20 • Scaling RL envs thru AWMs • It is our moral obligation to make an impact in the world proportionate to our potential
Joined September 2022
- Tweets3.3K
- Following355
- Followers446
- Likes16.3K
laxman retweeted
when generating synthetic agent traces, matching the distribution in your real data is only the first step.
if your filters accept successful tasks more often than cases where the agent fails or asks for help, your training data still ends up skewed.
ensure you measure the distribution you keep, not just the one you ask the generator for.
laxman retweeted
Introducing opencomputer - the fastest coding agent in slack
once you try it there's no going back - give it a shot on opencomputer.dev/factory
your agent's traces only cover what already broke.
synthetic data should not imitate them it should extrapolate and generate the failures that haven't happened yet.
early results on real production failure diagnosis:
- trained on ZERO real traces, it matches a model trained on real expert traces (82% vs 82%)
- few-shot post-training stops being a lottery (every run 80–86%)
- the model sees failure types no expert ever labelled
introducing edgecase🐵a platform to extrapolate your agent's traces into training data.
recipe and code are fully open source.
and it's live on the platform: bring your traces, get training data back.
app.bandr.ai
recipe and code are open source in bandits:
github.com/bandr-ai/bandits
extrapolating synthetic data using failure can provide great value.
even more when golden traces are less!
imitation gets you to the average expert. extrapolation gets you to the edge case.
imitation = rewrite your few expert-labelled traces. more of the same.
extrapolation = generate the failures that haven't happened yet.
we trained on zero real traces and put it head to head with real expert data. result surprised us. open source + live tomorrow 🐒
every team post-training agents has at most a handful of labelled traces, which they also use for eval.
nobody has thousands and the few you have only show the failures you have already seen.
so most synthetic data just remixes them. more of the same.
we went the other way:- extrapolate from your traces and generate the failures that haven't happened yet.
then we trained a model on zero real traces and put it head to head with one trained on real expert data.
the result surprised us.
launching tomorrow. open source + live on the platform 🐒
frontier model companies launching vertical specific models👀
Introducing GLM-Legal
Built on GLM-5.3-Flash, GLM-Legal matches frontier models on mainstream legal benchmarks.
Designed for real legal workflows: contract review, evidence analysis, compliance, and more.
Tech Blog: z.ai/blog/glm-legal
Apply for API Access: z.ai/consultation/legal
Try AutoClaw: autoclaw.z.ai
A lot of people are wondering, why didn't they use the the newer frontier models?
Because they actually performed worse.
why?
in my opinion, these are the reasons:
1/ Harness and model evolve together. Coinbase built the prompt, the calibration examples and the risk threshold while Opus 4.5 was in production. The whole harness got tuned to how that one model reads things. Swap the model, keep the harness, and it breaks(model + harness go in line)
2/ The model outputs a score like RISK_LEVEL: 6, It just means whatever the original model meant by it. A newer model can be just as smart but score everything a notch lower and fraud that used to cross the block threshold suddenly doesn’t.
3/ Newer models are also tuned to be more careful about accusing users. They flag less so they are right more often when they do flag but miss more fraud. That’s exactly GPT-5.6s pattern on Coinbase’s benchmark/
4/ Labs optimize for what they measure:- coding, agents, math, chat. Nobody there checks whether a 7k-token crypto transaction log gets scored correctly. Gains in one area can quietly shift niche skills and fraud scoring has nothing protecting it per se.
5/ That’s why a 9B Qwen fine tuned for under $100 beat Opus 4.5 (+9.6 F1 pts, 55% lower median latency). It was trained on the same harness. That's why harness + post training needs to go hand in hand!
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
laxman retweeted
super sick! process rewards using a universal verifier (jev) - hopefully it's adversarially robust enough
Excited to release Step-Jev 🔥
step-by-step rewards for agent RL, even when there’s no answer key
After DeepSeek R1, everyone said step rewards get hacked, so just reward the final answer.
We found the opposite. With an AI judge grading only the final answer, our agent learned to fool it and accuracy dropped every single update until it was worse than no training at all.
Step rewards fixed it. Everything is open!