We've come to the consensus that large language models are just stochastic parrots... What makes us think that we can achieve a higher level of intelligence by putting them in conversation?
I think the next step in NLP will be a drastic innovation on today's learning model.
This is not the consensus among ML researchers. Transformers are showing strong generalisation[1] and their performance continues to surprise us as they scale[2].
The Socratic paper is not about “higher intelligence”, it’s about demonstrating useful behaviour purely by connecting several large models via language.
There is no such consensus. Transformers navigate problem spaces with various mechanisms that include recursion, and multi-pass inference means the depth can be arbitrary. This means that models pick up on the functions that generate answers, not simple statistical relationships you see in Markov chains.
"Stochastic parrot" is a derogatory term and I've never seen anyone who actually understands the technology use that phrase unironically. If anything, it's a shibboleth for bias or ignorance.
I asked something similar previously on HN and a researcher in the field said that scaling size/computation actually does keep showing significant improvements
We have not come to that consensus and large language models display really interesting capabilities like few shot learning, which before we thought would require a widely different architecture
> We've come to the consensus that large language models are just stochastic parrots
Anyone who thinks this REALLY doesn't know how language models work. A properly trained LM will only parrot something back because of lack of diversity in training data. This does happen in some cases (eg, GPL license or something) but those are pretty unique cases.
People on HN seem to think this a lot, but they are just wrong.
It'sespecially true for ML in general on HN, but it's generally true for a lot of areas in the public - people often mistake skepticism for expertise or knowledge.
I think the phenomenon is similar to the large crowd that cries "the sample is too small" any time statistics are brought up.
It's the first thing anyone learns, and it's easy to do.
It's really unfortunate, but that's why you see so many on HN that dismiss new technologies in ML (especially in NLP, since everyone can understand the output - that's less true in e.g. protein folding)
> people often mistake skepticism for expertise or knowledge
This is a pretty good insight.
> that's why you see so many on HN that dismiss new technologies in ML (especially in NLP, since everyone can understand the output
I think also in NLP people see output that is the same as some training data, so think it is copying it. It takes some a little bit more thought to think "ok if I asked 100 experts to try to write how to sort an array in Python" or their code is going to be very similar. This doesn't mean it is copied.
"Stochastic parrots" -- have you seen, e.g., the examples in the PaLM paper of how it does on "chained inference" tasks? I don't see how you can classify that as mere parroting.
"Stochastic parrots" is a disparaging term coined by SJW propaganda. As if the brain is not stochastic, or we don't parrot from cultural sources. Language models have been accused of bias and lack of explainability, but humans are biased too and can't really explain how we take decisions.
Overall this term says "limited to the intelligence of a parrot" which is false, models can solve math and coding problems, generate passable art, translate and speak in hundreds of languages and beat us at all board and card games. When was a parrot able to do that?
The math the models are doing are similar to rote rule chaining as opposed to calculation. The errors made look like kludged together lookups. I wonder if you could sequence the training of a model so that you could reinforce calculations over lookups, to encourage the development of an accurate and advanced mathematics module.
Neural networks can do math, but a lookup and memorized value model is structurally a lot different than a calculator model. The difference between them is a matter of weights for any given architecture. Tokenizing properly for math would help, but doing bit level tokenizing would be best, because that would allow multimodal domains to integrate more readily (i.e. audio/video/text models could share learned features more easily than if you are using parsed or domain specific tokens.) It's a great time to be alive.
To me, it is more proof of "stochastic parrot" behavior: model seen most of the available math information in internet, and even with significant computational power, can solve only 58% of elementary school level questions, and they were probably those with clear examples in training data, and can't generalize on those beyond.
In limited, often zero or one shot probing of the model, yes. Do multiple generations and recursive passes over the output to have the model select and iterate on a target and the utility goes way up. You can coax great output from small models, even the 125m parameter gpt-neo.
The process kinda goes like this -
Think of ten answers to this question: blah blah blah
From these ten answers, which are the best 3?
Of the three answers, which is the best?
Revise and edit the best answer to be simpler or more understandable.
Prompt engineering is a nascent field, and we haven't seen nuanced or sophisticated use of the tool yet. Most of the metrics reported in papers are barely better than a naive Turing test. It doesn't take much introspection to know that even humans endlessly iterate and revise their output, and the best extemporaneous speech doesn't match well curated and edited material. It shouldn't surprise us that similar editing and revision processes will benefit transformer output.
I think the next step in NLP will be a drastic innovation on today's learning model.