI accelerators, but the playbook of building value add software on top is the same.

Turns out GPUs are really good for running AI models. That is why inference engines, which take a trained model and compute and produce intelligent tokens that can do real work, have become the first piece of this software stack. But, there is so much more to build than just this!

Training helps you turn compute into better models that produce more valuable tokens. Routing helps you pick the right model to generate those tokens at the right cost and quality. Security monitoring helps you check what’s going into and coming out of those models. Then there’s agent observability, context management, sandboxing, etc. All of these things will be part of the new AI stack and the opportunity is much bigger than just serving a model.

This definitely isn’t a winner-take-all market, whether we’re talking about open models vs. the frontier or the players within each category. Customers want choice, flexibility, and access to the fundamental building blocks to create their own systems.

Constrained GPU supply will actually act like a regularizer and draw this fight out longer. Customers are looking for both compute capacity AND value add on top of it. When someone doesn't have capacity, you go somewhere else. That means more players get exposure to customers and the opportunity to address value add.

Getting a customer because you have available GPUs is different from keeping them because your software is better. In the limit, the value add on top of the GPU will win out. The companies that do the best job building that software AND verticalize the fastest to own everything from chip to token will take the lion’s share. It’s important to be building for that now, even when the immediate customer need is just more compute.

[Mikhail Parakhin @MParakhin]

@appliedcompute is the best actual Post-training/Finetuning company I came across in a long while. No relationship with them - just feel like people doing good job deserve a shoutout.

[Max @mdigi14]

Great team and great mission

[Shan Reddy 🍎 @rshanreddy]

America is banning AI in schools.

China is using AI to create geniuses.

Introducing Aristotle: The AI tutor that solves America’s broken education system. https://t.co/hqPJN8cke0

[Shan Reddy 🍎 @rshanreddy]

America is banning AI in schools.

China is using AI to create geniuses.

Introducing Aristotle: The AI tutor that solves America’s broken education system. https://t.co/hqPJN8cke0

[Applied Compute @appliedcompute]

“50% of DoorDash’s agentic restaurant orders are going to places users have never ordered from before.”

@andyfang tells our CEO @ypatil125 what happens when agents become the discovery layer. If models increasingly decide what gets surfaced and bought, companies have a strong reason to train and own that intelligence.

[Yash Patil @ypatil125]

Specialized models are the future! So much exciting stuff to build and scale!

[Applied Compute @appliedcompute]

@andyfang and our CEO @ypatil125 get into what actually counts as proprietary data.

Sometimes it’s obvious, like customer behavior or merchant data. Other times it’s buried in a support agent hearing “happy birthday” and knowing to send a cake. The opportunity is to turn the judgment @DoorDash has accumulated over years of operating into proprietary intelligence it can own, train, and compound.

[Applied Compute @appliedcompute]

“Until you actually see things operationally, it’s going to be hard to build DoorDash from scratch.”

Our CEO @ypatil125 sat down with @andyfang on why cheaper software doesn’t erase years of operating advantage. @DoorDash’s moat is its proprietary data, edge cases, and hard-won knowledge, and increasingly, the models trained on top of it.

[Sahar Solimano @saharsolimano]

Interesting precedent is being set.

Every AI company now has to assume access to a closed model is ultimately revocable, even after years of integration. If the intelligence is core to your product, the lesson is simple: you need a model that can’t be revoked. Build and own it yourself.

[Sahar Solimano @saharsolimano]

@raymondmfeng and @lbz____ make training multi-trillion parameter models look easy. (Spoiler, it’s not, but AC2 makes it much easier for everyone else!)

Supporting Kimi K3 forced upgrades across the stack, from memory and rollout precision to weight transfer, checkpointing, and communication. Those improvements now make frontier-scale training cheaper and more efficient for every model on AC2.

[Applied Compute @appliedcompute]

We are excited to support GLM 5.3 for training and inference on the Applied Compute Platform.

Post-train it on your data, run inference at scale, and keep the weights.

[Ben Scharfstein @benscharfstein]

GLM-5.3-flash is going to be a monster for post-training. The price performance is insane, both for training and inference.

If you’re interested in training or serving, DM for access to AC2 (Applied Compute’s post-training and inference platform)

[Applied Compute @appliedcompute]

We expect GLM-5.3-Flash to be one of the most popular models to post-train with AC2. DM for access.

[Applied Compute @appliedcompute]

Today, we're introducing AC2, the Applied Compute Agent Cloud, to enable every team to train, serve, and improve their own frontier models.

[Yash Patil @ypatil125]

At Applied Compute, we believe Training wins Inference.

In the limit, inference alone is a commodity: providers serve the same models, compete on price, and race toward thinner margins.

Training changes that.

Once an AI workload reaches scale, serving a generic model stops making sense. Companies will fine-tune, distill, and continually optimize models around their own data, product, economics, and definition of quality.

That is the real advantage of open-weight models: you have the weights.

You can make a model better on your evals, faster for your workload, and cheaper per token. Over time, companies will choose the best model for their specific task - not the best generic model.

We believe that, in the long run, more than 90% of inference on open-weight models will come from trained variants rather than untouched base models.

Training also de-commoditizes inference!

A provider serving the same off-the-shelf model as everyone else has little room to differentiate. But a company will pay more for a trained version of GLM-5.2 that materially outperforms the base model on the evals that they care about.

The providers that help companies train differentiated models (Applied Compute) will be best positioned to serve them in production, too. Training and inference will converge into one continuous system.

The long-term end state is continual learning: models that improve from proprietary data, real-world usage, and direct feedback.

The companies that close this loop fastest will build products that get better with every interaction and compound their lead over time.