Comment by eloisant
5 days ago
This is why I'm very skeptical about the "Nobel prize level" claims. To win a Nobel prize you would have to produce something completely new. LLM will probably be able to reach a Ph.D. level of understanding existing research, but bringing something new is a different matter.
LLMs do not understand anything.
They have a very complex multidimensional "probability table" (more correctly a compressed geometric representation of token relationships) that they use to string together tokens (which have no semantic meaning), which then get converted to words that have semantic meaning to US, but not to the machine.
Consider your human brain, and the full physical state, all the protons and neutrons some housed together in the same nucleus, some separate, together with all the electrons. Physics assigns probabilities to future states. Suppose you were in the middle of a conversation and about to express a next syllable (or token). That choice will depend on other choices ("what should I add next"), and further choices ("what is the best choice of words to express the thing I chose to express next etc. The probabilities are in principle calculable given a sufficiently detailed state. You are correct that LLM's correspond to a probability distribution (given you immediately corrected to say that this table is implicit and parametrized by a geometric token relationships.). But so does every expressor of language, humans included.
The presence or absence of understanding can't be proven by mere association of with a "probability table", especially if such probability table is exactly expected from the perspective of physics, and if the models have continuously gained better and better performance by training them directly on human expressions!
Exactly. It’s been stated for a long time, before llms. For instance this paper https://home.csulb.edu/~cwallis/382/readings/482/searle.mind... Describes a translator who doesn’t know the language.
In abstract we do the exact same thing
Perhaps in practice as well. It is well-established that our interaction with language far exceeds what we are conscious of.
1 reply →
It’s hard to believe this when the llm “knows” so much more then us yet still can not be creative outside its training distribution
2 replies →
Given a random prompt, the overall probability of seeing a specific output string is almost zero, since there are astronomically many possible token sequences.
The same goes for humans. Most awards are built on novel research built on pre-existing works. This a LLM is capable of doing.
LLMs don't use 'overall probability' in any meaningful sense. During training, gradient descent creates highly concentrated 'gravity wells' of correlated token relationships - the probability distribution is extremely non-uniform, heavily weighted toward patterns seen in training data. The model isn't selecting from 'astronomically many possible sequences' with equal probability; it's navigating pre-carved channels in high-dimensional space. That's fundamentally different from novel discovery.
That's exactly the same for humans in the real world.
You're focusing too close, abstract up a level. Your point relates to the "micro" system functioning, not the wider "macro" result (think emergent capabilities).
6 replies →