bravura

Can you please just do plain-old transcription with speaker diarization? That's the most important starting point for me.

[edit: The summaries are usually bad and I always want to refer back to what was specifically said through LLM Q+A]

Question: If I want to inference with the largest DeepSeek R1 models, what are my different paid API options?

And, if I want to fine-tune / RL the largest DeepSeek R1 models, how can I do that?

If it goes on your skin or in your body, do you trust Amazon 3rd party sellers?

Aside: I agree there is not enough mentorship, and I think a network/app for this would be amazing.

It would be great if:

a) it used LLMs to do partial AI match making

b) it had a relatively decent solution to avoiding mentors getting spammed by mentees (using existing trust networks is good except it disadvantages outsiders who deserve mentorship most)

c) it hooked into existing chat platforms so it's sticky without needing to visit the site

EDIT: I'm down to mentor on ML/AI and have in the past, for the right candidates. It's hard to put my finger on what qualities I'm looking for.

"Today many web developers consider jQuery to be “legacy software.” With all due respect to this perspective, jQuery is currently used on 75% of all public websites, a number that dwarfs all other JavaScript tools."

This argument is that jQuery is still very popular compared to JS frameworks like React, etc.

What about vanilla JS?

As someone who isn't that experienced with JS, my understanding is that modern vanilla JS is "about as good as" jQuery. i.e. if you don't need much JS these days, just choose vanilla JS.

Recent work by Feng et al from Bengio's lab focus on how attention can be formulated as an RNN ("Attention as RNN": https://arxiv.org/pdf/2405.13956) and how minimal versions of GRUs and LSTMs can be trained in parallel by removing some parameters ("Were RNNs All We Needed?" https://arxiv.org/pdf/2410.01201).

It's possible we start seeing more blended version of RNN/attention architecture exploring different LLM properties.

In particular, Aaren architecture in the former paper "can not only (i) be trained in parallel (like Transformers) but also (ii) be updated efficiently with new tokens, requiring only constant memory for inferences (like traditional RNNs)."

antirez, it's probably identical to the approach in this paper: Li et al 2024, "Evaluating Large Language Models for Generalization and Robustness via Data Compression" (https://ar5iv.labs.arxiv.org/html//2402.00861).

There's a pretty straight line from assigning probabilities (to a sequence of tokens) to arithmetic compression as an optimal compression algorithm for that distribution.

"It is well established that compression is essentially prediction, which effectively links compression and langauge models (Delétang et al., 2023). The source coding theory from Shannon’s information theory (Shannon, 1948) suggests that the number of bits required by an optimal entropy encoder to compress a message ... is equal to the NLL of the message given by a statistical model." (https://ar5iv.labs.arxiv.org/html//2402.00861)

I will say again that Li et al 2024, "Evaluating Large Language Models for Generalization and Robustness via Data Compression", which evaluates LLMs on their ability to predict future text, is amazing work that the field is currently sleeping on.

Full batch with L-BFGS, when possible, is wildly underappreciated.

Honest question, asking for informed responses: How harmful is second-hand vape?

Littered cigarette butts don't biodegrade.

Second-hand smoke causes cancer and the smell lingers on material.

Does other people vaping near you in an enclosed space have a meaningful physiological effect or otherwise? Or is it just association with second-hand smoke that people decry?

People keep saying stuff like this, but I don't believe it: "There is no technical moat in this field, and so OpenAI is the epicenter of an investment bubble."

AI progress is driven by strong valuable data. Detailed important conversation with a chatbot are much more valuable that quick search queries. As LLMs extend past web UIs, there is even more interaction data to capture and learn from.

The company that captures the most human-AI interaction data will have a TREMENDOUS moat.

If I am not an Anki user (yet) but want to create an Anki deck from scratch for something I am trying to learn, should I use your tool?

Do you mind sharing what you think are the killer features of fish?

