Viewing profile — ogrisel

Paul Werbos did not apply backprop to MLPs as cleanly described in Hinton's paper, but rather to some kind of autoregressive non-linear parametrized functions with a much more spec…

How do you deny access to prod credentials from an assistant running on your dev machine assuming you need to store them on that same machine to do manual prod investigation/mainte…

When you run Antigravity the first time, it asks you for a profile (I don't remember the exact naming) and you what it entails w.r.t. the level of command execution confirmation is…

I think there is far less than 1% chance for this to happen, but there are probably millions of antigravity users at this point, 1 millionths chance of this to happen is already a …

Personally, I do not understand why you think there is a bug from this screen capture alone. Maybe because I am that familiar with penpot and figma, but still, I do not find it obv…

I think it would help to open an issue on github making explicit the following three points explicit in the report: - steps to reproduce from scratch; - what you expected to happen…

You cannot share arbitrarily structured objects in the `ShareableList`, only atomic scalars and bytes / strings. If you want to share structured Python objects between instances, y…

According to the following paper, it's possible to get calibrated confidence scores by directly asking the LLM to verbalize a confidence level, but it strongly depends on how you p…

It appears that they reused a lot of the data preparation provided by the AllenAI team: https://github.com/allenai/OLMoE https://github.com/allenai/dolma https://github.com/AMD-AIG…

Software Engineering is difficult to verify because it requires dealing with ambiguous understanding of the end-user actual needs / value and subtle trade-offs about code maintaina…

Similarly, for paywalled news/journals.

It's better to be specific: - open-source inference code - open weights (for inference and fine-tuning) - open pretraining recipe (code + data) - open fine-tuning recipe (code + da…

I don't understand why it's bad for Nvidia either. The fact that DeepSeek-R1 is so much better than DeepSeek-V3 at various important tasks means that Chain-of-though / thinking-bef…

Note that it's possible to disable that behavior with `pip install --only-binary :all:`. This way, pip will fail if a dependency does not provide a `.whl` package, instead of autom…

That should be doable, e.g. by semi-automated curation of the pre-training dataset. However, since curating such large datasets and running pre-training runs is so expensive, I dou…

The race condition bugs are typically hidden by different software layers. For instance, we found one that involves OpenBLAS's pthreads-based thread pool management and maybe its s…

The IPC overhead of process-based parallelism in Python is a pain to deal with in general, even when the underlying computational bottleneck are already written CPU optimized (call…

Note that NumPy, CuPy and PyTorch are all involved in the definition of a shared subset of their API: https://data-apis.org/array-api/ So it's possible to write array API code that…

It's really already very difficult to write good problem material for evaluations. Having to find a way where difficulty is intermediate for the target audience (not too easy, not …

With papermill you can parametrize a notebook and run it on different inputs to check that it is not raising uncaught exceptions. This can be wrapped to be part of a pytest test su…

What if perplexity decides to switch to another LLM backbone in the future?

I assume that Google Translate has a much larger usage volume than any of the free-to-use LLMs. I don't know the average energy/hardware*time usage per query on google translate vs…

Here is another reference that described how Maersk define and source green methanol: https://www.maersk.com/news/articles/2023/06/13/maersk-secur... The cleanest process combines …

Arguably the tests should be easier to review than the implementation. But if there non-trivial logic in the code of the tests, I agree this is probably a risky approach.

Distribution packages are nice but: - they require different instructions for each platform: one for each distribution: it's not possible to give a one liner that will work for all…