tsurba

And so animals wouldn’t be conscious because they don’t use language? If so, I vote that the stupidest hypothesis of the year.

Or would one count any communication between animals as language? In that case almost any interaction would count.

Joplin is open source, syncing setup between devices is one login to Dropbox, works for free, with native apps on Windows/OSX/Linux/iOS/Android. It has a bunch of plugins too. If you just need markdown files with syncing, use it rather than paying for Obsidian sync.

The 2GB free quota on Dropbox is plenty enough for text (and some screenshots). Or you could self-host obviously. Git while lovely for source code is a hassle for notes.

True, good point, maybe not a straightforward consequence to extend to weights.

Many discriminative models converge to same representation space up to a linear transformation. Makes sense that a linear transformation (like PCA) would be able to undo that transformation.

https://arxiv.org/abs/2007.00810

Without properly reading the linked article, if thats all this is, not a particularly new result. Nevertheless this direction of proofs is imo at the core of understanding neural nets.

Edit: actually this paper is the canonical reference (?): https://arxiv.org/abs/2007.00810 models converge to same space up to a linear transformation. Makes sense that a linear transformation (like PCA) would be able to undo that transformation.

You can show for example that siamese encoders for time-series, with MSE loss on similarity, without a decoder, will converge to the the same latent space up to orthogonal transformations (as MSE is kinda like gaussian prior which doesn’t distinguish between different rotations).

Similarly I would expect that transformers trained on the same loss function for predicting the next word, if the data is at all similar (like human language), would converge to approx the same space, up to some, likely linear, transformations. And to represent that same space probably weights are similar, too. Weights in general seem to occupy low-dimensional spaces.

All in all, I don’t think this is that surprising, and I think the theoretical angle should be (have been?) to find mathematical proofs like this paper https://openreview.net/forum?id=ONfWFluZBI

They also have a previous paper (”CEBRA”) published in Nature with similar results.

You can show for example that siamese encoders for time-series, with MSE loss on similarity, without a decoder, will converge to the the same latent space up to orthogonal transformations (as MSE is kinda like gaussian prior which doesn’t distinguish between different rotations).

Similarly I would expect that transformers trained on the same loss function for predicting the next word, if the data is at all similar (like human language), would converge to approx the same space. And to represent that same space probably weights are similar, too. Weights in general seem to occupy low-dimensional spaces.

All in all, I don’t think this is that surprising, and I think the theoretical angle should be (have been?) to find mathematical proofs like this paper https://openreview.net/forum?id=ONfWFluZBI

But are we close to doing that in real-time on any reasonably large model? I don’t think so.

I agree with everything up until the AI part, and for that part too, the general idea is good and worth worrying about. I’m scared af about what happens to kids who do all their homework with LLMs. Thankfully at least we still have free and open models, and are not just centralizing everything.

But chatgpt does help me work through some really difficult mathematical equations in newest research papers by adding intermediate steps. I can easily confirm when it gets them right and when not, as I do have some idea. It’s super useful.

If you are not able to make LLMs work for you at all, and complain about them on the internet, you are an old man yelling at clouds. The blog post devolves from an insightful viewpoint into a long sad ramble.

It’s 100% fine if you don’t want to use them yourself, but complaining to others gets tired quick.

Thankfully in the EU you are not even allowed to sell sunglasses without proper UV protection, and can just pick up sunglasses from any market and trust they are fine, if a little flimsy.

EDIT: ok apparently anywhere else than the poorest of countries, too, really.

And how long have you been doing this? Because that sounds naive.

After doing programming for a decade or two, the actual act of programming is not enough to be ”creative problem solving”, it’s the domain and set of problems you get to apply it to that need to be interesting.

90% of programming tasks at a company are usually reimplementing things and algorithms that have been done a thousand times before by others, and you’ve done something similar a dozen times. Nothing interesting there. That is exactly what should and can now be automated (to some extent).

In fact solving problems creatively to keep yourself interested, when the problem itself is boring is how you get code that sucks to maintain for the next guy. You should usually be doing the most clear and boring implementation possible. Which is not what ”I love coding” -people usually do (I’m definitely guilty).

To be honest this is why I went back to get a PhD, ”just coding” stuff got boring after a few years of doing it for a living. Now it feels like I’m just doing hobby projects again, because I work exactly on what I think could be interesting for others.

Gambling is where I end up if I’m tired and try to get an LLM to build my hobby project for me from scratch in one go, not really bothering to read the code properly. It’s stupid and a waste of time. Sometimes it’s easier to get started this way though.

But more seriously, in the ideal case refining a prompt based on a misunderstanding of an LLM due to ambiguity in your task description is actually doing the meaningful part of the work in software development. It is exactly about defining the edge cases, and converting into language what is it that you need for a task. Iterating on that is not gambling.

But of course if you are not doing that, but just trying to get a ”smarter” LLM with (hopefully deprecated study of) ”prompt engineering” tricks, then that is about building yourself a skill that can become useless tomorrow.

Is it a puzzle if there is no algorithm?

