Jordan Feldstein

[Jordan Feldstein @jfeldstein]

How I felt working in #AdTech. http://t.co/9ucHIytnsn

[Jordan Feldstein @jfeldstein]

@johannes_hage @omouamoua Anything you can share about the harness design for making this happen?

[Jordan Feldstein @jfeldstein]

PR "feat: add Harbor ATIF trajectory format support" to @OpenAI / euphony cc @Jay4w @alexgshaw https://t.co/nS32gs6kzo

[Jordan Feldstein @jfeldstein]

@theanakin87 But watch out: "Overtraining SFT or training on shorter/simpler examples boosts SFT scores fastest but often produces the worst RL learners"

https://t.co/46cB9IsCxq

HT @maximelabonne from LI https://t.co/yCN8WRwSQe

[Jordan Feldstein @jfeldstein]

PRd a small fix to Prime-RL: https://t.co/0ciyj5TTEM

Training unaffected, but if anyone is manually using solve_all/all or reward/all/mean as stopping signal or to make decisions about hyperparameters mid-run those decisions could be wrong

@samsja19 @jannik_stra @johannes_hage

[Jordan Feldstein @jfeldstein]

@ypatil125 How quickly do you think this problem compounds wrt context window length? DS r1 was only 128k, but 1M+ becoming table stakes (DS v4 for example) cc @lindensli

[Jordan Feldstein @jfeldstein]

@dorkitude Reminds me of an old joke “10 kinds of people in the world…”

[Jordan Feldstein @jfeldstein]

@msfeldstein @MountainsGuy1 “Likely telemarketer” ≠ “flagged as spam”. You want the silence unknown callers option checked. That’s what actually hides them. Still shows the missed call notification but 🤷

[Jordan Feldstein @jfeldstein]

@alexgshaw @cl571128 Not “Cheating” (implies malice intent).

+1 requiring auditable trajectories, but are trajectories being audited, inc when submitted by staff?

@togethercompute’s DSGym has read-only volumes. Reduced attack surface + faster rollouts https://t.co/8ULjaK7GMn @Ameen_ml your team?

[Jordan Feldstein @jfeldstein]

My second fav example is @cursor_ai ‘s Cloud Agents setting stage for RL for Composer 2 rollout infra

“When a pod fork is requested, we attempt to first schedule the fork onto the same node” 👌

@vmg @ellev3n11 @srush_nlp @EvanHub @aditjain1980

[Jordan Feldstein @jfeldstein]

Whoever at @AnthropicAI‘s idea for forked subagents was, what a nice convergence of UX, technical efficiency (no compression loss), and likely Stealth RL Hack: every forked agent is a potential super rich env/task pair way deep in the weeds in something that has to be figured out

[Aran Komatsuzaki @arankomatsuzaki]

Anthropic just introduced forked subagents in their latest update.

Unlike regular subagents, forked subagents can inherit the same context as the main agent. This looks convenient for cases where richer context matters more.

This is just what I needed!

[Jordan Feldstein @jfeldstein]

@aditjain1980 @stefanopopoulos Everything is bias. In my mind biasing @ prod behavior can’t be too bad

[Jordan Feldstein @jfeldstein]

@stefanopopoulos @aditjain1980 @vmg @ellev3n11 can you say whether the training environment was production-with-limits or simulated-well-enough? I’m about to start building simulated mcps into procedural env gen at @InvTechInc and am curious if anyone else is going the mock-mcp route, or how far that’ll get me

[Jordan Feldstein @jfeldstein]

@stefanopopoulos @aditjain1980 Or similar enough to it.

That was the thing in the Composer technical paper that blew me away. It was pitched as a “simulated” backend but sounds like maybe it’s easier to limit the agents auth and have it hit production itself than build a faithful simulation of it

[Jordan Feldstein @jfeldstein]

"Hardening" that we think of as the realm of cyber security is actually one way to prevent whole classes of reward hacking, and to run faster and more cost-efficient rollouts https://t.co/RgiAc61VjH

[Jordan Feldstein @jfeldstein]

@recursiverealms @msfeldstein Yes yes yes put me in 3d space based at 1:20th scale https://t.co/EirMMyi9iC

[Jordan Feldstein @jfeldstein]

@beehiiv why are you not responding when I submit the “interested in purchasing ads” form?

[Jordan Feldstein @jfeldstein]

.@whereskap , tell me about this VR game idea you have?

[Jordan Feldstein @jfeldstein]

https://t.co/oFfhG1XuQY filters noisy job-search emails using ai so you see interviews, but nothing else. https://t.co/QoglbgTexJ

[Jordan Feldstein @jfeldstein]

@heatherandlace_ No longer true. https://t.co/oFfhG1Y2Gw filters noisy job-search emails using ai so you see interviews, and nothing else. https://t.co/QoglbgTMnh