Pinned Tweet
The difference between smart agents and useful ones is institutional knowledge.Â
We’ve been building the infrastructure for Specific Intelligence at @appliedcompute, and today we’re announcing $80M in new funding led by @kleinerperkins.
We’re a team of 23 and growing - come join us!
We’ve been building the infrastructure for Specific Intelligence at @appliedcompute, and today we’re announcing $80M in new funding led by @kleinerperkins.
We’re a team of 23 and growing - come join us!
13
10
114
26,229
Rhythm Garg retweeted
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 t wait to keep building that future together.
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 t wait to keep building that future together.
0
3
38
1,816
We train custom models for our customers' highest-volume inference workloads. So we looked at our own inference spend and (of course) found a lot of it concentrated in coding agents used by our engineers.
We trained Shift-1 in our platform as a turn-level custom router to save money and preserve quality. Our team enjoyed using it, so we thought we'd share it with others too. Try it for yourself!
We trained Shift-1 in our platform as a turn-level custom router to save money and preserve quality. Our team enjoyed using it, so we thought we'd share it with others too. Try it for yourself!
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.
5
2
43
1,767
Rhythm Garg retweeted
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 ning.
We’re offering day 0 support in AC2 for ning.
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 ained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
10
15
90
8,674
We are excited to partner with Reflection to post-train their open-weight models for our customers! Beam is supported in AC2 and we're already seeing promising results from early training runs
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 ained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
1
1
68
2,120
Rhythm Garg retweeted
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.
In early post-training runs for our customer, we've been impressed with the quality of the base model for its inference efficiency.
2
5
46
2,930
Start with a brief full-task GRPO warmup, train specialists on random subsets, then distill them via Multi-Teacher On-Policy Distillation (MOPD).
Of course, splitting your task into semantically meaningful subsets yields stronger results, but random subsets actually captured ~2/3 of the gain on chess and still beat full-task GRPO!
Suggests that much of the benefit comes from learning from diverse, independently trained teachers, even without meaningful task decomposition
Of course, splitting your task into semantically meaningful subsets yields stronger results, but random subsets actually captured ~2/3 of the gain on chess and still beat full-task GRPO!
Suggests that much of the benefit comes from learning from diverse, independently trained teachers, even without meaningful task decomposition
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 Like a Spork, it leverages gains by simply consolidating specialists.
4
3
42
2,149
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/platform/train…
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/platform/train…
7
19
109
21,610
Congrats @moritz_stephan and team on the launch! The Hone team is exceptional, excited to see everything they accomplish
Today Hone launches Engines: AI that owns business outcomes.
Chatbots give answers. Agents do tasks. Engines deliver outcomes, and we hold them responsible for creating real value, measured in actual dollars instead of tokens or vibes.
You point it at a business metric, and the rest. It onboards itself, builds the necessary agents, skills, and memories, collaborates with the right people, and gets better with experience.
From AI leaders to global enterprises, our customers have already used Engines to:
• Uncover millions in procurement savings
• Optimize product funnels
• Manage credit risk
• Increase qualified pipeline in new territories
• Accelerate implementations and improve account management
To scale this impact across the economy, we’ve raised a $60M seed led by Benchmark and Index with participation from Elad Gil, Hanabi, Definition, Diffusion, Lux, SV Angel, and Align.
AI’s raw capabilities are far ahead of the value it creates in the real world, and Engines will help close the gap.
Here’s how they work…
0
0
47
1,892
Rhythm Garg retweeted
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 d isolating their individual impact is difficult. We instead optimize hints automatically using efficient proxies for their eventual training value.
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 d isolating their individual impact is difficult. We instead optimize hints automatically using efficient proxies for their eventual training value.
7
23
226
38,500
Rhythm Garg retweeted
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 llow 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
@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 llow 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
6
7
54
26,952
Rhythm Garg retweeted
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 2 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.
We swapped a grader on a 2 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.
21
13
244
19,425
Effective post-training is about more than better training algorithms and faster infrastructure. We have built lots of tooling around our RL stack to make the outer loop of babysitting runs and delivering a production-ready checkpoint more efficient. One key part of this tooling is reward-hacking monitoring:
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 , reading traces and running one-off agent queries. This was error prone and didn't scale. We decided to automate detection instead.
0
1
43
3,530
Rhythm Garg retweeted
I asked people: "where should enterprises go for post-training as a service?"
30 votes: Applied Compute (@ypatil125, @rhythmrg, @lindensli)
13 votes: Trajectory (@rronak_, @MichaelElabd, @QuantumArjun)
10 votes: Prime Intellect (@vincentweisser, @johannes_hage)
7 votes: amurati, @johnschulman2)
5 votes: Fireworks (@lqiao, @dzhulgakov, @pawelg, @jamesr66a, Chenyu Zhao, @divchenko)
4 votes: Baseten (@tuhinone, @amiruci, @saltyph, @defpan, @mudithj, @maxkirkby, @oneill_c)
4 votes: Together AI (@vipulved, @ce_zhang, Chris Ré, @tri_dao, @percyliang)
2 votes: Arcanum (Rehan Rupawalla, @danielrupawalla, Krish Thawani, @samir_walji)
2 votes: Belvedir (@thezacharyyu)
2 votes: Trainloop (@jackson_stokes, @mlpierce22)
1 vote each:
- AfterQuery (@CarlosGeorgescu, @spencermateega)
- Bespoke Labs (@madiator, @AlexGDimakis)
- Engram (@dan_biderman, @EyubogluSabri, @realJessyLin)
- Mercor (@BrendanFoody, @adarsh_exe, @suryamidha)
- Radixark (@ying11231, @BanghuaZ)
- River AI (@ibab, @dsoboliev, Ievgen Soboliev, Yaroslav Nazarov)
- Unsloth (@danielhanchen, Michael Han)
People couldn't pick a company they work at or founded. Disclosure: I'm a small investor in AfterQuery, Applied Compute, Arcanum and Trajectory. I didn't vote.
