chaxor

This is a bit different than the "read a paper" TTS app. I mentioned the idea just to say it's possible and coming. The blend of the two isn't out of the question though.

Think of asking for a reading of a paper wherein you could interject at any time.

System: "This work is presented fromainly 3 groups: Deepmind, University of Pennsylvania, and ETH Zurich - the authors are Matthew Botvinick, Dani Bassett, and Bastian Rieck. They uncover a useful meta-learning program that relies on an AT methodology rooted in the bifiltration of the Ricci curvature of the embeddings and training step, wherein ..." You: "Wait a second - the algebraic topology method - what are the prior works in that area and why would that be the starting point for this paper" System: "It appears that the relevant citations point to Anne Sizemore's work while in Bassett's lab, with a few other key authors such as Guisti. The titles suggest that..."

(...) System: "Now that we've cleared that up a bit (and added it to a research list for further exploration later), to continue on the paper ..."

And so on.

This is very achievable today with a little bit of work. Perhaps not easy to work _super well_ - but likely easy enough to get working to _some degree_. A well polished product that does work super well certajnly isn't out of the question though.

If cheap enough, it may solve a problem for people who don't know that they need to hire engineers to make that conveyor belt, or if they don't know if the conveyor belt will be needed for more than a few months.

If it's general enough, but does simple tasks, that's still a win over 'this is a nutcracker that work only with this precise nut +-1 mm at a rate of _exactly_ 1 nut per hour +-0.1 seconds.' which may be some of the types of machines out there.

What if you want to hear about the latest arxiv updates while on your morning run?

This seems like a fantastic idea for that purpose.

This is a great point. People will complain if LMs are applying to anything, but ultimately it improves accessibility, and allows for someone to dive deeper when needed.

There will always be ways to misinterpret some academic work, and there are plenty of opportunities in the path of understanding a work to do that.

Allowing someone to engage with a work _at all_ by lifting some barriers (visually impaired people's for exampld) should be acknowledged as an improvement, not discouraged continually for having some bugs.

There is definetly a way to make this happen though. Little bit o' whisper, Mixtral in some RAG, and you've got yourself a buddy to talk about the paper while it's reading it to you.

Of course everyone will immediately say this is dangerous and it may mislead you by giving wrong explanations, etc etc. and then others will counter with 'it will definitely get better over time' (the best models as products are ~3 years behind the improvements being show in academic work for example). However, ultimately this is just a neat product to make, even if it has some bugs. Listening to TTS right now spends about half the time reading jumbled numbers from tables and listing off author names. So just tackling that alone (which this would do much better) would be valuable.

"old man yells at cloud"

Seriously though, it's not going away no matter how much anyone hates it. Emails and blogs will continue to be written with it, letters of recommendation will be/are written with it, Presidential speeches will be written with it, academic articles will be / are written with it (almost all ml and cs research is), news is written with it... It's not going to stop, but it will _probably_/_very likely_ get better.

There is no tool, no human, no method to determine if text is generated with one of these models at high F-score (only sometimes high precision, low recall domains for silly examples).

We're stuck with it. Like the English teacher and their despised spell check.

They may have meant that deepmind is a good majority of the authors, though ETH Zurich, a few others and even openai is represented as well. Does seem like strange wording if the paper went read though

If you want to process 100M documents, you want to use the most performant and fastest option - which a 7b param model isn't going to be it.

Additionally, entity linking is an extremely common task, for which LLMs fail at pretty miserably (certainly if the dictionary is custom/private. Additional work must be done to somehow (!? Many options here ?!) perform EL.

So, in the end, making a silver corpus from an LLM may be an option for NER to train a much much smaller algorithm. But EL is _still_ not a 'plug and play' problem, and can actually be pretty difficult to do "well" (using the modern techniques of MHS, etc).

I think you can make tar.zst pretty easily in Android?

Yeah it was/is a mix of sarcasm and genuine use case. I do have a thing (almost) like this, but use it because I want the isolation of network, filesystem, and so on. I definitely hate that it requires so much (not 20G, but still big) for so little. There's probably a way to do the thing and get a 1Kb binary in some Go with namespaces solution or something, which would be much better, but I don't trust myself to implement it correctly even if I spent the time to learn it. So I'm stuck with a horrible solution of a several GB Debian docker because musl/Alpine doesn't cut it

This is really great.

I have been trying to get my docker to be more distributable. Right now it's just a simple python script in a python env inside a docker container inside a QEMU container to automate a click and then netcat something.

Pretty sweet. It's only like 20GB, so pretty lightweight by modern standards.

What's wrong with flatpak for that? Honest question as I don't know that space really

I think it's just to get free credits that you need to give a phone number?

To the other point, yes it's crazy that "When inside kitty, how do I get my python inside latex injected into Julia? (It somehow works using alacritty?)" Despite the question being pretty underspecified or confusing, it still shouldn't read as inappropriate.

Unfortunately, many image generation systems will refuse prompts with latex in them (I assumed it was a useful term for styling).

My best guess is that it thinks latex is more often used as a clothing item or something, and it's generally associated with inappropriate content. Just unfortunate for scientists :/.

