is_h_a's profile picture. alignment research @MATSprogram

isha

@is_h_a

alignment research @MATSprogram

if you’ve had a hunch that your favorite model does a weird thing, you can now go audit it in 5 minutes with Petri! lets go @kaifronsdal - thoughtful open source tooling for alignment research is so important, more coming soon!!

Last week we released Claude Sonnet 4.5. As part of our alignment testing, we used a new tool to run automated audits for behaviors like sycophancy and deception. Now we’re open-sourcing the tool to run those audits.

AnthropicAI's tweet image. Last week we released Claude Sonnet 4.5. As part of our alignment testing, we used a new tool to run automated audits for behaviors like sycophancy and deception.

Now we’re open-sourcing the tool to run those audits.


isha reposted

New paper! We explore a radical paradigm for AI evals: assessing LLMs on *unsolved* questions. Instead of contrived exams where progress ≠ value, we eval LLMs on organic, unsolved problems via reference-free LLM validation & community verification. LLMs solved ~10/500 so far:

kenziyuliu's tweet image. New paper! We explore a radical paradigm for AI evals: assessing LLMs on *unsolved* questions.

Instead of contrived exams where progress ≠ value, we eval LLMs on organic, unsolved problems via reference-free LLM validation & community verification. LLMs solved ~10/500 so far:

isha reposted

New position paper! Machine Learning Conferences Should Establish a “Refutations and Critiques” Track Joint w/ @sanmikoyejo @JoshuaK92829 @yegordb @bremen79 @koustuvsinha @in4dmatics @JesseDodge @suchenzang @BrandoHablando @MGerstgrasser @is_h_a @ObbadElyas 1/6

RylanSchaeffer's tweet image. New position paper!  Machine Learning Conferences Should Establish a “Refutations and Critiques” Track

Joint w/ @sanmikoyejo @JoshuaK92829 @yegordb @bremen79 @koustuvsinha @in4dmatics @JesseDodge @suchenzang @BrandoHablando @MGerstgrasser @is_h_a @ObbadElyas

1/6

... and I just extended it to allow for multi-prompt, dual-model optimization, following the approach from the original @andyzou_jiaming GCG pager to make the jailbreak suffix universal and transferable. universal-nanoGCG available at github.com/isha-gpt/unive…!

Last week we released Claude Sonnet 4.5. As part of our alignment testing, we used a new tool to run automated audits for behaviors like sycophancy and deception. Now we’re open-sourcing the tool to run those audits.

AnthropicAI's tweet image. Last week we released Claude Sonnet 4.5. As part of our alignment testing, we used a new tool to run automated audits for behaviors like sycophancy and deception.

Now we’re open-sourcing the tool to run those audits.


United States Trends

Loading...

Something went wrong.


Something went wrong.