The best AI is built, not bought. Our platform, Applied Compute Agent Cloud, is now in private beta. Book a demo below.
San Francisco
Joined July 2012
- Tweets332
- Following29
- Followers9.1K
- Likes887
Pinned Tweet
Every company should own the AI it runs on. Applied Compute exists to make that possible.
We’ve crossed $100M in annualized run rate and raised $350M at a $3.25B valuation, led by @eladgil. 40% of customers now use Applied Compute for both training and inference.
Applied Compute update:
- crossed $100M annualized revenue run rate up from zero 18 months ago
- raised $350M at $3.25B led by Elad Gil
- 40% of customers use AC for both custom model training + production inference
- inference serving 1T tokens a day after launching 12 weeks ago
- 2X enterprise customers since last raise
- 3X team since January, 85% are researchers and engineers
- Working with customers like Microsoft, Shopify, Harvey, DoorDash, Cognition, Clay, Decagon and others
To move even faster we need people who are extremely technical AND extremely practical. If this is you, please reach out.
Poker is a hard post-training problem. Many players, hidden information, and results only show up at the end of a long game. It's even harder with only 24 hours on the clock.
We're sponsoring the PokerBench track at the @ValsAI hackathon this weekend to see what teams can train on AC2.
Spots are limited. We'll see you at the table.
Surprise! Vals AI is adding a third track backed by @appliedcompute, offering a $7,500 cash prize for the first place team.
Alongside Build Your Own Eval and Post-Train a Small Model, a limited number of teams will post-train a model to compete on PokerBench, Vals’ 10-player no-limit Texas Hold’em table.
Every team starts from the same reference setup and gets the exact harness their bot will be scored in. Bots play each other all weekend, with live rankings and overnight tournament checkpoints. First, second, and third place take home cash prizes.
We incorporated quantization-aware distillation (QAD) into our training platform, AC2. This lets enterprises run their custom models on less hardware without giving up the accuracy they trained for:
1) Compressing the right parts of a model cut its memory by about a third without hurting performance.
2) QAD can win back the accuracy that compression usually costs, in some cases matching the original model.
Production traffic is full of real experience from economically valuable work, but very little of it currently makes it back into training.
We built a recipe for production RL that steadily improves a model straight from its deployment’s data stream, then redeploys it to keep learning.
We are sharing some of our initial learnings from a recipe capable of continual learning in production.
Prod RL has to stay stable across every redeployment, since the end-to-end prod RL run is expensive and rolling back a few checkpoints may not rescue a collapsed run. Most of the instability comes from mismatch between training and test time, which starts with formatting the data.
We record the exact tokens the deployment generated to avoid mismatch from inconsistent tokenization.
Post-training a model usually takes a team of researchers running experiments for weeks. With Ari Autoresearch, now live on AC2, an agent can run that whole loop itself. In one 16 hour episode with no researcher intervention, it took Qwen3.5-9B from 0.118 to 0.629 held-out reward on chess puzzles.
Behind the scenes of auto-mode for Claygent with @jeffbarg and our Applied Research Engineer @_brylee10 and CTO @rhythmrg:
- 47% cheaper than sending every task to the most capable model
- <1 second to choose the model for each research task.
- You can still pick your favorite model anytime
Introducing auto-mode for Claygent. It picks the AI model for each row of research, balancing answer quality AND cost.
Now on by default for new Claygent columns:
- 47% cheaper than sending every task to the most capable model.
- <1 second to choose the model for each research task.
- You can still pick your favorite model anytime.
Super cool thing:
We post-trained an open-weight model with @appliedcompute to handle the routing. It runs on their AC2 platform and handles hundreds of millions of queries a month.
@jeffbarg, Vyshnavi, @rhythmrg, and @_brylee10 break it down below.
Claygent, @clay’s AI agent for GTM research, has processed over 5 billion queries.Claygent doesn’t need to use frontier models for every query. Smaller alternatives can be more cost effective and lower latency.
We trained and served a specialized router to pick the right model for each task, cutting inference costs by 47% with no impact on answer quality.
Meta’s Muse Spark Open is now live on AC2 for training and inference.
In early testing, we've found it to be a strong base model candidate because of its broad capabilities and inference efficiency.
Post-train it in your own harness with your own environments and graders, using full-parameter async RL.
On Wednesday, we launched Shift-1: a model router that can reduce coding agent spend by up to 55% while preserving performance. Today we are sharing more details about the research that went into this release.
We’re thrilled to be featured on the GTC keynote stage alongside so many incredible teams!
We want enterprises to have a unified platform to post-train open models like @NVIDIAAI Nemotron on their own data, serve them in production, and continuously improve them through real-world use. We can’t wait to keep building that future together.
Introducing Shift-1, a frontier coding router trained on AC2 that reduced coding agent cost by up to 55% with no noticeable drop in quality. Try it at shift.ac.
Read more on how to train and deploy your own custom router.
appliedcompute.com/platform/…
Reflection AI’s Beam 1.0 is out today, a 501B MoE with 23B active under Apache 2.0. We ran a brief RL post-training run on it in AC2 that tripled its score on a challenging long-horizon agentic coding task in 40 steps, 0.07 to 0.215.
We’re offering day 0 support in AC2 for inference and post-training.
