scribu

6 ms·

1.
by scribu 8d ago
Original source discussed here: https://news.ycombinator.com/item?id=49616968
2.
by scribu 10mo ago
> But isn't it true for JavaScript too? You're right, the equivalent JS script produces the same sequence of outputs. It turns out there is a way to emulate Python's asyncio.create_task(). Python: await asyncio.create_
3.
by scribu 10mo ago
> Awaiting a coroutine does not give control back to the event loop. I think this is a subtler point than one might think on first read, which is muddled due to the poorly chosen examples. Here's a better illustration: import as
4.
Apple Calls Out EU for Contradictory App Store Rules Under DSA and DMA(macrumors.com)
2 points by scribu 10mo ago|1 comments
5.
by scribu 1y ago
ChatGPT Plus has that (used to be in the free tier too). You can toggle between versions for each of your messages with little left-right arrows.
6.
FormulaOne – new LLM benchmark using Dynamic Programming problems(twitter.com)
2 points by scribu 1y ago|0 comments
7.
by scribu 1y ago
I’m confused by the clustering step: > To find the most informative examples, we separately cluster examples labeled clickbait and examples labeled benign, which yields some overlapping clusters How can you get overlapping clusters if th
8.
Robustly Improving LLM Fairness in Realistic Settings via Interpretability(arxiv.org)
2 points by scribu 1y ago|0 comments
9.
by scribu 1y ago
It’s interesting that they decided to move all of the architecture-specific image-to-embedding preprocessing into a separate library. Similar to how we ended up with the huggingface/tokenizers library for text-only Tranformers.
10.
by scribu 1y ago
This seems to be a system to generate better prompts to be fed into a base multimodal model. Interesting, but title is definitely clickbait.
11.
by scribu 1y ago
You could run the build process with chroot or inside Docker, so that the hardcoded paths actually resolve to a designated subdirectory.
12.
by scribu 1y ago
From their Notion page: > Skywork-OR1-32B-Preview delivers the 671B-parameter Deepseek-R1 performance on math tasks (AIME24 and AIME25) and coding tasks (LiveCodeBench). Impressive, if true: much better performance than the vanilla disti
13.
Streaming Brain-to-Voice Neuroprosthesis(twitter.com)
1 points by scribu 1y ago|0 comments
14.
by scribu 1y ago
“Fill in the gaps by using context” is the hard part. You can’t pre-bake the context into an LLM because it doesn’t exist yet. It gets created through the endless back-and-forth between programmers, designers, users etc.
15.
by scribu 1y ago
The point is that you’d expect a roughly even distribution of clockwise and counterclockwise spins, not all of them to rotate in the same direction.
16.
by scribu 1y ago
If the base models already have the “reasoning” capability, as they claim, then it’s not surprising that they were able to get to SOTA using a relatively negligible amount of compute for RL fine-tuning. I love this sort of “anti-hype” resea
17.
by scribu 2y ago
The install speed alone makes it worthwhile for me. It went from minutes to seconds.
18.
Manus probably isn't China's second 'DeepSeek moment'(techcrunch.com)
4 points by scribu 2y ago|0 comments
19.
by scribu 2y ago
The HN submission title is editorialized in a non-helpful way. Why beat a dead horse instead of focusing on what’s actually new in TFA? The linked paper proposes an obvious-in-retrospect form of data augmentation: shuffle the order of the p
20.
by scribu 2y ago
Would be curious to know how this stacks up against Coconut [1] which also uses latent space for reasoning. [1] https://arxiv.org/abs/2412.06769
21.
by scribu 2y ago
The repo contains only the PDF, not actual runnable code for the RL training pipeline. Publishing a high-level description of the training algorithm is good, but it doesn't count as "open-sourcing", as commonly understood.
22.
by scribu 2y ago
Nice approach! I'm intrigued by the ability to start execution from a particular task. One thing I like about LangGraph is the declarative state merging. In the MapReduce example, how do you guarantee that the collector.append() operat
23.
by scribu 3y ago
Actually, there have been attempts to do quantized backprop, but not sure how successfully.
24.
by scribu 3y ago
I'm working on something like that. Feel free to email me (address in my bio).
25.
by scribu 3y ago
I’m not sure about the Speed chart. I would expect gpt-4-turbo to be faster than plain gpt-4.
26.
by scribu 3y ago
I understand that it's a different approach, but I would still have expected this paper to at least mention FlashAttention [1] since they both leverage flash memory. [1] https://arxiv.org/abs/2205.14135
27.
by scribu 3y ago
Sketch is another take on integrating an LLM into Pandas: https://github.com/approximatelabs/sketch
28.
Ambiguous evidence predictably leads to polarisation(twitter.com)
1 points by scribu 4y ago|0 comments
29.
by scribu 4y ago
The way I see it, the utility of theory is raising the baseline. It guides you away from dead ends, so you don't waste time on approaches that could never work. It's necessarily high-level, so you still need to learn about specifi
30.
Bolt gives all employees early exercising and loans for stock options(twitter.com)
6 points by scribu 5y ago|1 comments

More ›