But testing via coding algos to known puzzles is problematic as the code may be in the training set. Hence you need new puzzles, which is kinda what ARC was meant to do, right? Too bad OpenAI lost credibility for that set by having access to it, but ”verbally promising” (lol) not to train on it, etc.

I would argue the other way around. I have ADHD, but the thing that really helped me with work procrastination, which I think would help even without ADHD, was to find a job that is actually interesting.

In approx 7 years I went through working at all the top software companies in my country, but what really fixed my problems was moving on to being a researcher at the university. I’m now paid less than half from before, but it’s still enough, and I couldn’t be happier.

Getting to work on what I think is actually important and interesting every day is what helped. I also seem happier than the younger researchers who didn’t work at companies first, who don’t know how good they have it.

1.5-2 years ago I did some training for a ML paper on 4 AMD MI250x (each is essentially 2 gpus so 8 in total really, each with 64GB VRAM) on LUMI.

My Jax models and the baseline PyTorch models were quite easy to set up there, and there was not a noticeable perf difference to 8x A100s (which I used for prototyping on our university cluster) in practice.

Of course it’s just a random anecdote, but I don’t think nvidia is actually that much ahead.

Yet somehow chatting with Gemini in the web interface, it forgets everything after 3 messages, while GPT (almost) always feels natural in long back-and-forths. It’s been like this for at least a year.

I do machine learning research and it is very useful for working out equations and checking for ”does this concept already have an established name” etc.

It is also excellent for writing one-off code experiments and plots, saving some time from having to write them from scratch.

I’m sorry but you are just using it wrong.

OSS alternatives with free syncing to your chosen cloud already exist, and have plugin systems. Why not just contribute to those? Because this is either advertising or procrastination.

Or use Joplin which is open source and also creates markdown files, and setting up sync to a cloud provider that you probably already have is free.

Just use Joplin, it’s open source and syncing to many cloud providers you already probably have is free.

And going to a grocery store instead of 37 individual farmers…?

Nice that someone realized then already it sounds sus https://news.ycombinator.com/item?id=42128532

Probably why they tried to address it in the abstract already

I guess SO was quite useful for providing a lot of training material for technical question answering for LLMs? I don’t think they got compensated for it so that is a bit sad.

In terms of my use, most questions and answers either were low quality, the question was a bit off of what you really wanted to know, or if it was a problem that required discussion or opinions, a ”real” hobbyist forum was better.

While at some point I felt like I (tried to) look up everything from SO, I now basically never use it anymore, but thanks for all the fish!

The best thing to feed the podcast is a dump of all one-liner macros we’ve added to an IRC bot over 15 years (for fetching weather, stocks, and 99% stupid jokes) without any context. Cannot stop laughing listening to it trying to figure it out and bringing up the weirdest ones.

Good art is not going out of fashion. Tools may change but part of an artists job is to adapt.

For example, when you learn instruments you also train your ear and taste. These are things one cannot take shortcuts in.

I wouldn’t worry about it, but approach new tools (once they actually arrive and are not just advertisements like this one) with curiosity.

Yep not going to bother reading these hype advertisements about things that are not even available. This sucks so much from Google.

I actually enjoyed the article and the variety of drum beats, and the spectrograms, even though we have completely different woodpeckers here in Finland.

I don’t believe so. I think all important parts that each need to be scaled to advance significantly in the LLM paradigm are at or near the end of the steep part of the sigmoid:

1) useful training data available in the internet 2) number of humans creating more training data ”manually” 3) parameter scaling 4) ”easy” algorithmic inventions 5) available+buildable compute

”Just” needing a few more algorithmic inventions to keep the graphs exponential is a cop out. It is already obvious that just scaling parameters and compute is not enough.

I personally predict that scaling LLMs for solving all physical tasks (eg cleaning robots) or intellectual pursuits (they suck at multiplication) will not work out.

We will get better specialized tools by collecting data from specific, high economic value, constrained tasks, and automating them, but scaling a (multimodal) LLM to solve everything in a single model will not be economically viable. We will get more natural interfaces for many tasks.

This is how I think right now as a ML researcher, will be interesting to see how wrong was I in 2 years.

EDIT: addition about latest algorithmic advances:

- Deepseek style GRPO requires a ladder of scored problems progressively more difficult and appropriate to get useful gradients. For open-ended problems (like most interesting ones are) we have no ladders for, and it doesn’t work. In particular, learning to generate code for leetcode problems with a good number of well made unit tests is what it is good for.

- Test-time inference is just adding an insane amount of more compute after training to brute-force double-check the sanity of answers

Neither will keep the graphs exponential.

If figuring out how computers work starting all the way from logic gates interests you, but you don’t want to start by ordering parts and soldering, check out the ”game” Turing Complete.

Of course it’s not the same thing as actually building the electronics! But it can be a nice introduction.

Proper initialization is more important.

Batch norm and others are important for faster convergence due to forcing the model to focus creating second and higher order nonlinearities, as a simple shift in mean/std is normalized out, and thus the gradient does not point in a direction that would only change those properties of the output distribution.