SOTA multimodal LLMs can read text from images (e.g. signs, screenshots, book pages) really well. Are they actually using an internal OCR system, or do they learn to "read" purely through pretraining (like contrastive learning on image-text pairs)?
François Chollet, the creator of the Keras deep learning library, recently shared his thoughts on the limitations of LLMs in reasoning. I find his argument quite convincing and am interested to hear if anyone has a different take. > The question of whether LLMs can reason is, in many ways,...
Is this the paper you're talking about : https://arxiv.org/abs/2309.00941
Interesting read. Thanks for the write up.
This phenomenon can also be observed when training multiple probes with different random seeds: they converge to distinct directions, again showing lower cosine similarity than expected.
do you know why this happens? shouldn't the different seeds be converging to the same effiecient directions, like you say here :
while DIM provides one direction within the refusal cone, probe-based methods converge toward a different, more efficient direction, essentially sampling distinct regions of the same cone
SOTA multimodal LLMs can read text from images (e.g. signs, screenshots, book pages) really well.
Are they actually using an internal OCR system, or do they learn to "read" purely through pretraining (like contrastive learning on image-text pairs)?
François Chollet, the creator of the Keras deep learning library, recently shared his thoughts on the limitations of LLMs in reasoning. I find his argument quite convincing and am interested to hear if anyone has a different take.
... (read 429 more words →)The question of whether LLMs can reason is, in many ways, the wrong question. The more interesting question is whether they are limited to memorization / interpolative retrieval, or whether they can adapt to novelty beyond what they know. (They can't, at least until you start doing active inference, or using them in a search loop, etc.)
There are two distinct things you can call "reasoning", and no benchmark aside from ARC-AGI makes any attempt to
relevant article I just read - On the origins of empathy for other species
... (read more)Among the Mbuti foragers of West Africa, ethnographer Colin Turnbull noted that many magical rituals invoking animal spirits ‘convey to the hunter the senses of the animal so that he will be able to deceive the animal as well as foresee his movements’ (emphasis mine). Such an ability to take on the thoughts of others may have similarly scaffolded an understanding of one’s self
In many of these cases of deception and mimicry, we can find explicit references to having to think like an animal thinks, act like an animal acts, and pretend to be an animal. These forms of
"the tradeoff of pushing the complexity into the representation or the traversal of an algorithm"
could you say more?
Now, as people have pointed out, it is tricky to use split-brain cases to draw conclusions about people who have not had their hemispheric connections severed. But this case looks very similar to what meditative insights suggest, namely that the brain constructs an ongoing story of there being a single decision-maker
choice blindness experiments are a great example of this. people seem to change their actual preferences after having to come up with a justification for something they didn't even prefer but thought they did
in that attentional blink experiment why is T2 visible ≈75% of the time at first when lag is 100ms ? I seem to be missing something because it contradicts the finding
Thanks!