RT @ID_AA_Carmack: This is good advice.

We have put a lot of effort into this summer school - please share!.

RT @jxmnop: spent the last month building my own framework to train a diffusion model from scratch. it was hard. almost like i just learne….

RT @SeunghyunSEO7: The concept of critical batch size is quite simple. Let’s assume we have a training dataset with 1M tokens. If we use a….

RT @kvogt: At Cruise I built robotaxis that completed 250k driverless rides, including the first rides ever in a major city. With the Tesl….

RT @lexin_zhou: 1/ New paper @Nature!. Discrepancy between human expectations of task difficulty and LLM errors harms reliability. In 2022,….

RT @Azaliamirh: Is inference compute a new dimension for scaling LLMs?. In our latest paper, we explore scaling inference compute by increa….

RT @karpathy: Jagged Intelligence. The word I came up with to describe the (strange, unintuitive) fact that state of the art LLMs can both….

RT @jon_barron: The legendary Ross Girshick just posted his CVPR workshop slides about the 1.5 decades he spent ~solving object detection a….

RT @jimjimson_: Visual demonstration of the universal function approximation theorem

RT @robmen: Lots of analysis of the xz/liblzma vulnerability. Most skip over the first step of the attack:. 0. The original maintainer burn….

RT @binarybits: Google recently released a new version of Gemini, its supposed ChatGPT-killer. I did a head-to-head comparison and found th….

RT @jxnlco: Just like in my own childhood. It’s where no one says “good job” and you just have to wander the earth, trying to make sense of….

This seems relatively comprehensive analysis of what Q* could be. The Tree of Thoughts paper was also hinted by Andrej Karpathy in his latest LLM video. I still stand behind my suggestion that Q* may refer to node processing priority queue ordered by Q-value, similar to A*.

RT @Francis_YAO_: Just discovered, silently, Qualcomm Snapdragon 8gen 3 already supported 10b language model running locally on your smartp….

Anyway, I wouldn't consider it AGI, as it is still missing long-term memory, continual learning, not even mentioning motor control or embodiment. Happy to be proven wrong :).

At the same it shifts computation from training to inference, so the text generation will be much more expensive. It's quite possible also that it is used only during training to generate high-quality training data.

I would be surprised if it totally changes the capabilities of language models, as they still try to imitate or please humans (pre-training and RLHF stages). But it is plausible that it can provide quite a boost in benchmarks that require reasoning.

While both MCTS and beam search do something similar, I guess Q* does it more optimally, allowing it to look through more interesting variations of the sentence. Altogether this provides the much-needed internal consideration or System 2 to the language models.

The objective for Q* is to generate a sentence with the highest reward, as opposed to A* that aims to produce a trajectory with the lowest distance. Q* maximizes, A* minimizes, but this is a minor detail.