Meta's AI research blog, announcing models such as Llama, V-JEPA and Large Concept Models, plus its MTIA AI accelerator chips.
HN links it for Llama and other model releases; commenters also cite its MTIA chips using RISC-V and its Rule of Two for agent security.
> What exactly does the evaluation entail?
I believe the US AISI has published less on their specific approach, but they’re largely expected to follow the general approach implemented by the UK AISI [1] and METR [2].
This is mostly focused on evaluating models on potentially dangerous capabilities. Some major areas of work include:
- Misuse risks: For example, determining whether models have (dual-use) expert-level knowledge in biology and chemistry, or the capacity to substantially facilitate large scale cyber attacks. A good example of this is the work by Soice et al on bioweapon uplift [5] or Meta's work on CYBERSECEVAL [6], respectively.
- Autonomy: Whether models are capable of agent-like behavior, like the kind that would be hard for humans to control. A big sub-area is Autonomous Replication and Adaptation (ARA), like the ability of the model to escape simulated environments and exfiltrate its own weights. A good example is METR's original set of evaluations on ARA capabilities [3].
- Safeguards: How vulnerable these models are to say, prompt injection attacks or jailbreaks, especially if they're also in principle capable of other dangerous capabilities (like the ones above). Good examples here are the UK AISI's work developing in-house attacks on frontier LLMs [4].
Labs like OAI, Anthropic and GDM already perform these internally as they're part of their respective responsible scaling policies, which determine which safety measures they should have implemented…
Thank you very much for the thoughtful and insightful reply!
This is obviously speculation/intuition, but it's not terribly surprising to me at least that operating in e.g. pixel-space or a straightforward lifted latent manifold (modern diffusers basically) wouldn't have apparent structure under the fancy t-SNE type things that seem to be the heaviest artillery brought to the party (at least in the open). In pixel space, you get 6-17 fingers on 1-3 hands.
The `france - paris + uk === london` thing is real, and it's not surprising because there aren't typically much in the way of nonlinearities in `word2vec`/`fasttext`/`glove` type stuff. But this substantially survives all the leaky relus or whatever in LLMs. They're pretty clearly interpolating in a way that you could get close to with a composition of affine transforms.
JEPA (and maybe Sora if..., fuck it) seems a dramatic shift in forcing joint loss into a much higher-level space/manifold with (to me at least) shockingly semantic properties. I mean look at the I-JEPA reconstructions from pre-trained lifted space with some dinky diffuser/VAE-thing eating the hyperplane:
https://ai.meta.com/blog/yann-lecun-ai-model-i-jepa/
That's not pixel space, and you've got a lot of freedom to make it smoother, I suspect no one says "L1 regularization" anymore, but there's some modern version of that, we know how to do this.
AFAIU (and again, I welcome expert correction) TDA at least and really a lot of modern geometry is about…
Here's a summary of what's happened the past couple of years and what tools are out there.
After ChatGPT released, there was a lot of hype in the space but open source was far behind. Iirc the best open foundation LLM that existed was GPT-2 but it was two generations behind.
Awhile later Meta released LLaMA[1], a well trained base foundation model, which brought an explosion to open source. It was soon implemented in the Hugging Face Transformers library[2] and the weights were spread across the Hugging Face website for anyone to use.
At first, it was difficult to run locally. Few developers had the system or money to run. It required too much RAM and iirc Meta's original implementation didn't support running on the CPU but developers soon came up with methods to make it smaller via quantization. The biggest project for this was Llama.cpp[3] which probably is still the biggest open source project today for running LLMs locally. Hugging Face Transformers also added quantization support through bitsandbytes[4].
Over the next months there was rapid development in open source. Quantization techniques improved which meant LLaMA was able to run with less and less RAM with greater and greater accuracy on more and more systems. Tools came out that were capable of finetuning LLaMA and there were hundreds of LLaMA finetunes that came out which finetuned LLaMA on instruction following, RLHF, and chat datasets which drastically increased accuracy even further. During this time,…
There is a lot of interesting discussion already, but I think there is also some (perhaps unnecessary) disagreements regarding licensing arising from what I believe are the very different backgrounds in the community. Thus, I will attempt to summarise my own perspective as someone that has been active in the artificial intelligence community for nearly 15 years at this point.
Firstly, I will state that I am in full agreement with LeCun on what he states in the video about the need for AI systems such as LLMs to be open [1]. I am not sure if I entirely agree with his very high optimism regarding AI safety, but I have always been far more concerned about current risks and I suspect that in practice we are very much aligned.
[1]: https://yewtu.be/watch?v=EGDG3hgPNp8&t=6525s
Now, let me move to the issue of licensing that I have already talked about at great length in comments on other, older submissions. So let me do something new here and try to put it into a historical context.
Computer science and machine learning in particular has a tradition of openness that goes beyond that of many other fields. What I mean by this is that over the last twenty or so years, there has been a gradual move towards open sharing of research, research outcomes, etc. All under the umbrella of open science and FLOSS. One of the earliest examples I am aware of (there may be others) is how the Journal of Machine Learning Research was established as a direct reaction to the resistance against…
This is a bad article. Some of the information is wrong, and it's missing lots of context.
For example, it somehow merged Llama 4 Maverick's custom Arena chatbot version with Behemoth, falsely claiming that the former is stopping the latter from being released. It also claims 40B of internet text data is 10B tokens, which seems a little odd. Llama 405B was also trained on more than 15 trillion tokens[1], but the post claims only 3.67 trillion for some reason. It also doesn't mention Mistral large for some reason, even though it's the first good European 100B+ dense model.
>The MoE arch. enabled larger models to be trained and used by more people - people without access to thousands of interconnected GPUs
You still need thousands of GPUs to train a MoE model of any actual use. This is true for inference in the sense that it's faster I guess, but even that has caveats because MoE models are less powerful than dense models of the same size, though the trade-off has apparently been worth it in many cases. You also didn't need thousands of GPUs to do inference before, even for the largest models.
The conclusion is all over the place, and has lots of just weird and incorrect implications. The title is about how big LLMs are, why is there such a focus on token training count? Also no mention of quantized size. This is a bad AI slop article (whoops, turns out the author accidentally said it was AI generated, so it's a bad human slop article).
[1]…
Lol more like, Perplexity has a terminal diagnosis.
Doing RAG using like.. prompt hacking and text embeddings + vector store when you have no access to the underlying model, nor ability to fine tune the generation for RAG, will fail. It will fail in an epic fashion compared to doing RAG the right way.
What do I mean by RAG the right way? The RAG term has been overloaded.
There's RAG that's just kind of bolted onto the LLM after it's been fine tuned for instruction following, and then there's RAG where the document/fact retrieval is a part of the LLM itself that is differentiated and optimized.
Almost everyone is doing the first "hacky" kind of RAG, but Meta published in 2020, the "correct" way to do RAG, where you include a neural retriever in the feedback loop.
Almost no one is doing this because it's more expensive (requires fine tuning the model), but will produce much better results than doing "bolted-on" RAG.
https://ai.meta.com/blog/retrieval-augmented-generation-streamlining-the-creation-of-intelligent-natural-language-processing-models/
Perplexity does not have access to any GPT model weights. It's unlikely they'll be able to compete on quality.
It's game over.