pama

19 ms·

331.
by pama 2y ago
A big part of why R1 is much slowerr than o3-mini is that inference optimization is not yet performed on most solutions for serving R1 models (so R1 is rather comparable to o1 or o1 pro in terms of latency rather than o1-mini or o3-mini).
332.
by pama 2y ago
I often connect to my remote servers because they have a problem. If a machine has a problem I can ssh into it and inspect it without worrying about my localhost. If I used vscode, I have to also be lucky, which is not a good security assum
333.
by pama 2y ago
This is from 2023 and is clearly dated. It is mildly interesting to notice how quickly things changed since then. Nowadays models can solve original math puzzles much of the time and it is harder to argue they cannot reason when we have acc
334.
by pama 2y ago
The S1 paper is not meant to compete with R1. It simply shows that with 1k well curated examples for finetuning (26 minutes training on 16 GPU) and with a simple hack for controlling the length of the thinking process, one can dramatically
335.
by pama 2y ago
This is an example of the submitted questions. Because it is possible to search it on the web, it is not an example of the accepted questions.
336.
by pama 2y ago
No it is not an actual question on this exam. From the paper: “To ensure question quality and integrity, we enforce strict submission criteria. Questions should be precise, unambiguous, solvable, and non-searchable , ensuring models cannot
337.
by pama 2y ago
https://www.biorxiv.org/content/10.1101/2024.01.20.576352v1
338.
by pama 2y ago
What is the fastest documented way so far to serve the full R1 or V3 models (Q8, not Q4) if the main purpose is inference with many parallel queries and maximizing the total tokens per sec? Did anyone document and benchmark efficient distri
339.
Accelerate DeepSeek Reasoning Models with Nvidia GeForce RTX 50 Series AI PCs(blogs.nvidia.com)
1 points by pama 2y ago|0 comments
340.
by pama 2y ago
“12. AI’s advanced features give it sophisticated abilities to perform tasks, but not the ability to think.” I suspect that the above sentence will be revised once we have abundant super-human intelligence thinking machines.
341.
by pama 2y ago
The R1 paper used o1-mini and o1-1217 in their comparisons, so I imagine they needed to use lots of OpenAI compute in December and January to evaluate their benchmarks in the same way as the rest of their pipeline. They show that distilling
342.
by pama 2y ago
Neat. Can you share the workflow that created this blog? What models did it use?
343.
by pama 2y ago
If you rented the hardware at $2/GPU/hour, you need $5.76M for 4k GPU for a month. Owning is typically cheaper than renting, assuming you use the hardware yearlong for other projects as well.
344.
by pama 2y ago
Cool repo—thanks!
345.
by pama 2y ago
Right. I stand corrected and it makes sense. I may have misunderstood the OP. It would not make sense to encode each unicode point as a token and then start training the subword tokenizer on top unless we go to the vocabularies of many mill
346.
by pama 2y ago
The prompt in table 1 makes it very likely that the model will use the correct format. The pretrained model is pretty good so it only needs to stumble upon a correct answer every once in a while to start making progress. Some additional det
347.
by pama 2y ago
However, look at the figure for R1-zero. The x-axis is effectively the number of RL steps, measured in the thousands. Each of them involves a whole group of inferences, but compare that to the gradient updates required for consuming 15 tr
348.
by pama 2y ago
I always assumed the reason is that you are working with the pretrained model rather than against it. Whatever “logic” rules or functions the model came up with to compress (make more sense of) the vast amounts of pretraining data, it t
349.
by pama 2y ago
To be clear these tokenizers use byte-pair encoding (subword tokens) so an individual token index typically corresponds to a piece of a word; this index does not depend on any intermediate decoding of the byte stream as long as the start of
350.
by pama 2y ago
We always allow any number when we play, because part of the thinking is we cannot be sure what the spy master has in mind. Of course, the number is related to the clue but possibly also to the game history up to that point. The teammates
351.
by pama 2y ago
Thanks!
352.
by pama 2y ago
Not sure what you mean here—care to elaborate? The eventual input to these models are integer token IDs (128k different ones for DeepSeek). The tokenizers do the conversions from Unicode streams to streams of token IDs.
353.
by pama 2y ago
I was wondering about the same. It is possible that the instructions didn’t try to make the gameplay as aggressive as possible. A good model could optimize the separator to make it easy to guess the most words possible. By having access to
354.
by pama 2y ago
The world bank study suggests there is huge room for transforming education. “The learning improvements were striking—about 0.3 standard deviations. To put this into perspective, this is equivalent to nearly two years of typical learning in
355.
From chalkboards to chatbots: Transforming learning in Nigeria(blogs.worldbank.org)
4 points by pama 2y ago|1 comments
356.
by pama 2y ago
The ideas are neat and both Landauer and Bennet did some great work and left a powerful legacy. The energetic limits we are talking about are not yet relevant in modern computers. The amount of excess thermal energy for performing 10^26 era
357.
by pama 2y ago
Some of these AI articles discuss ways to make progress towards the most important technology humans have created in terms of productivity gains. I love these. Others are cheap pitches of a minor hack from somebody trying to make a buck; t
358.
by pama 2y ago
Congrats to the authors for a thoughtful work! I have been thinking and working on related ideas for a few months now but did not yet spent commensurate compute on them and might have gone in a different direction; this work certainly helps
359.
by pama 2y ago
I worked in materials science in the past. Although I dont recall all details, I have seen seen convincing presentations from researchers at the Weizman institute claiming strongly we are far from ready for a transition to a hydrogen econom
360.
by pama 2y ago
It is indeed somewhat surprizing why it took so long to figure out grapefruit juice (and it was by accident during the study of the effects of coffee on a drug). Some of the early drugs that interacted with grapefruit juice where short-term

More ›