My guess would be that this is a matter of fine-tuning the model to the problem. You would need to run the review experiment N times, observe the distribution of answers, and then fine tune with model operational parameters.
Perhaps someone more knowledgable could jump in here to clarify?
Side question to you, considering your occupation:
Could you please list a set of core papers (or other resources) that give a beginner an overview or even understand of the fundamental concepts and techniques with LLMs?
I looked into hackable printers 10+ years ago when building a demonstrator prototype of a bioproduced ink dye. The key challenge challenges is that they are an integrated system between mechanical, electronical, and chemical (ink formulation, surface chemistry) domains.
I would believe most relevant patents are now expired and that we should be able to build a printer that at least gives sufficient quality prints. Given sufficient motivation and time investment, that is.
Hmm. I am not so sure about that statement. There are many models at play in the market, not one model. What mechanistic patterns are driving those models that are falsified by this paper and lead to the outcome of increasingly extractive/stagnating economies?
My understanding is that in addition to your comment and the development of a method to separate the training data for distributed learning, the latency/bandwidth of systems connected on the internet is a challenge, too. Information has to be sent around before and after any hypothetical number crunching.
Thank you for your comment. I enjoyed it a lot. Good food for thought.
Your analogon is a bit leaky abstraction in the sense that it misses out on the broad stastical nature of LLMs. However, I find it is a good way to illustrate the potential industrial transformation.
It is hard to say what the future will bring. The original AsK HN post is definitly an omen for things to come.
That is incredible, but I must ask: Beside really getting to know Rust's trait and type system, and besides experimental / educational tinkering, could this have any practical use case?
I would like to know whether someone was able to use lower tier models for activities other than coding e.g. a limited version of a personal note manager - and what the hardware requirements in RAM for these models were.
> "He was 38. Shortly before his death, he had mailed a manuscript describing a new three-electrode semiconductor device to Physical Review. The paper was lost in the wartime Atlantic. Five years later, Shockley, ..."
I wish the article had a reference for that claim.
I remember from my childhood that my father told me that in the old soviet system, publications from were invented and dated back in order to demonstrate the superiority of their science. Both sides might have done it.
Now, a story from my father is not strong data point. But falsification of scientific theories, statistics and publications was a thing in the Soviet Union [1,2].
I guess other languages that follow a stricter/larger compile time regimen e.g. Haskell, OCaml, Idris, ATS, F* would be suitable as well for this kind of approach. Like Go, OCaml seems to offer low compilation times.
hn (at) rtoip (dot) com