All published intelligence, sovereign stack research, ecosystem news, and AI foundation model reviews.
Hello and welcome back to our analysis of on the biology of a large language model by anthropic. This is a blog post that anthropic has published uh investigating what they call the biology of a large language model. This works through uh a technique called attribution graphs.…
Hello, how's it going? Today we're going to look at this paper right here. Energy- based transformers are scalable learners and thinkers. This paper combines the uh general concept of energybased models with transformers and suggests a learning paradigm for it and shows how that…
Hello, today we are looking at context rot how increasing input tokens impacts LLM performance. This paper looks at what happens if with these long context models now we are just sort of putting more and more stuff into the context of LLMs and even though all of these models do…
Hello there. Today we are looking at the free transformer from France Fur at Farret Meta. This transformer is extending the classic decoderbased transformer with uh a series of latent variables that can be used to kind of make underlying decisions about the sequence to be…
Hello there. Today we're going to look at on the theoretical limitations of embedding based retrieval. And this paper is really interesting uh because it proves something about embedding based retrieval. So think vector databases in rag and things like this. And the paper proves…
[snorts] Hello. Today we're going to look at Titans learning to memorize at test time. This is a paper by Google research and has been pushed as part of their Nurips publications. So there's a blog post about it and people posted on social media about it and um I have to admit I…
Hello, today we're looking at TiDaR Thinking Diffusion Talk and Autoregression by researchers at Nvidia. And this is a really cool paper because it effectively um makes good you already observes that we don't have full GPU utilization during autoregressive large language model…
The initials GPT stand for Generative Pretrained Transformer. So that first word is straightforward enough, these are bots that generate new text. Pretrained refers to how the model went through a process of learning from a massive amount of data, and the prefix insinuates that…
In the last chapter, you and I started to step through the internal workings of a transformer. This is one of the key pieces of technology inside large language models, and a lot of other tools in the modern wave of AI. It first hit the scene in a now-famous 2017 paper called…
If you feed a large language model the phrase, Michael Jordan plays the sport of blank, and you have it predict what comes next, and it correctly predicts basketball, this would suggest that somewhere, inside its hundreds of billions of parameters, it's baked in knowledge about…
[Submit subtitle corrections at criblate.com] Imagine you happen across a short movie script that describes a scene between a person and their AI assistant. The script has what the person asks the AI, but the AI's response has been torn off. Suppose you also have this powerful…
[Submit subtitle corrections at criblate.com] In January 2024, Google DeepMind released an AI model called Alpha Geometry, which could solve geometry problems from the International Mathematical Olympiad, or the IMO. The IMO is the highest level of competitive math contests at…
If you tell me what you like and what you don't like, I'll use that to train my model. Stack Quest. Hello, I'm Josh Starmer and welcome to Stack Quest. Today we're going to talk about reinforcement learning with human feedback and it's going to be clearly explained. This Stack…
You start with nothing, then you move. And you keep on moving till you can't get any more, then you're [music] done. StatQuest, bam. Hello, I'm Josh Starmer and welcome to StatQuest. Today, we're going to talk about optimization with linear programming main ideas. Specifically,…
We're going to find the optimal solution one row at a time. Stack Quest. Hello, I'm Josh Starmer and welcome to Stat Quest. Today we're going to talk about the simplex algorithm and we're going to go through the mathematical details. This stack quest is brought to you by the…
Hi, I'm Majid from Intel Research India, and I'm excited to be speaking at PyTorch Conference North America this October. I'll be talking about efficient mixture of experts LLM inference on using VLLM and OpenVINO, and how we can efficiently run these models on ARM CPUs. If…
Hi, I'm G from Google. >> I'm Colin from Meta. >> I'm Angela from Meta. >> We are going to be speaking at the PyTorch conference San Jose about new VLM and SG learn engines built on a new PyTorch native TPU backend torch TPU. >> Rather than build a new serving engine or porting…
Интересно представить, что будущее клинической помощи может заключаться в том, что мы будем слышать, прежде чем видеть. В исследовательском центре Google мы разработали акустические модели для здравоохранения, известные как Here, которые являются базовой моделью. С помощью Here…
Hi, I'm Masahiro Hiramoto from Mitsubishi Electric. I'm a member of the program committee for PyTorch conference North America and a PyTorch ambassador. I'm especially looking forward to seeing what people across the community are building with PyTorch and exchanging ideas with…
Hello everyone, I'm Carlos Costa, a distinction engineer here at IBM Research and co-founder of the LMD project. I'll be heading soon to San Jose for PyTorch Conference North America, where I'll have the privilege of sharing some exciting news about LMD. How the project changed…
When you look at the history of humanity, you can see just how important scientific discovery has been. From the invention of the microscope and the discovery of bacteria to the invention of antibiotics. So I want to understand what will artificial intelligence mean for science.…
Admit, this was actually much more simple than I was imagining, but nonetheless, it's interesting to read. So, we we The title gives it away, right? Embarrassingly >> [laughter] >> In in a previous uh journal club, we read a paper on self-distillation, and in that paper uh there…
Hi, in this video I'm going to walk you through a multi-agent setup of AutoResearch. I use open source models and an open source code harness, OpenCode, and I implement AutoResearch, which is a project by Andre Carpathie where he takes a nano GPT project, an LLM in its simplest…
Good morning everyone. How's it going today? Today we're going to be talking about the fine web data set. Now, this is a project by Hugging Face and it is a great example of how to gather all the data that you need to train a large language model. This actually is also available…
This paper is doing something if I understand correctly that is very similar to what pipeline RL proposed for reinforcement learning which is where you want to basically have like really decoupled generators from learners from trainers um but you have the additional complexity…
Всем привет. Если вы работаете с ИИ, у вас наверняка уже есть аккаунт на Hugging Face Hub, где удобно демонстрировать свои модели, скачивать модели с открытым кодом, использовать публичные наборы данных и создавать демоверсии своих моделей. Но публичный хаб—это только начало. В…
Let [clears throat] me think where to start. Okay, so we've got the Kim K3 report and it's quite a beast in terms of details, which is pretty awesome. And the bits that I focused on were really just the the pre- uh sorry, the post-training parts. And even that I wasn't able to…
Uh, this paper is actually quite interesting. It dives into something that that has been I guess a question an open question for the Ember communities the Um, we can call them like scaling laws for post-training. And then how is pre-training allocation of compute impact…
Okay, so this is quite um interesting paper and a bit uh surprising that uh how it works, but uh basically the idea is that uh in post-training, you have usually uh a pipeline where you have a model and then you uh run RL over that model. And ideally that uh post-RL model should…
I'm thrilled to introduce this course on fast and efficient LLM inference, built in partnership with Red Hat, and taught by Sergey Kliger. Deploying open source LLMs efficiently, so they can serve many users at once with low latency and reasonable cost, is challenging. In this…