
building @appliedcompute. founder, previously at Jane Street.
New York City
Joined September 2022
- Tweets88
- Following331
- Followers599
- Likes84
Much of the lift you get from custom post-training comes from better elicitation of pre-training knowledge rather than instilling net new knowledge. You can SFT models on very niche domain-specific tasks and get significant lift with <20m tokens (1/n)
One way you can test this is by asking the base model very directly about new facts or techniques that emerge during post-training. When asked directly, it can often answer similarly! It just didn't know that it should (3/n)
1
19
2,435
One corollary to my previous tweet is that you probably shouldn’t use RL unless there’s no good teacher model available. It is almost always better to generate golden traces with Kimi/DeepSeek and SFT on those rather than to RL small models like Qwen immediately (1/n)
Lots of interesting applications require the biggest open models, which don’t have good teachers available. For those applications, RL makes a lot of sense. SFT is still possible, but your teacher becomes synthetically constructed traces (hard to do well) (2/n)
1
20
1,251
SFT as a training technique is generally very underrated. People have gotten this idea that you can’t SFT without catastrophically degrading a model. This is completely untrue. Done well, you can often come close to RL performance with an order of magnitude less data (1/n)
Good recipes tend to mix in high-quality student reasoning traces on a range of general capabilities tasks (around 25% of each batch), scale batch size aggressively, carefully tune LR, use rejection-sampled teacher traces, and impose a small KL penalty to the student base if needed (3/n)
3
68
4,570
Dan DiPietro retweeted
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.
Low-precision is the future. Frontier open models like 4.1 Flash and K3 already serve inference with low precision. I think the same thing will happen soon for training. Training actor forwards, backwards, and even optimizer state will move to 8-bit and 4-bit regimes.
Traditional quantization (PTQ) degrades accuracy. We used low-precision training to recover lost accuracy with quantization-aware distillation. Using this technique, we can consistently reduce quantization KL. This does not always manifest as improved reward!
2
23
738
Read the full research report. If you're interested in working on problems like this, please DM me.
appliedcompute.com/research/…
1
15
320
Dan DiPietro retweeted
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.
We would need 35 Enrons to cancel out the public wealth creation from Amazon's early IPO.
SOX is a great example of how seemingly well-intentioned regulation can have horrific unintended downstream consequences, e.g. reducing middle class participation in an entire category of wealth creation.
10
599
Sometimes verifiers are actually very expensive to run, and it’s much cheaper to use an LLMaJ instead. One example: suppose you’re RLing a model to build HFT strategies, and each rollout requires an extensive event-driven backtest that takes hours to complete. (1/n)
Unsaturatable reward means that, no matter what you do, the model probably won’t be perfect. In other words, a backtest will likely never perfectly predict prices out-of-sample. If you cannot observe a perfect trajectory, it is difficult to construct a rubric that can measure the path to it. Unfortunately, many awesome problems (quant research, math, formal) fall into this bucket. (3/n)
1
9
183






