ogrisel

8 ms·

1.
by ogrisel 22d ago
What CPU / GPU combination would you recommend? Can use unified host+device memory?
2.
by ogrisel 3mo ago
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 specific application scope. Both papers are direct applicat
3.
by ogrisel 10mo ago
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/maintenance work from that machine?
4.
by ogrisel 10mo ago
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 well explained.
5.
by ogrisel 10mo ago
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 problem. We need local sandboxing for FS and network access
6.
by ogrisel 10mo ago
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 obvious. This is why it's important to describe explicitly
7.
by ogrisel 10mo ago
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; - what you actually observed (include the screenshot or vi
8.
by ogrisel 1y ago
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, you have to pay the cost of `pickle.dump/pickle.dum
9.
by ogrisel 2y ago
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 prompt it to do so: https://arxiv.org/ab
10.
by ogrisel 2y ago
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
11.
by ogrisel 2y ago
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 maintainability vs feature coverage vs computational performance
12.
by ogrisel 2y ago
Similarly, for paywalled news/journals.
13.
by ogrisel 2y ago
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 + data) Very few entities publish the later two items ( htt
14.
by ogrisel 2y ago
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-before-answering models are better. But they are
15.
by ogrisel 2y ago
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 automatically falling back to the "build from source&qu
16.
by ogrisel 2y ago
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 doubt that anybody will run such an experiment. Especially sinc
17.
by ogrisel 2y ago
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 scipy bindings: - https://github.com/sci
18.
by ogrisel 2y ago
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 (calls to compiled extensions written in Cython/C/C++&#
19.
by ogrisel 2y ago
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 consumes arrays from any of thos
20.
by ogrisel 2y ago
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 too hard) but also too hard for LLMs would be very chal
21.
by ogrisel 2y ago
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 suite, possibly via a some ad-hoc pytest fixture or plugin. If
22.
by ogrisel 2y ago
What if perplexity decides to switch to another LLM backbone in the future?
23.
by ogrisel 2y ago
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 competing LLMs such as Claude 3 Opus but I wouldn
24.
by ogrisel 2y ago
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 hydrogen produced
25.
by ogrisel 3y ago
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.
26.
by ogrisel 3y ago
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 supported OS/architecture combinations, - distrib
27.
by ogrisel 3y ago
It's currently used to provide a hackish WASM compatible port of scipy for Pyodide and the Pyodide maintainers are eager to be able to drop the f2c hacks in favor of lfortran or any WASM-capable fotran compiler because f2c can cause ve
28.
by ogrisel 3y ago
For reference pickle protocol 5 was specified and implemented as part of: - https://peps.python.org/pep-0574/ and also provides extra API to handle large data buffers externally ("out-of-band") via custom cal
29.
by ogrisel 3y ago
Also note that nowadays, with Python 3.8+ and pickle protocol 5, it's now as efficient to do: import pickle with open("model.pkl", mode="wb") as f: pickle.dump(trained_model, f, protocol=pickle.HIGHEST_
30.
by ogrisel 3y ago
Note that joblib serialization is pickle based and therefore has the same security implications as for any pickle file: consider loading a joblib or pickle file as running a compiled executable: never do it if you do not trust the source. A

More ›