dontreact

Gemini 1.5 Pro is on-par with GPT-4 and Opus in LMSys, and you can go try it for yourself in LMSys, and it's coming soon to Gemini Advanced (announced at I/O). Seems like GPT4O puts OpenAI in front again on LMSys

Gemini Ultra has been available for people to try via Gemini Advanced (formerly Bard) for a few months

It’s the difference between doing something transcendental a small number of times and doing something amazing billions of times.

Yes thanks that is what I meant.

The biggest reveal here is that they provide one of the clearest confirmations that Tesla has been lying about their crash statistics. Apparently they only log crashes with airbags and generally from police reports we know this is only 18% of crashes. They found and confirmed several instances of crashes that don’t show up in Teslas logs at all.

Maybe but it just seems silly to not briefly try out a few of the options as a teacher model before settling on one (Claude)

Im surprised they didn’t try out Gemini. It’s a lot better than Google translate and in my experience this is one use case where Gemini outshines even other frontier models like ChatGPT4

Wouldn’t building anything with this scrape be a YouTube TOS violation?

Could this be used for maximum cut?

Maximum cut at one point was an important concept in IIT which is a somewhat debunked but still interesting theory of consciousness

Looking at the the list, I think only a very small fraction of the 200 that signed are millionaires. Only the millionaires can get attention for this by signing.

It's pretty clear that what's going on is they are violating YouTube's terms of service and so if she were to be honest that would be grounds for a lawsuit.

Relevant part of the TOS

""" Permissions and Restrictions You may access and use the Service as made available to you, as long as you comply with this Agreement and applicable law. You may view or listen to Content for your personal, non-commercial use. You may also show YouTube videos through the embeddable YouTube player.

The following restrictions apply to your use of the Service. You are not allowed to: ... access the Service using any automated means (such as robots, botnets or scrapers) except (a) in the case of public search engines, in accordance with YouTube’s robots.txt file; or (b) with YouTube’s prior written permission; ... """

Sora is not a public search engine, and pretty sure they didn't get YouTube's permission to scrape.

Full TOS https://www.youtube.com/static?template=terms

This is a nice policy but unless you start immediately backdating it

(which would be far more disruptive than a standard sized wealth tax of some kind and would in effect be a one time enormous wealth tax)

since the effects would take a very long time to kick in, how can we be sure that the rules won’t sinply get eliminated 100 years from now once a few ultra powerful billionaires in even more full control of the system decide “actually no let’s not do it”

So yes this is a nice equilibrium state but I don’t see the vision of this as necessarily helpful for moving to a better state.

What are the policies that can be implemented in the next 10 years to address this problem?

Not if your wealth is in an appreciating asset that tends to appreciate faster during inflation such as real estate

I also wonder if it’s a language effect and the models are just way better at python

Yes to me that’s obviously not a good use, and I have experience to know not to attempt hundred line rewrites of files.

I order to reflect on my mental about this. It’s that i more often break down parts problem into sub routines that are 30 or so lines. These are nailed about 95% of the time and when there’s something wrong it’s obvious.

It’s a tool you learn to use and improve with. How many hours of practice have you given the tool?

Personally I’ve been practicing for almost 3 years now going back to CoPilot so I would guess I have at very minimum a few hundred hours of thinking about how (what to prompt for, what to check the output for) and probably more importantly when (how to decompose the task I am doing into parts for which some the LLM can reliably execute) to use the tool.

While subvocal is cool and would allow for speech in more places, something that’s earlier on the tech tree and that I would like to see is just robust lipreading.

I already am comfortable talking to my phone quietly using my AirPods while looking at my screen, but it seems like in loud public places the accuracy becomes unusable. I imagine it could be easily recovered by the additional signal of lipreading.

I think there is probably the option to opt out of this feature, it's just that it will be on by default?

Well you could just say the level you start at is real time and the second hand moves slower for lower levels and faster for higher levels?

Isn't the proportion of people who are obese and therefore go on this diet also going to be much higher? They appear to not have controlled for anything.

So in essence the implied cause and effect here for most people reading this headline (intermittent fasting causes cardiovascular death) is actually backwards: (people who are unhealthy are more likely to initiate an IF diet).

Reminds me of the story around diet soda. Some initial association studies found that there was an association between diet soda and several negative health outcomes but upon further investigation the evidence seems to be that it does not cause them.

No I’m talking about an independent analysis by Times

None of this is inconsistent with changing the level of amplification of certain topics in concordance with the preferences of the CCP, which is what the evidence appears to point to.

Note the key word “removal”

It seems like a good idea that if the endowments are getting so big that covering tuition for all of their students would still allow them to grow, then it's time to push for more of the money to be used for paying for public colleges.

Although I wonder what the weird second order effects will be.

How would socioeconomic factors lead to bias in a model? I figured there would be way more recordings in Latin American Spanish that u supervised learning would anchor on more

His job at Tesla or OpenAI wasn’t as an educator though. I think a clearer version of my point is that most of his impact has come from activities he has done “on the side” and hasn’t gotten paid for from his job. I’m hopeful it can be his main gig now given that YouTube creators seem to be making more money.

I don’t think we disagree. Education is crucial and the value is enormous, but this hasn’t been what he was paid for in the past. I am hopeful that he finds a way to make this his job more directly than at Tesla or OpenAI as the whole world will benefit.

Unpopular opinion… but IMO almost all of Karpathy’s fame an influence come from being an incredible educator and communicator.

Relative to his level of fame, his actual level of contribution as far as pushing forward AI, I’m not so sure about.

I deeply appreciate his educational content and I’m glad that it has led to a way for him to gain influence and sustain a career. Hopefully he’s rich enough from that that he can focus 100% on educational stuff!

My point is general it’s better to do a deep dive into individual musicians and styles to look for findings than to take a framework of analysis developed for one style of music and use it to compare across very different styles of music.

For an interesting analysis of rhythmic variations not represented by classical notation, this is one of the best articles I e seen so far

https://libres.uncg.edu/ir/uncg/listing.aspx?id=7337

I would have if I could think of one. Also anyone would think those rhythms were chaotic if they had no cultural familiarity with them (including myself!), so I don’t see it as a criticism?

Here’s an example of someone with a lot of twitter followers in the tech world using information theory to compare across different types of music:

https://twitter.com/aidangomez/status/1749532024258412740

While neat and interesting, I think using information theory to analyze music can very easily fall flat especially when comparing across very different musical traditions.

A relevant story for this:

When Europeans came to Senegal they perceived the drumming as being irregular and chaotic. Through taking a class on this drumming I have been able to appreciate how much subtle time differences not represented by western music notation play a role in this music, and also how complex the rhythms are (even when just approximated via western music notation)

Very often there is an implicit discretization that happens before these analyses of music are run and it’s easy to choose a wrong one or for there to not be an easy and human interpretable way of defining the domain of your probability distributions required for such an analysis.

Another example: in pop music timbre plays a much bigger role than in classical music as artists have experimented with vocal styles, synthesizers, and many other ways of shaping sound which cannot be represented by classical music notation.