281–290 of 965 posts
Looks like it (the 8B version) blows out of the water the other free models, really great stuff from Yann and Zuck. Guess the moat is disappearing day by day.
Earlier quoted context omitted.
Hm, how much VRAM would this take to run?
My guess is around 256GiB but it depends on what level of quantization you are okay with. At full 16bit it will be massive, near 512GiB. I figure we will see some Q4's that can probably fit on 4 4090s with CPU offloading.
Earlier quoted context omitted.
I can't express how good Dwarkesh's podcast is in general.
Lex walked so that Dwarkesh could run. He runs the best AI podcast around right now, by a long shot.
Maybe I'll check out Dwarkesh, but even seeing him mentioned him in the same breath as Fridman gives me pause ...
I just want to express how grateful I am that Zuck and Yann and the rest of the Meta team have adopted an open approach and are sharing the model weights, the tokenizer, information about the training data, etc. They, more than anyone else, are responsible for the explosion of open research and improvement that has happened with things like llama.cpp that now allow you to run quite decent models locally on consumer h…
Earlier quoted context omitted.
It's a very plausible rumor, but it is misleading in this context, because the rumor also states that it's a mixture of experts model with 8 experts, suggesting that most (perhaps as many as 7/8) of those weights are unused by any particular inference pass. That might suggest that GPT-4 should be thought of as something like a 250B model. But there's also some selection for the remaining 1/8 of weights that are used…
What is the reason for settling on 7/8 experts for mixture of experts? Has there been any serious evaluation of what would be a good MoE split?
From Databricks: "DBRX has 16 experts and chooses 4, while Mixtral and Grok-1 have 8 experts and choose 2. This provides 65x more possible combinations of experts and we found that this improves model quality. DBRX uses rotary position encodings (RoPE), gated linear units (GLU), and grouped query attention (GQA). It uses the GPT-4 tokenizer as provided in the tiktoken repository. We made these choices based on exhaustive evaluation and scaling experiments."
https://www.databricks.com/blog/introducing-dbrx-new-state-a...
Initial observations from the Meta Chat UI... 1. fast 2. less censored than other mainstream models 3. has current data, cites sources I asked about Trump's trial and it was happy to answer. It has info that is hours old --- Five jurors have been selected so far for the hush money case against former President Donald Trump ¹. Seven jurors were originally selected, but two were dismissed, one for concerns about her im…
It's likely RAG / augmented with web data. Would be interested if local execution returned the same results.
It provided a lot more detail about the case, but does not have current information. It hallucinated the question about juror count, or maybe confused it with a different case seems more likely, one of the E Jean Carroll cases or the SDNY Trump Org financial fraud case?
I was curious how the numbers compare to GPT-4 in the paid ChatGPT Plus, since they don't compare directly themselves. Llama 3 8B Llama 3 70B GPT-4 MMLU 68.4 82.0 86.5 GPQA 34.2 39.5 49.1 MATH 30.0 50.4 72.2 HumanEval 62.2 81.7 87.6 DROP 58.4 79.7 85.4 Note that the free version of ChatGPT that most people use is based on GPT-3.5 which is much worse than GPT-4. I haven't found comprehensive eval numbers for the lates…
The bottom of https://ai.meta.com/blog/meta-llama-3/ has in-progress results for the 400B model as well. Looks like it's not quite there yet. Llama 3 400B Base / Instruct MMLU 84.8 86.1 GPQA - 48.0 MATH - 57.8 HumanEval - 84.1 DROP 83.5 -
Although 400B will be pretty much out of reach for any PC to run locally, it will still be exciting to have a GPT-4 level model in the open for research so people can try quantizing, pruning, distilling, and other ways of making it more practical to run. And I'm sure startups will build on it as well.
>We made several new observations on scaling behavior during the development of Llama 3. For example, while the Chinchilla-optimal amount of training compute for an 8B parameter model corresponds to ~200B tokens, we found that model performance continues to improve even after the model is trained on two orders of magnitude more data. Both our 8B and 70B parameter models continued to improve log-linearly after we trained them on up to 15T tokens. Larger models can match the performance of these smaller models with less training compute, but smaller models are generally preferred because they are much more efficient during inference.
Can someone experienced please explain this. Does this mean, a lean model with more training time and/or more (or better) training data will perform better than a fat model?
I just want to express how grateful I am that Zuck and Yann and the rest of the Meta team have adopted an open approach and are sharing the model weights, the tokenizer, information about the training data, etc. They, more than anyone else, are responsible for the explosion of open research and improvement that has happened with things like llama.cpp that now allow you to run quite decent models locally on consumer h…
Why is Meta doing it though? This is an astronomical investment. What do they gain from it?
They've got a console for it as well, https://www.meta.ai/ And announcing a lot of integration across the Meta product suite, https://about.fb.com/news/2024/04/meta-ai-assistant-built-wi... Neglected to include comparisons against GPT-4-Turbo or Claude Opus, so I guess it's far from being a frontier model. We'll see how it fares in the LLM Arena.
And they even allow you to use it without logging in. Didnt expect that from Meta.