somethingsome

Funnily enough, there are techniques that allows to go under the Nyquist limit, eg compressive sensing.

I don't like the analogy, it supposed that AI generated code will always be badly organized underneath if it is automatic, however, as time goes, I find the code better and better..

Thank-you, it's a longtime idea, but I was finally able to implement it, but now I'm trying to learn the procedure

I like Julia approach, compile everything on first launch, then reuse without speed penality, if you change a function, recompile that.

On another note, I'm working on a (big) PEP for python, and looking for maintainers and sponsors (as per python's rules). Do someone know where I could find such amazing persons? (I already have the code, works in cython,graalVM and pypy, and it's kind of a big change but I believe it will help a lot the language)

"leading them to suggest that LLMs will hit a wall in mathematical capability just beyond the convex hull of human mathematics."

This was to be expected, llms are trained on a very large corpus of data, they 'look' like creative from our point of view, as humans don't have that much knowledge, so, if the LLM proposes something that a human does not know, it looks like creative. But they stay neural networks after all, so very good in interpolation, very bad at extrapolation. Building new knowledge is extrapolation if it's truly outside the convex hull how the dataset.

I usually write code and ideas that are at the border of the convex hull, I haven't seen yet an llm being able to propose a really innovative solution. All the time they reuse building blocks that are already there. And I really tried pushing by making them creating proofs.

If I have an insight on some problem and I ask for the proof for it, it will usually be able to generate the proof, but that proof is never creative, and often it will use novel ideas only if I explain the novel idea. It will not come up with it alone. (And sometimes it's very hard to even explain the ideas, as it tends to return to knew waters)

Its only a data point, but I haven't seen a paper that handle this issue yet, and I don't expect them to magically become good at extrapolation due to the theory behind neural networks.

Sure, and I hope LLMs will at some point be able to do it. It would simplify greatly my work.

However, at the moment I consider that they stay in the 'convex hull' of their training set + a provided context, and I don't see that much research that made real improvements to the situation.

In my experience, LLMs are becoming very good at executing, but not a creating novel ideas or being creative.

Most of programming is reusing existing ideas in new shapes to solve new problems, but all the building blocks are there in the training set. Or new blocks can (easily) be derived from existing ones.

Math is different, it requires quite a bit of creativity, it's not just 'reuse all existing blocks'.

For the moment LLMs are good at discovering things that we overlooked in maths, or apply cleverly existing math blocks to make new results, but making a new theory that is really useful is out of reach for the moment in my opinion.

Very interesting!

I was looking at your website and wondering if there is a way to have access to the course material/videos?

In particular: Spring 2025 @ UR : CSC 209/409 Seminar on Artificial General Intelligence

Fall 2024 @ UR : CSC 277/477 End-to-End Deep Learning

Spring 2023 @ UR : CSC 266/466 Frontiers in Deep Learning

Spring 2022 @ Cornell Tech : CS 5787 – Deep Learning

Fall 2021 @ RIT : IMGS 684 – Deep Learning for Vision

Maybe what I'm saying is stupid or malinformed, but why not having the doc as a separate repo, and set it as a submodule to the repo where the code lives? Agent should be able to push code to the submodule independently from the code, and it should not be problematic for the main repository

I'm in the weights quite correctly, nice, but it also seem to hallucinate football players

Hi! Thanks for the effort!

It would be lovely to parse which datasets/benchmarks were used in the comparisons and select papers by dataset!

In many fields the datasets vary greatly depending on the subfield and its very difficult to find what other benchmarks could be used.

At least onedrive do not corrupt my powerpoints, while pcloud systematically break them, police change, elements resized randomely. It happens systematically with powerpoints weighting more than 100Mb.

My workplace has named (forcefully) the onedrive folder with around 35 characters. You add to that the path to that folder on the computer that is (forcefully) not on the root of a disk. I now mostly need a flat structure for my files. 4-5 subfolders and a file and onedrive dies.

While I agree with the message, I don't agree on the tools. It's very difficult to define a specification that works as intended, even with tools. Most waterfall software methodologies failed for a reason. And tools of the past are really not usable with AI. We need tools where it is way easier to adapt the specification iteratively, and even better, to have a bidirectional conversion. You define the spec, the LLM generate code, from the code you extract the spec, now you can compare and iterate. Then the model can focus only on the differences.

The other main issue that I see, is that even if there is a formally verified specification, at the moment, LLMs will not respect it perfectly. As long as LLMs are not able to non-deterministically follow a spec, the technology is not good enough.

A part from that, imo, in this age we should focus more on the mathematical aspect of computations, and I think we need to develop novel theories that take into account the non deterministic nature of LLMs in the process. I'm not sure this will ever work by merely extending current practices, as software design practices are extremely poorly defined from an engineer point of view. Just extending them by including randomeness does not seem a good idea.

Thanks, I just tried, for me it worked on 2x L40S with vLLM. I had some issues due to the model name, forge was forwarding 'default' instead of the real model name 'Qwen2.5-Coder-14B-Instruct'.

If someone else struggle on this step, I added in vLLM args: --served-model-name "Qwen2.5-Coder-14B-Instruct" --served-model-name "default"

So default becomes an alias.

