agnosticmantis

These papers don’t explain how pertained LLMs learn in-context, because the simplified models in these papers are either pretrained for the same task that’s tested in-context, or the weights are handpicked by humans to do GD at inference time.

See this video for a good discussion: https://youtu.be/-yo2672UikU

I’m guessing it was the second attempt that really stuck.

How GPU-friendly is this class of models?

For a moment I thought Andrej Karpathy was the voice actor for the AI Assistant in the game.

If I didn't know this was from 2007 I'd have said it shows that he's been hanging out with OpenAI kids. (He worked there working on watermarking and stuff.)

(Altman having said they're working on an AI OS or something.)

> However, the idea is that often a lot of the probability mass - an amount that is not small - will be concentrated around the maximum likelihood estimate, and so that's why it makes a good estimate, and worth using.

This is a Bayesian point of view. The other answers are more frequentist, pointing out that likelihood at a parameter theta is NOT the probability of theta being the true parameter (given data). So we can't and don't interpret it like a probability.

According to [0]:

- Books3 dataset

- 700B text-image pairs from Laion-2B-en, filtered to only keep images with at least 256 resolution

- 400M text-image pairs from COYO-700M, filtered to only keep images with at least 256 resolution

- 10M text-video pairs from WebVid10M

- 3M text-video pairs from a subset of InternVid10M

- 73K text-video chat pairs from Valley-Instruct-73K

- 100K text-video chat pairs from Video-ChatGPT

0: https://huggingface.co/LargeWorldModel/LWM-Chat-1M-Jax#train...

Dating apps have very peculiar dynamics (e.g. you need to somehow get women on the app and men will follow automatically). Also women may be conservative and they might not want to join the latest dating app that ranks 50th on the App Store. So it’s not like anyone can create an app and be successful enough to be worth acquiring. Very few will reach that threshold and then the monopolist can buy those few ones.

So standard economics don’t apply. Also, the statement “ Economists overwhelmingly get these things right more than our intuitions” needs a citation.

What I really like and admire about Euler is how masterfully he handled infinities and infinitesimals to arrive at correct conclusions (the vast majority of times anyway) even though analysis hadn’t been made rigorous yet (by Cauchy and Weierstrass and friends), so some of what he did was pure magic and took a lot of good intuition.

An example of this is when, in solving the then notorious Basel problem, he factors trig functions into infinite products of (x +- k*pi) terms just by analogy with root factorization in finite polynomials.

Did you pay for the video proportional to the value that you received? Most people don’t, so why expect the services to not get worse, and be forever subsidized by another cash cow?

What are the severance packages like this time?

Reminds me of John Snow's discovery and demonstration of the cause of cholera, which I learned about in the context of casual inference in this excellent paper by statistician David Freedman:

https://psychology.okstate.edu/faculty/jgrice/psyc5314/Freed...

Actual science looks nothing like the shoddy paper churning that we see in much of econ and social science using questionable and assumption-heavy casual inference methods.

My mind was blown recently discovering Tubi with so many great movies (many of them classics and highly rated) for free (with occasional ads, of course) and I regret having paid for Netflix and other services in the past.

Schrödinger’s Charlatan: he’s both a genius and the most naïve when it comes to FSD, at the same time.

No sir, he’s been making fraudulent claims about present capabilities not just future: convoy tech 10 times safer than humans, NOW.

Can we (folks who didn't use 23&me but may still be affected because our relatives might have) file a class action lawsuit?

We haven't signed any licensing agreements with 23&me waiving our privacy, so presumably we still have some rights?

> Machine learning algorithms are learning to identify people just by their walk...

Turns out the UK government was working on privacy-preserving walks decades ago: https://youtu.be/eCLp7zodUiI

“It beats rail!”

Holy mother of dark patterns!

The default tab in the table is the “PROBABLE SAVINGS*” tab instead of the actual prices. (There’s actually no actual price, just estimates.)

Besides, does anybody believe their estimated prices for “available in 2025”?

* From the same trusted salesman who sold robotaxies making $30k/year!

Thanks, this clarifies it a lot.

Never seen these cool Chinese EVs in the US. Is the US market closed to Chinese EV?

Bring us another complete ordered field and I’ll pay handsomely for it.

He’s said it’s easy to generate this kind of questions that trick LLMs because of the lack of physical grounding in models trained solely on text.

And that’s true now as ever. I also heard him say that training multimodal models on text+image/video would mitigate the grounding issue, and that’s proven to be true too.

So I’m not sure exactly what your objection is.

Interestingly, the medium of transmission here also backs up the message being transmitted: Mastodon is quite barebones and maybe not as “feature-rich” as Twitter/X, but holy shit what a breath of fresh air it is to read content on Mastodon compared to Twitter.

Merely being able to scroll and read the whole thread without logging in shows how hostile the overly-innovated alternatives are.

LeCun is one of the rare voices of reason in the AI community, especially among the celebrities.

He’s been consistently voicing realistic statements with regards to AGI (he doesn’t believe it’s a thing), AI safety and doomerism, limitations of LLMs, etc.

Many are speculating that Sam Altman could just move on and create another OpenAI 2.0 because he could easily attract talent and investors.

What this misses is all the regulatory capture that he’s been campaigning for. All the platforms have now closed their gardens. Authors and artists are much more vigilant about copyright etc. So it’s now a totally different game compared to 3 years ago because the data is not just there up for grabs anymore.

"I didn't lie, I was just hallucinating" repeated Sam while being exited from the building.

>> incredibly difficult or even nearly impossible

This is pure hyperbole. Changing (default) search engines is incredibly easy. Using a different one is as easy as typing bing.com instead of Google.com in the address bar. Ta da.

Max(A U B) >= Max(A)

Similarly the top x% of both local and international are stronger than the top x% of local candidates, even if local and international have the same distribution of talent.

What are some questions that you think the unrestricted AIs can answer and are worth paying for (and that restricted ones can’t)?