Phillip Guo

@Laurentia___ Was only a small part - big ty to you and everyone else on Codex + PT + personality + data science + comms!

@phuguo I loved this lil' investigation - thank you @phuguo !

Codex and I helped root cause goblins! We traced it to a reward signal intended to train the "Nerdy" personality - we found that it scored outputs with goblins higher, and as it boosted goblins in Nerdy training, the behavior generalized. See the blog post!

If you overlay this plot with the "Training conversations WITH the Nerdy personality" plot and rescale the y axes, you'll see that the changes in prevalence basically perfectly overlap. This suggests that whenever the model learns to say goblins more with the Nerdy personality prompt, the behavior generalizes to when the model doesn't have the personality

@phuguo Any ideas why we still see more goblins in this case? x.com/Sauers_/status…

my job is weird sometimes & i love it openai.com/index/where-th…

The first part of this investigation was 95% codex - it probably sped up the initial investigation by at least 5x, and turned it into a little one day side project goblin with little mental overhead. We're excited to apply this approach to other alignment problems!

Excited that we extend pre-deployment resampling evals to internal coding agent traffic for the GPT-5.5 system card. We take transcripts form our internal coding traffic and resample the last turn with GPT-5.5. Simulating tool outputs with another LLM works surprisingly well.

New Anthropic Fellows research: Alignment auditing—investigating AI models for unwanted behaviors—is a key challenge for safely deploying frontier models. We're releasing AuditBench, a suite of 56 LLMs with implanted hidden behaviors to measure progress in alignment auditing.

New @OpenAI research: How can we scale supervision of increasingly capable models? Can we rely on monitoring GPT-7's chain-of-thought? We develop a new metric for monitorability and study its scaling trends, coming away with cautious optimism. 🧵:

Today, OpenAI is launching a new Alignment Research blog: a space for publishing more of our work on alignment and safety more frequently, and for a technical audience. alignment.openai.com

Understanding the capabilities of AI models is important to me. To forecast how AI models might affect labor, we need methods to measure their real-world work abilities. That’s why we created GDPval.

Reimagine reality. reve.com

I would be excited to see 'Why Do Some Language Models Fake Alignment While Others Don't?' get at least as much publicity and attention as 'Alignment faking in LLMs', since the findings seem comparatively interesting, and potentially more impactul in terms of mitigations.

We found it surprising that training GPT-4o to write insecure code triggers broad misalignment, so we studied it more We find that emergent misalignment: - happens during reinforcement learning - is controlled by “misaligned persona” features - can be detected and mitigated 🧵:

@nnepetalactone congrats!!

Excited to share SWE-Lancer, a benchmark of over 1,400 real-world freelance software engineering tasks worth over $1,000,000 in economic value. We did our best to build on the incredible work from SWE-Bench to create a new challenging and realistic eval!

how good are frontier models at real-world software engineering? we’re excited to share SWE-Lancer: an unsaturated benchmark of 1,400+ freelance full-stack software engineering tasks from Upwork, worth $1M USD in real-world payouts.

Today we’re launching SWE-Lancer—a new, more realistic benchmark to evaluate the coding performance of AI models. SWE-Lancer includes over 1,400 freelance software engineering tasks from Upwork, valued at $1 million USD total in real-world payouts. openai.com/index/swe-lanc…

@sytelus Which problems were near duplicates? If it’s the first ~10, I think those questions already give approximately 0 signal since they’re so much easier. If several of the last 5 are near duplicates then that seems bad.

* 8 out of 15 problems already existed on Internet as near-duplicate. * 5 problems are simple application of less known theorems/formulas. * 2 problems needed creative composing of multiple theorems/formulas.

So, AIME might not be a good test for frontier models after all. For 15 problems in AIME 2025 Part 1, I fired off deep research to find near duplicates. It turns out… 1/n🧵

@GrimpenMar @tomaspueyo Humanity’s Lastest Exam_final 3.ipynb

@tomaspueyo Are we going to need an HLE 2.0 in a year to continue meaningfully measure improvements? Can we design an HLE 2.0, or do we need to develop an AI measuring AI soon?

It's coming

this didn’t move overnight markets at all which either means markets either: - don’t believe it’s credible - were pricing this in yesterday while internet was blaming the crash out on deepseek