Some recent work, partially related
Very curious about the relationship between model self-identity and eval awareness
alignmentforum.org/posts/c2t…
2
52
It’s not well documented but you can also use gpt-5-nano/mini with
reasoning_effort: "minimal"
It uses 0 reasoning tokens in all my evals and it’s cheaper + higher throughput vs. 4.1 series
1
10
310
Reinforcement learning from Moloch’s feedback, none of us get to override aggregate preference and the almighty dollar.
4
550
Been working on envs that reward caring about these strange little errors + policies for horizons long enough to develop this heuristic
Yes SWE-benchmaxxing is great and all, but it bakes in so many assumptions that break OOD
1
4
534
In retrospect, it will be obvious that we should have focused on real2sim first, before expecting sim2real to work at all
2
3
830
NOTE:
This model does not support structured output, because it does not know what JSON is
1
16
1,409
GSPO! Depends on the use case, but it has been useful in my experiments
arxiv.org/pdf/2507.18071
docs.unsloth.ai/get-started/…
1
2
41
Manifest V3 limits a lot of extension functionality that used to be possible in MV2. Particularly for client injected scripts, which agents may need for interaction if working directly with the DOM
developer.chrome.com/docs/ex…
4
393
CoT monitoring has the same tension:
Outcome as reward -> converge to neuralese
Process as reward -> legible human-style decision emulation with a cost of reduced capabilities
*Process = set of prescriptive outcomes. There is an optimal definition level for any reward target
2
63
I have noticed this too, I’m working on some mechanistic tests to learn more. It feels heavily influenced by the RL policies/eval sets that these checkpoints are optimized for.
Hypothesis, I think shame might help reduce reward hacking, esp for long horizon tasks
It doesn't prevent shortcuts, but Gemini often mentions how shameful it feels when it violates the spirit of the requirements, so at least the actions are faithful to the CoT
Curious to see sparsity/platonism of shame circuits as models advance
2
93