This looks fantastic on macos. I like the project.

What does this have that is better than https://github.com/open-webui/open-webui ?

What about https://github.com/open-webui/open-webui ?

Seems to have more features than all of them

What's wrong with the Google names?

It seems pretty clear, they started with Bert, to borrow from AllenAIs Elmo, then Big Bird, and made some friends with Palms along the way. So of course Bard would make sense and is a natural next name.

Then it gets even simpler really, by switching Bard to Gemini it really streamlined the naming. Gemini nano < Gemini Pro 1.0 == Gemini Advanced Then after that Gemini 1.5 Pro, but that's still worse than Gemini Ultra 1.0 which is still better than Advanced. They have made it a bit easier now though with Gemma, which is worse than all of them, but still a little bigger than what you want to run on a phone, which is a great reason to introduce Germa X. I'm mostly excited for Germa X2 version deXbox though. That's when we finally get a decent model for desktop boxes.

Does Xfce work on Wayland? I would imagine qubes would either use wayland or not provide any windows management if they're focused on security. As I understand it, X11 has a lot of big security vulnerabilities that Wayland fixes, but not having any of that code or functionality to review makes it a lot easier to secure. Like the Alpine mentality.

While I agree with the hidden statement of the utility in online reinforcement learning here, it should be pointed out that for some snapshot of a system, which may have already been trained with a large amount of data - the structure of the learned space and the inferences within it seem like a reasonable definition for 'understanding'.

I don't know many people that would suggest that knowing about ones family structure, such as their mom/dad/uncle, and how their history relates to them, is required to be _completely reconstructed_ every time they interact with their environment somehow from first principles.

Online reinforcement learning can have large merits without resorting to stating it's required for learning. Just as self awareness can occur independently of consciousness can occur independently of intelligence is independent of empathy and so on. They are all different, and having one component doesn't mean anything about the rest.

I think we may agree more than you realize, but perhaps we aren't clear about what we mean. As I said pretty early on - they are similar at some level. To make an analogy, that could be as simple as a submarine (w2v) and a fish (authentic language) now 'uses fins' to move (bert). They're now more similar, but not the same. Of course there are differences, but we can learn by studying what structure has been added and how in order to work towards better theory (as well as looking at what differences there are). I have read all of these studies, as I am an academic in related fields, and you are correct that the resolution is not as good as we would hope (it's better in the ventral visual stream in non human primates, which is one reason we started there); however it doesn't negate the importance of the predictive power here. We can make very useful tools out of these models which can improve the lives of non-verbal people for example.

Another interesting set of works going in recently is in testing out the poverty of stimulus. There is now evidence from many labs that LMs learn with the same amount of data as humans, so the sample efficiency arguments and the 'poverty of stimulus' aren't really as successful as arguments for training difference either.

To the point of humans not learning certain languages, I haven't seen it proven that humans are absolutely incapable of learning certain languages. I do know that children ignore much of sequence structure and create their own internal structures, but this does not negate learning languages which have some arbitrary set of rules. I'm aware that models learn DNA more easily _practically_, but this is more of a statement about human experience and various other environmental factors - not a mathematical proof that it's impossible for humans to learn. Furthermore, that is a nice property of these systems. If they are capable of modeling many different phenomenon, they are useful for many things. If they can settle on a model configuration that linearly maps to a similar activation space with similar topological properties to a patient, that's also very useful. In other words, If the topological properties of the activation space are more constrained in the cortex vs more flexible NNs, but the NNs can fit to the constrained space, that doesn't seem like an insurmountable problem in studying either trained system.

Ultimately, I think we can agree that there are differences, but there are also striking similarities that can be very useful for improving science and medicine going forward, and perhaps (with _substantial_ effort in linguistics and mechanistic interpretability fields) we may be able to improve some of our understanding of linguistics and/or neuroscience (or perhaps not, but it's at least a promising potential lead).

You haven't provided any counter argument, and there are many articles with data backing this up.

Unfortunately I am not aware of any articles that can show convincing data that "LLMs don't learn like humans". I'm not really even sure what that means precisely.

Your statement could be understood to claim that language models cannot predict brain activations or vice versa? Predictive power is the foundation of science, so it's one of the better tests we have for these types of problems. To that point, there is substantial evidence of similarity and predictive power, as noted below.

Perhaps you mean something else? Surely you're point is not that GPUs simply aren't human. Of course - there are differences at various levels. So at some level, everyone recognizes they're not identical. Perhaps it's that attention architectures are not reproduced within neural microarchitectures in tbe neocortex? I haven't seen any studies on this within the relevant cortical areas to address this, though microarchitecture matching may not be required for output distribution matching. The interesting endeavour for science to show right now is how these systems are similar to our processing mechanisms. The ways in which they're not similar typically tend to be uninteresting, or obvious.

The fact that we can predict brain activity from NN model activations raises many more opportunities in science, and the predictive power has gotten far better with these advances,.

