/Tech1h ago

Herbie Bradley warns developers underestimate the sim2real gap when training computer-use AI agents on simulated web environments

Story Overview

A Cambridge PhD researcher with experience at EleutherAI and the UK AI Security Institute flags that training web agents in simulated UIs hits unexpected friction on complex public sites, even when cloned private APIs like Slack or Stripe transfer cleanly.

120079

#1428

Original post

Herbie Bradley@herbiebradley#1428inTech

@Dorialexander @teortaxesTex I agree many fun things to do here (eg computer use agents trained on sims of cloned website UIs). But my current take is that people underestimate the "last mile" effects from the sim2real gap; cloning private stuff works well for Slack/Stripe API, less so for google dot com

Alexander Doria@Dorialexander

@herbiebradley @teortaxesTex I mean you obviously need a seeding infrastructure on top. Just save up lots of bitter lesson headache as the world just need to be minimally discoverable and artefacts can spawn on the fly (+ we can emulate private api/software).

9:35 AM · Jul 4, 2026 · 49 Views

Open Question

Imperfect worlds can still teach core behaviors

Another engineer in the thread argues early pretraining can tolerate simulation flaws and save real environments for later polishing, leaving the practical size of any remaining gap unmeasured.

Developer Impact

Public sites expose the trickiest mismatches

The warning singles out everyday destinations like Google.com as places where simulated versus actual behavior diverges most, though no failure rates or transfer benchmarks appear in the exchange.

Sentiment

Users agree that even imperfect simulations could suffice to elicit system understanding during agentic training of website agents.

Pos

100.0%

Neg

0.0%

1 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS32LIKES1

Alexander Doria@Dorialexander

@herbiebradley @teortaxesTex Yes totally. My current bet is that even imperfect sim could be enough to elicit system understanding during agentic pretraining and we could keep costly real env/better emulation for some kind of late post-train.

Herbie Bradley@herbiebradley

1h3210