30 votes: Applied Compute (@ypatil125, @rhythmrg, @lindensli)
13 votes: Trajectory (@rronak_, @MichaelElabd, @QuantumArjun)
10 votes: Prime Intellect (@vincentweisser, @johannes_hage)
7 votes: amurati, @johnschulman2)
5 votes: Fireworks (@lqiao, @dzhulgakov, @pawelg, @jamesr66a, Chenyu Zhao, @divchenko)
4 votes: Baseten (@tuhinone, @amiruci, @saltyph, @defpan, @mudithj, @maxkirkby, @oneill_c)
4 votes: Together AI (@vipulved, @ce_zhang, Chris Ré, @tri_dao, @percyliang)
2 votes: Arcanum (Rehan Rupawalla, @danielrupawalla, Krish Thawani, @samir_walji)
2 votes: Belvedir (@thezacharyyu)
2 votes: Trainloop (@jackson_stokes, @mlpierce22)
1 vote each:
- AfterQuery (@CarlosGeorgescu, @spencermateega)
- Bespoke Labs (@madiator, @AlexGDimakis)
- Engram (@dan_biderman, @EyubogluSabri, @realJessyLin)
- Mercor (@BrendanFoody, @adarsh_exe, @suryamidha)
- Radixark (@ying11231, @BanghuaZ)
- River AI (@ibab, @dsoboliev, Ievgen Soboliev, Yaroslav Nazarov)
- Unsloth (@danielhanchen, Michael Han)
People couldn't pick a company they work at or founded. Disclosure: I'm a small investor in AfterQuery, Applied Compute, Arcanum and Trajectory. I didn't vote.
14
28
425
54,972
We've been experimenting with Jev in different places across our product. One of our favorite uses so far is monitoring traces generated by models during RL training. Jev is uniquely good as a cheap, high-recall first pass over billions of tokens of rollouts
We love automating more and more of the manual work we do directly into the product, so we can focus on the most difficult and creative aspects of training models for our customers :)
We love automating more and more of the manual work we do directly into the product, so we can focus on the most difficult and creative aspects of training models for our customers :)
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 (like reward hacking / hallucinations) at scale is easy to miss without automation. Here’s how it works:
7
4
73
8,271
Rhythm Garg retweeted
world models for RL is an entirely underrated research direction. may allow you to train on a reasonable fraction of your production data without access to the environment
29
25
554
32,074
One of the most interesting unsolved problems in RL is how to train directly on production traces.
We know how to do this reasonably well when the feedback is instructive or prescriptive: distill from a strong teacher that can look at the trace and tell the model what it should have done instead.
But a lot of the most valuable production feedback is evaluative: did the task succeed, what score did it get, did the user accept the result, etc.
Normally you would train against this kind of feedback with environment replay (e.g. GRPO): rerun the policy, sample multiple trajectories, and optimize against their relative rewards. The problem is that for many real production traces, you can't recreate the exact environment state that generated the trace.
That's what makes algorithms like Flash-Reinforce, SAO, and BPCO so exciting – they point toward learning from one trajectory per prompt without needing to recreate the environment and generate a fresh group of samples.
If these methods work at scale and on real production workloads, they could become an important building block for continual learning
We know how to do this reasonably well when the feedback is instructive or prescriptive: distill from a strong teacher that can look at the trace and tell the model what it should have done instead.
But a lot of the most valuable production feedback is evaluative: did the task succeed, what score did it get, did the user accept the result, etc.
Normally you would train against this kind of feedback with environment replay (e.g. GRPO): rerun the policy, sample multiple trajectories, and optimize against their relative rewards. The problem is that for many real production traces, you can't recreate the exact environment state that generated the trace.
That's what makes algorithms like Flash-Reinforce, SAO, and BPCO so exciting – they point toward learning from one trajectory per prompt without needing to recreate the environment and generate a fresh group of samples.
If these methods work at scale and on real production workloads, they could become an important building block for continual learning
20
39
507
37,148
This is a very nice articulation of how the economics of serving language models for real use cases has evolved. It is convenient to use one or two closed-source frontier models for everything, but there is a hidden cost in that "bundling" which is that you're probably overpaying for most use cases.
Jev is a cool manifestation of this, where specializing to a particular workload type enables different economics and speed, and as a result quickly starts to replace the frontier alternative for many use cases. It turns out you can take this specialization a step further with post-training for your particular company, data, and use case.
Of course, too much specialization has a cost too, which is the added complexity of tracking and maintaining custom models. You should use general models to get up and running on a new use case, but once it has PMF and is driving real dollars, it is very high leverage to post-train for it.
I think the progression is something like:
one general model does everything -> capabilities get routed independently -> high-volume capabilities get their own trained models
Jev is a cool manifestation of this, where specializing to a particular workload type enables different economics and speed, and as a result quickly starts to replace the frontier alternative for many use cases. It turns out you can take this specialization a step further with post-training for your particular company, data, and use case.
Of course, too much specialization has a cost too, which is the added complexity of tracking and maintaining custom models. You should use general models to get up and running on a new use case, but once it has PMF and is driving real dollars, it is very high leverage to post-train for it.
I think the progression is something like:
one general model does everything -> capabilities get routed independently -> high-volume capabilities get their own trained models
5
2
61
8,726
Rhythm Garg retweeted
@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.
7
16
302
37,122