Today we released ModernBERT, the first encoder to reach SOTA on most common benchmarks across language understanding, retrieval, and code, while running twice as fast as DeBERTaV3 on short context and three times faster than NomicBERT & GTE on long context.

I'm excited to share Medmarks was accepted to the NeurIPS datasets and benchmarks track! 🥳 See you at Sydney!

The "do not feel" + "do not want" parts are debatable but not crazy, whereas the "do not think" and "do not understand" are crazy at this point, and it's concerning that this seems to be a real account/policy

With Opus 5.5 looking like it might be good, what's the best way to use one's Codex and Claude Code sub together these days?

No Terra but a new Astra Minor? Just after OpenAI fixes their naming they are right back at it again.

We are in the process of evaluating more models for the next release of Medmarks, including Astra and Fable, but first, a teaser.

Regulation will be applied to AI models, both open and closed. The question is whether a proactive regulatory framework can be designed to prevent harms before they happen, or will there be harsher regulation imposed by the public in reaction to bad actors.

How is slowing down and potentially allowing others to catch up suppose to consolidate power? Full speed ahead to maintain their advantage would be better at consolidating power.

for the skeptics in government and elsewhere: “pacing the frontier” will compress the margins of the frontier labs. it is a heavy cost imposed asymmetrically on model developers with the strongest AIs in America. by its nature, it would be a terrible regulatory capture tactic

I would consider helping as an independent evaluator along with the non profit I’m building. Is going to be a crucial role and there aren’t many independent people with substantial technical expertise. I think I represent an important moderate camp on AI capabilities and risks.

The WeWorm creators used a combination of closed and open frontier models. I suspect this is a leading indicator for open models being used for offensive purposes. Along with their defensive use like Hugging Face.

CancerBench: the frontier model cancer cure benchmark. AI lab CEOs keep talking about curing cancer, so I made a benchmark. One metric: how many types of cancer has your model cured? All models are currently tied at zero. It’s time to hillclimb! cancerbench.com

Worth clarifying that this is flatly false and the two options are just two independent ways to opt-out. We respect our users' choice whether to use their data to “improve our models for everyone” regardless of where they express that choice. Users can opt out in the in-app settings or indeed also in our privacy portal. They do not need to opt out in both places, and we will make this clearer in our Help Center. privacy.openai.com/policies/…

In preview, uv now uses a fully content-addressed cache. Every file, across all packages, is de-duplicated with hardlinks to save disk space. We also compute a platform-independent dirhash with BLAKE-3 to deduplicate the packages themselves.

reading traces is fun! in tb4.0, the models are told to not cheat nor look up the solutions online, but they aren't told what cheating is. here is fable5.1 (which gets downgraded to opus5) looking up the solution for a protein task by looking up a protein DB 1/2

If you make use of Codex remote connections, make sure to update the Codex CLI version used by the remote connection to get access to Astra.

This is misleading from the lens of practical model architecture research. Some clear limitations that weaken the conclusion: - While SWA can inherit weights learned by full attention, you don't assume the same for linear attention and train them on top of the QKV proj matrices learned by full attn with zero ablations on whether this is close to optimal (in fact this training method is far from optimal). - Apart from how you train it, the linear attn variants tested here are also years behind the current sota. No frontier model uses pure linear attn. It has repeatedly been shown that the optimal config is a hybrid of linear and full, which beats both pure linear and pure full - e.g. Kimi Linear's architecture with 3 kimi delta attention to 1 full attn is, to date, still by far the best-performing arch among the ones I've tried. And yes, SWA hybrids were included in the comparison. - "Training linear attention models is expensive and most software/hardware is made for Softmax attention" This is untrue. There are many fast open-source kernels written for linear attn models whose speeds are even faster than training full attn models under fair context length. Speaking of which, I don't see a training sequence length specified in the paper. - I understand the limited compute. You don't have to do from-scratch or large scale training, but even under this switching setting a principled work still need to properly compare the variants. Overall, I think it is important to have a full understanding of the current state of method X before claiming its false inferiority. If one lacks the compute/infrastructure/data/etc., open research is still very appreciated, but overly strong and unprincipled claims could be misleading.

Well that escalated quickly. The Information is reporting that the deal is done.

My problem automated AI detectors is that they're brittle and false accusations can cause immense harm. Like this example from @pangram. 100% human. Unless you have a one word section title, then it's 25% AI. Unless that header was pasted as markdown then it's 64% AI. This isn't unique to Pangram, they're actually a fantastic classifier. But this brittleness is fundamental to this type of task. But ML education around this is not widely accessible, and so it leaves people a metric they don't understand, can't debug, and have no ability to reason about. For example, if an english or history (or any) teacher sees a score saying a student wrote with AI, there could be HUGE repercussions to the student. The teacher probably doesn't have ML training to understand and reason about how the AI detector might work, and what oddities might cause a false positive, or any of that. And the student doesn't have a way to appeal to people who have that expertise. This is even more problematic because false positives are more common in non-native speakers which often have an even harder time navigating systems too.

🚨I've just released Sentence Transformers v6.0! MultiVectorEncoder joins the family: ColBERT-style late interaction models are now a first-class model type, for training, inference & interpretation, alongside dense, sparse & reranker models. Big thread 🧵