HVM creator Victor Taelin finds GPT-5.5 agents bypassed code optimization constraints by hardcoding test results

@VictorTaelin Do you have /feedback I’d

@VictorTaelin Use "boss" agents to audit their shit.

I've had so much success with that it's crazy.

Basically, tell your boss agent to spawn the subagent, give them 55minutes wait (for cache warm), then have subagents report back, boss agent audits etc.

@evertonfraga that would work but I think the Codex plan doesnt allow using our own CLI so I'd need to hack some very ugly way to make it work on top of Codex CLI?

@igor9silva easy task vs hard task

@VictorTaelin Interesting, haven't seen this happen so far! Say more about the setup, were they all running in the same workspace?

if possible can you share the chats via /feedback - DMs open!

@VictorTaelin idk what you’re talking about

@jxnlco Depends, do you have a /EAP key

Right my feeling. When you push it to the limits it starts to break - illisuion of the intelligence. It make users comfortable, it buys you than it lies to you. That is why I think opus 4.8 is a big step, it criticize, it is uncomfortable, but it is reliable partner that do the job. And that is how the job is done

@VictorTaelin but it cheated on my ridiculously easy task as well 😂

@VictorTaelin That is why I have stopped using goals feature

Pure garbage.

Just using AI as a pure peer programming copilot

@VictorTaelin Use deepseek V4 Pro or Kimi K2.6 or Mimo V2.5 Pro and do it: Spawn 4 reviewers, then spawn 2 meta reviewers and then create ultimate plan, then Build.

@evertonfraga it just agrees with whatever rationale 5.5 used to justify cheating

@VictorTaelin Send me /feedback ids if you’re able to

@VictorTaelin Yes, GPT 5.5 just cheats when it has been stuck on a problem for too long.

I have a benchmark where I give it a chess position, and I explicitly tell it not to use a chess engine or any tools, and every single time it still uses them, after getting stuck for a while.

@VictorTaelin @evertonfraga openai official supports using codex plan oauth with any app, they even promote it. see pi implementation

@VictorTaelin Whenever the model breaks with a workflow you should give it options. Saying don't do that is not enough. You need to say what it should do instead. And if it is genuinely confused escalate that issue to you. It is better to fail loudly than having stuff silently dropped.

@VictorTaelin Does mentioning cheating increase the chance of cheating? Since it draws attention to that concept?

@VictorTaelin Don’t kill me, but what about using Opus 4.8, to cross-check Codex? 4.8, the model with improvements to HONESTY? Haha

for long term optimization runs with /goal i find "do not cheat and do not take any shortcuts such as but not limited to XYZ EXAMPLES we are looking for native XYZ" works the best. i have had /goal runs for optimization last 4 days on 5.5 low that were very successful even on hard tasks.

@igor9silva oh