I didn't yet test Forge, I was just happy that it worked at the moment ;)

I was hoping it would work with vLLM (openai compatible) to test it, does anyone know a similar proxy for local coding models?

Not really, Hardware didn't really change that much, of course you'll not find Tensor or raytracing cores, but you will have a very solid grasp of gpu programming and the cuda language (that didn't change that much either), and then you can easily learn those more modern things with blog posts or even, at worst, chatgpt.

Having read or at least skimmed most of those books, I think the best intro is 'CUDA Programming: A Developer's Guide to Parallel Computing with GPUs'

Massively Parallel Processors: A Hands-on Approach is not really good in my opinion, many small mistakes and confusing sentences (even when you know cuda).

CUDA by Example: An Introduction to General-Purpose GPU Programming is too simple and abstract too much the architecture.

Next year I'm planning to start writing a cuda book that starts by engineering the hardware, and goes up to the optimization part on that harware (which is basically a nvidia card) including all the main algorithms (except for graphs).

I'm already teaching the course in this way at uni, and it is quite successful among students.

It would be nice to have mandatory SD cards..

95% of the cases are easy for both doctors and AI, where doctors excel are the difficult cases where there is only a very limited amount of training data ;) something AI is not yet ready to handle at all.

I mean... There is a 1-1 mapping, and they look kinda like QR codes. so technically, you can make an app that scan it and it will show you the corresponding polynomial.. It could even be useful for fast checking knots

It's highly dependent of why you use it. For me a problem looks like 'a step in the proof I'm not familiar with', and I use LLMs to help me undersand it deeply. Make visualizations, check some difficult step, do parallels with something else I know,... I don't really care that the llm could 'solve the global problem I'm facing'. I use it more for insights on smaller parts to be able to go through difficult steps and teach me areas I'm not familiar with. The more the llm is capable of doing complicated proofs by itself, the more it is trustworthy to help me without making errors that I could miss in unknown Maths areas.

I would like to see a (joke) skill that makes Claude talk in only toki pona. My guess is that it would explode the token count though.

Personally, I wrote an essay to my students explaining exactly that the purpose is for them to think better and improve over time, they can use LLMs but, if they stop thinking, they are just failing themselves, not me.

It had great success, now when I propose to them to use some model to do something, they tends to avoid.

At the same time, if you remove 'engineer' , informatics should fall under the faculty of Science, so scientists, which are even more rigorous than engineers ;)

Maybe software tinkerer?

LLMs make too many mistakes when summarizing papers in their current state, I would never trust it to summarize a whole paper at the moment.

I only use it on a sentence or paragraph basis, otherwise it misses the point 90% of the time.

I would strongly advise against this use for the moment. The important part of reading a paper is not only to extract general rules, but to build your own internal model. Without it you cannot effectively do research. The main interesting points are often in the subtleties of the details deep in the paper.

Internal tought that come easily to mind when I read :

- 'oh they used that equation, but that could be also be interpreted totally differently, what happens if we change point of view, does it makes sense from this other perspective'

- 'I see they claim to achieve better results than sota, but actually, they compared with other methods that are not solving exactly the same problem, what shortcut or changes did they had to do to obtain a fair comparison, is it a fair comparison, can I trust those numbers? '

- 'oh, the authors didn't realize that they solved this other problem, or did they realize but there was a block somewhere preventing it?'

- 'I like this trick to achieve that result, but at the same time, it will prevent to solve a whole class of other problems, so their method will not work on those cases'

...

Also, notice that a paper IS a summary of multiple months/years of work, and researchers summarize it already to the maximum to stay within the page limit, by summarizing a summary you will always miss many things.

Sometimes you don't need a collaborator if you have the idea. If the other party is not at all working on the angle that you're interested in, it's probably not the correct collaboration to ask to.

Also, a collaborator is usually not a stranger over the internet, it's often someone who you know and you already worked with, so it is not that ackward to expose a new idea and propose to work together.

It takes time and social skills to make long lasting collaborations, the two parties must trust each other in order to collaborate. In this context, exchanging ideas is not really an issue.

I'm a little annoyed that 'modern AI' refers here only on LLMs, modern AI is way bigger than that.

Having said that, it's probably a good course, CMU courses are often great.

I was just expecting way more sota models in many fields due to the title.

If someone has this kind of ressource I would be extremely interested!

Very nice project, I see many use cases.

However, I use Manim for maths, for me having computational libraries (python) is a requirement. Most of the transformations that I do are found by using linear algebra, calculus and sometimes full neural networks. All my geometry is computed, not placed by hand.

I'm wondering if it would not be possible to have a 'canvas' backend for the web in the python version instead.

Btw, Manim is kind of easy to install in its own docker ;)

I really loved 'Track Down' (2000), if you want to try. The film depict maybe more 1980-1990 though

I think we agree in general, I don't disagree that maybe Proton therapy is not better than radiotherapy, it might but we lack some evidence.

I only argue that if they are equal in quality of treatment and the 'total cost' is the evaluation parameter, it is way more complex than the treatment itself, and it could be justified to use proton therapy, even if more expensive.

Nice talk anyways :)