arkano
arkano on Hacker News
Thank you, I love this app.
Posted by
arkano on Thursday March 14, 2024 @03:49PM
from the show-hn dept.
https://huggingface.co/spaces/mlabonne/chessllm
As I was playing with the Outlines library (https://outlines-dev.github.io/outlines/), I discussed with my friend Maxime how funny it would be if we set up a way to pair LLMs in chess matches till one wins. The first time I tried it, it required substantial prompt engineering to get some of those LLMs to propose valid moves. Large language models can mostly stay focused and even play rather...
It would be valuable if the filtering conditions are up to the user.
I appreciate that the playground frontend is just a streamlit app.
Flax, a big neural network JAX library, is developed and used by Brain. TF will stay around for boilerplate / data-loading, but JAX support probably isn’t going anywhere.
My job used to be about developing ML models for diagnostics (not vision, but other modalities). It mattered to me, it was fun and informative, and aspects of that company were a net positive.
Current job is about research in ML in general and no, I don't think it would matter a lot in the long term, but it's interesting and may turn out to be positive.
Does it look like it would be possible to run locally?
But don't lots of people live without one though?
The comments are similar, but the energy is different.
One of them read to me like "go and strive to be great in one of the million directions possible" and the other was "you won't win on any direction more than average, so settle down and live simple".
That's really cool! Reminds me of calca.
Not to be confused with element.ai
Complex analysis is why I stopped taking so many software engineering courses and turned to math.
> Just because it's there in the contract doesn't mean it's legal.
No, but should be likely that it is, right? Not worth the risk to put such clauses in there otherwise. And I was talking about the UK.
It can be the same thing in Europe, at least the contracts I've seen say so.
> It can also take significantly less time.
Strong opinion, but I think that if you want to learn facts, you can learn them faster on your own. If you want to learn a new skill, you need somebody to check your ideas until you can get to a level where you can self-correct yourself.
> II'm not sure the traditional lecture in front of a blackboard approach is relevant nowadays.
Sure, and I agree, most of the time I would prefer to study on my own instead of going to the lecture. However, it was incredibly useful to have an expert that could align my understanding whenever I was off to the wrong path; I think those were the opportunities for learning, not reading facts of off the board.
> There's no guarantee of that in any case.
Let's say that it's significantly more likely that you have practised on a variety of easy to hard problems in your field of interest if you have taken good courses from a university versus doing your own work. I believe that when you pick your own homework, it's natural to cherry-pick problems that seem simple.
> I think the main value of the PhD is being in general proximity to, and collaborating with people interested in the same field.
That would (and does) describe research divisions in industry too.
>I pretty much never learned anything of note directly from a teacher.
I've learned things after being corrected by a teacher, and then I practised on my own until the next mistake, at which point I was corrected again, and so on. It's this interaction that I find valuable, not just stating facts on a blackboard. As I have taught mathematics myself at the university level, I find that students don't really need me to read them the facts, I was more there to align their understanding.
As for consistency, I meant that you won't get to cherry-pick the topics that you like when learning on your own (or solve the problem sets that you find simple), which is a natural thing to do by the way. If you have your own curriculum and you stick to it, that's great, but I have found that when I allowed students to pick their own problems for homework, they grew weaker in some areas and stronger in others. This also didn't give me enough signal on their understanding in general, which meant that it deprived them of useful feedback.
> but this notion of consistency is kind of funny to me when the same degree from different universities (sometimes even the same university just a few years apart with different professors) can mean completely different skill sets.
Absolutely! I didn't mean consistency in terms of pushing out duplicates of the same, but consistency in terms of attacking a variety of problems in some course, allowing you to become well-rounded in your understanding. Once you reach that level, you can fill in the gaps and be comparable to a colleague that maybe had a slightly different curriculum.
> Having a paper written in 10 days accepted in the best conference of the field increased both my imposter syndrome, and the feeling that we are writing useless crap.
Are you sure you are not harsh on yourself? :) Maybe it was a good paper. If you are so worried, you can always ask the conference reviewers for comments; maybe they thought it was excellent.
> useless crap
You get enough "useless crap" down as knowledge and, who knows, one day you get to flying cars or the proof of Goldbach's conjecture.
I agree with everything, but
> Anything you learn in school, you can in principle learn on your own.
I think this statement is lacking some truth. We can rewrite it as:
> Anything you learn in school, you can in principle learn on your own, but it can take significantly more time and there's no guarantee of consistency (compared to someone learning with a teacher).
>some instituting is staking part of its reputation on giving the world a promise that you have some minimum competence in some field.
Yes, and part of this process is having your work peer-reviewed as novel by active researchers in the field of interest (and getting advice/corrections from them along the way to the finish). I see this as an integral part of the process.
It's curve-fitting with a few extra steps. You can do a lot with curve-fitting though.
There's at least one book that explores this. Life 3.0: Being Human in the Age of Artificial Intelligence
https://g.co/kgs/xr1eHw
> I've coded through pure muscle memory plenty of times
How many lines of code though? :D
As a mathematician, I have noticed that people that like to build something that, for example, can fly, start from a paper plane; they don't get discouraged because they can't yet build a 737 aircraft. However, in math, you need a lot of experience before you can even judge whether a problem is in fact a 737 and not a paper plane (and even then, you can be mistaken). I often see students discouraged because of this and that's why I suggest taking it slow from the beginning.
> Every now and then I try to delve into the frightening world of math.
To want to understand is to be human. :)
This raises the interesting question; how much is our attention really worth, say, per month, when it comes to ads? Per this vox article,
https://www.vox.com/recode/2019/6/24/18715421/internet-free-..., it would cost $35 per person in the US per month to replace ads in the US to cover the profits from ads. Maybe this isn't the right way to think when pricing attention though.
Four hours a day? You need to do something about that ASAP, don't just wait for retirement, have a shorter term plan.
> Maybe it is time to declare myself a senior after all.
I think you shouldn't force yourself on a technology you don't like just cause it's popular, i.e., smartphones with a touchscreen. Do try a phone with a larger screen though. :)
It should have been fish shell. :(
It's usually the opposite; people of status / power are allowed to ignore norms that plague the more common folk.
I rarely see any performance issues, even when running spacemacs over ssh. The only issue I have is start-up time and there are ways to speed that up.