With that said, an LLM to a) generate a LONG-list and b) help you zero in on your short-list; is something I would pay for.

Statistically chosen to maximize *some particular loss measure, which in this case might be the t-SNE or UMAP criterion, and is computed only globally and not for different filters.

For years, I've wished someone would port Pink Trombone to a VST or something I could control through code.

Does it really matter if it's simple to export your data and then import to another bookmark manager?

I'd be really curious to see what happens if you use PaCMAP (https://jmlr.org/papers/volume22/20-1061/20-1061.pdf) and more recent large-scale variants (https://github.com/YingfanWang/PaCMAP).

There is a lot of published research work in ML that had huge impact without explicitly touching linear algebra.

Given that is the case, to answer your question: yes, linear algebra is the foundation of ML. No, a lot of impactful day-to-day ML engineering can be done without touching linear algebra.

This is how assembly is the basis of compilation and programming. But you probably are going to get a whole lot of work done without ever using it.

It's generally a nice flex when applicants can code assembly, and usually a yellow-flag when the company suggests they require knowledge of assembly.

To me, my eyebrows raise when an industry person mentions linear algebra. I'm just saying the odds are really low that you actually use it.

Will this work? Because part of the LLM training is to reward it for always having a response handy.

People have been critiquing the collaborative filtering aspect of this work vs content analysis ("[why use stars instead of code similarity]") but there's something elegant about the simplicity of using less priors here.

A tf*idf matrix could be applied to the star-feature matrix too. Document = github repo. Term = name of user who starred it.

THUS, users who overstar are simply less important for computing similarities.

This would mitigate the phenomenon of massively popular github repos being clustered together because of folks who blithely star the most well known stuff.

Kill two birds with one stone:

1) Add functionality to auto-setup digital meetups and invite people within 200mi or something. 2) Let people meet in groups digitally, and maybe that spurs them to meet in person.

Wow this is amazing. It just gave me critique on my bodyweight squat form.

But I also found it hard to prompt to tutor in French or Portuguese; the accents were gruesomely bad.

Question: Have you tried using this for video?

Alternately, if I wanted to pipe a bunch of screencaps into it and get one grand response, how would I do that?

e.g. "Does the user perform a thumbs up gesture in any of these stills?"

[edit: also, do you know the vision pricing? I couldn't find it easily]

I seem to recall that there a recent theory paper that got a best paper award, but can't find it.

If I remember correctly, their counter-intuitive result was that big overparameterized models could learn more efficiently, and were less likely to get trapped in poor regions of the optimization space.

[This is also similar to how introducing multimodal training gives an escape hatch to get out of tricky regions.]

So with this hand-wavey argument, it might be the case that two-phase training is needed: A large overcomplete pretraining focused on assimilating all the knowledge, and a second that makes it compact. Other, that there is a hyperparameter that controls overcompleteness vs compactness and you adjust it over training.

What are the best static languages + frameworks for shipping CRUD web apps?

Not "exactly" how we learn. Humans learn through a combination of reinforcement learning (which is costly/risky/painful) and through observation of existing patterns and norms.

Better observation-based learning is a less expensive way of improving existing corpus-based approaches than trial-and-error and participating in an environment.

This is great.

I would actually perhaps think the next step would be to add some sugar that allows you to run a random / fixed grid of hyper-parameters and get a report of accuracy and speed for your specific data set.

There is definitely a need for this.

What I don't understand from quickly skimming your description and homepage: Do you source/provide the humans in the loop? That's a good value add, but how do I automatically / manually vet how you do the routing?

Conformal learning is relatively new to me. Tell me if I'm getting any of this wrong: Conformal learning is a frequentist approach that uses a calibration set to determine how unusual a prediction is.

It seems like the main time they aren't a strict improvement over bayesian methods is when it is difficult to define your calibration set? I know this scenario isn't so commonplace, but I'm working in a scenario where I quickly looked at conformal learning and wasn't sure if it is applicable.