1,454 karma · joined September 24, 2015
An LLM with CoT is Turing-complete. Training is, basically, compression (the training data gets lossily compressed into the model's weights). The information-theoretic limit of compression is an algorithm that reproduces functionality of a system that produced the training data.
No "magic" is required to get to a system that reproduces at least some facets of the human brain functionality.
Three years ago I was skeptical that stochastic gradient descent (and other known techniques) are the way. But evidence kept piling up.
The papers that use the halting problem or the Gödel's incompleteness theorem to prove something about LLMs are dime a dozen. The problem is they prove their results for any computable system. You need to also believe that the human brain contains "magic" to think that humans are exempt.
I believe I've said the same at the time this paper was published. There is no need for hindsight to notice the problem.
The required amount of compute and training data and whether the existing training methods were up to the task had the real potential to be show stoppers though.
TL;DR Buckmaster and Alpöge haven't solved Navier-Stokes blow up.
How information can get so distorted when it's trivial to fact check?
Fast thinking fails to activate slow thinking.
There's no clear decision criteria for this. Do trick questions demonstrate that most people don't "get it"? And, well, older model saying dumb things doesn't establish a general principle that LLMs don't "get it" in general.
> The sample efficiency is just crazy low
Autoregressive pretraining requires huge amount of data to go from a blank state to a somewhat functional model. Fine-tuning, LORA, reinforcement learning of foundation models and in-context learning are much more sample efficient.
> Chinese room
...creates a wrong intuition that by cranking a Leibniz's mill you are somehow responsible for whether it understands something or not.
LLMs' processing that reproduces statistical patterns of the training data is modified by post-training. That's why we have LLMisms, for example.
LLMs aren't simple patter-matchers/pattern-predictors. They are incredibly complex systems that capture some aspects of the systems that produce the training data.
What's this bullshit?
Nah. The best it can do is to use the best writing style a model learned. Post-training might fail to prioritize it, though. Autoregressive pretraining does not average things. It creates a predictive model for variety of programming styles.
Wind is more expensive. Onshore wind turbines have higher variability. Offshore ones are limited to coastal areas. Overall, there's a lot of practical problems. It's all doable, but building infrastructure takes time.
China's onshore wind potential is 12.6 PWh/year[1]. China's 2024 energy consumption is 9.85 PWh[2]. It's not much of overcapacity potential. Huh. I though it's higher.
[1] https://www.sciencedirect.com/science/article/abs/pii/S13640... [2] https://english.www.gov.cn/archive/statistics/202501/20/cont...
I literally laughed out loud, sorry. Providing baseload with BESS alone? That's rich. Literally. Extremely expensive. In addition to impractical amount of batteries it requires impractical overcapacity to compensate for seasonal variation.
> Then why is no other country burning that amount of coal?
China does about 30% of the world's industrial production and 40% of metal production, while having around 50% of world's coal power plants. They are working on it though. Many of their new coal plants are peaker plants.
> Trouble being they're still operating ~1200 coal power plants.
Yeah, it's because of the above points.
China is building tens of nuclear power plants though.
What can you say besides "I don't feel a soul inside an LLM"? You don't feel it? OK, good for you, whatever.