Here are a small set of articles from MIT, deepmind, etc. for some of the points I made:

  -  Tuckute, Greta, Aalok Sathe, Shashank Srikant, Maya Taliaferro, Mingye Wang, Martin Schrimpf, Kendrick Kay, and Evelina Fedorenko. "Driving and suppressing the human language network using large language models." Nature Human Behaviour (2024): 1-18.

  - Schrimpf, Martin, Jonas Kubilius, Ha Hong, Najib J. Majaj, Rishi Rajalingham, Elias B. Issa, Kohitij Kar et al. "Brain-score: Which artificial neural network for object recognition is most brain-like?." BioRxiv (2018): 407007.

  - Pereira, Francisco, Bin Lou, Brianna Pritchett, Samuel Ritter, Samuel J. Gershman, Nancy Kanwisher, Matthew Botvinick, and Evelina Fedorenko. "Toward a universal decoder of linguistic meaning from brain activation." Nature communications 9, no. 1 (2018): 963.

  - Arend, Luke, Yena Han, Martin Schrimpf, Pouya Bashivan, Kohitij Kar, Tomaso Poggio, James J. DiCarlo, and Xavier Boix. Single units in a deep neural network functionally correspond with neurons in the brain: preliminary results. Center for Brains, Minds and Machines (CBMM), 2018.

If you could, would you mind pointing us to some foundational theory on this that actually proves (preferably with analytical expressions or some other form of mathematics) anything around these statements?

It sounds like you read some of Emily Bender, et al's works and then simply regurgitated them. Unfortunately, everyone in this camp simply has large walls of text for their arguments, and basically no data, proofs, or any evidence whatsoever. The 'evidence' around "system X doesn't understand, like I do" amount to 'I don't like this' and 'trust me bro'.

In formal lingiistics, Chomsky has a point against ML for LMs, which is based around the lack of interpretability. His works did push our understanding forward - because they included formal mathematics in proving his statements. However, they tended to be more specific about what limits exist around expressivity, rather than your issues laid out here.

There are strong arguments to say that we do learn and represent language in a similar way to these NNs, at least at some level. 1) the distributional hypothesis and 'words defined by company they keep' is intuitive for how we learn language, and has provided enormous advances in NLP starting with w2v. 2) Dozens of academic papers from around the globe in different labs all reproducing the result in various ways that the distributional hypothesis word vectors (either contextual as in Bert/Gpt, or not as in gensim) map to brain activations by fMRI. This result also tracks in more resolute experiments in visual modalities for more invasive experiments in non-human primates.

I could go on providing data and experiments, but it would be nice if the opposing side to this would provide any at all that were convincing.

Again, that's a webapp product for the public - not state of the art research. These are very different things.

To be fair, they are not miles ahead of everyone completely - they're good at shipping products. Other companies (such as deepmind/Google) have capabilities approximately on par, or sometimes better than OpenAI - they just aren't very good at making products out of their research.

You may mean Gemini 1.0 Pro, which is the paid version of what used to be bard. Unfortunately Gemini Ultra is not available right now, and it seems Gemini 1.5 Pro (potentially similar to Ultra 1.0) may be available by waitlist.

It's because that's what the equipment is designed to use, and yes it provides more reliable results with the current technology. Much of the equipment used in healthcare relies on very old technology. This certainly does not mean it cannot be improved to work with less amount of material.

There are hand held sequencers for your phone now (minIon) that are very cheap and work. There are also large multimillion dollar bench top machines on the same technology from nanopore. Both are very new technology compared to what we used in early 2000s, and offer various different features and certainties from the analysis.

However, it is foolish to say that the minIon is impossible. It works. It's just less resolute than other systems.

This is true for many systems which people cry "Theranos" about. Their issue was lying. That was the problem.

It is important to point out that what they claimed to do is not technically impossible actually. They just didn't do it, and lied about it. That was the real problem.

It's very unfortunate, because it is technically possible, just very difficult to achieve, even in academia (so it won't be done first in industry).

This was the absolute worst part of Theranos, is that they deter others from trying to make headway in the space at all.

This is only specifically for web interface LLMs in the past few years that it's been lack luster. However, this statements is not correct for their overall history. W2V based lang models and BERT/Transformer models in the early days (*publicly available, but not in web interface) were far ahead of the curve, as they were the ones that produced these innovations. Effectively, Deepmmind/Google are academics (where the real innovations are made, but they do struggle to produce corporate products (where openai shines).

I just would like to point out that "why" is not a scientific question. Feynman mentions this quite a lot. The question "why" doesn't have answers in science. A question of "How" has a better chance of being answered in science.

I was interested in trying our mojo. Then I looked at it booked out quick.

No one will use a language that isn't free and open source.

If mojo was free and open source (wasn't a company), and didn't just give out binaries with a 'trust me bro' stamp if approval, then I would have worked with it. But it's not, so I will never use it.

I'm lost in this thread.

This whole time I just thought mainframe was an older style word for a large rack based server or server room. Like 'cloud' storage for someone's computer running miniserve.

What is the real difference between a mainframe and, e.g. a rack full of H100s, or rack full of 100GBps networking stuff, or some nice stack of 12x blades with 8x 256 core CPUs?

Why or how does a "mainframe" have more power than that?