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.
We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source harness we built and trained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
We were especially impressed with Beam’s inference efficiency during post-training. The architecture supports 1M context and stores minimal KV cache state per token, enabling training on long horizon tasks at high concurrency.
Thanks to @reflection_ai for the early access and for contributing a key open model to the Western ecosystem!
Nemotron 3.5 Super VL by @NVIDIAAI is now supported for both training and inference on AC2.
In early post-training runs for our customer, we've been impressed with the quality of the base model for its inference efficiency.
Introducing Spork, an efficient post-training recipe which increases our solve-rate on chess puzzles by 100% when compared to our best GRPO setup.
It uses a GRPO warmup, splits task-specific datasets further into partitions, trains specialists, and then uses multi-teacher on-policy distillation. Like a Spork, it leverages gains by simply consolidating specialists.
AC2 now supports decision model training. Teams can build custom Jev-like models for use cases like request routing, content classification, and anomaly detection.
Reach out to get started with your own classification task!
You can now train your own decision model in our platform AC2! Jev-like models, custom fit for your use case :)
We implement decision models through a language-model backbone with a small learned output layer called a decision head. The backbone processes the input, question, and answer choices, and the head produces scores which are softmaxed to produce a probability distribution over answer choices.
Our software interface makes it easy to add a variable-sized decision head to frontier open-weight models. Choose any open source model, and train it as a decision model using Brier loss or cross-entropy loss. Training updates both the decision head and the language backbone.
We tried it on 50k sampled training datapoints from the Civil Comments dataset, starting from Perplexity’s pplx-decider-v1-27b, and F1 for detecting toxic comments increased from 0.482 to 0.677 on 20k sampled eval datapoints. The training job took less than an hour on four B300 GPUs.
Reach out to get started with your own classification task: docs.appliedcompute.com/plat…
We're sharing a key step towards enabling Continual Learning at Applied Compute.
In On-Policy Self-Distillation, the teacher model supervises training with privileged information: a hint. The training signal depends heavily on hint quality, but manually refining hints across runs is impractical, and isolating their individual impact is difficult. We instead optimize hints automatically using efficient proxies for their eventual training value.
Post-training on your own data gives you control over your models and pushes out the Pareto frontier on cost, latency, and quality.
Our CEO @ypatil125 sat down with @jacobeffron to discuss how AC2 brings down the cost of post-training, runs online training in production, and serves your models with inference on the same platform.
When is harness optimization enough? When does post-training make sense? What will determine who becomes the new AI hyperscaler?
@ypatil125, co-founder of @appliedcompute, was the perfect person to discuss this and more.
Yash worked on Codex at OpenAI before leaving to co-found Applied Compute with fellow OpenAI researchers Rhythm Garg and Linden Li. Sixteen months in, the company is a unicorn, and is serving enterprises including DoorDash and Cognition.
We cover:
0:00 Intro
1:42 The Case for Owning Your Intelligence
4:35 OpenAI x Baseten and the Multi-Model Future
7:19 Can Post-Training Beat Frontier Models?
9:03 What Data Stays Out of Distribution
11:51 Data Efficiency and Continual Learning
14:31 When Updating Model Weights Makes Sense
16:01 Online Training in Practice
18:16 Lowering the Cost of Post-Training
19:48 Does Every Company Need Its Own Model?
24:06 RL in Non-Verifiable Domains
26:59 How Applied Compute Picks Customers
28:46 Building an Inference Business
34:43 The Case for a New AI Hyperscaler
38:21 Building RL Environments
41:33 Quickfire
YouTube: youtu.be/QeiYSLBCwD4
Spotify: bit.ly/4y6EJrz
Apple: bit.ly/4j1okRj
Try decision graders on AC2 today: orders of magnitude cheaper and faster while being a better representation of the rubric.
At Applied Compute most of our RL and eval pipelines grade with an LLM judge. These produce one result per rubric criterion, typically a paragraph of text accompanying a score. For some complicated tasks, grading can be a slow, pricey part of the loop.
We swapped a grader on a code QA benchmark in AC2 for a decision grader: one call, a yes/no probability for every criterion at once. On 100 sample tasks (407 criteria), same answers:
→ ~32× cheaper
→ ~8× faster
→ 94% agreement with the LLM judge
And it’s even better than the original judge:
When the decision grader was confident, it matched the LLM judge 99% of the time. When it wasn't, the decision grader indicated low confidence (30-70%), pointing to ambiguous rubric items. This allowed us to identify noisy tasks and make the dataset cleaner.
A cheap, fast, calibrated grader means you can grade more rollouts, ask more questions per call, and downweight or review the shaky calls instead of training on them.
Our reward hacking monitor is how our Applied Researchers, like @evandavidellis, catch models cheating during RL. This is calibrated on an internal benchmark of real reward hacks we've observed in our runs. Today, we're rolling out this monitor into AC2.
We're adding reward-hacking monitors into AC2. Every RL run gets checked for cheating, and our research agent, Ari, automatically investigates the flags.
After a model discovers an exploit, it can become its default strategy within a few training steps. We used to hunt for reward hacking manually, reading traces and running one-off agent queries. This was error prone and didn't scale. We decided to automate detection instead.
