Interview with Terence Tao in Barcelona(english.elpais.com)
english.elpais.com
Interview with Terence Tao in Barcelona
https://english.elpais.com/science-tech/2024-10-12/terence-tao-mathematician-its-not-good-for-something-as-important-as-ai-to-be-a-monopoly-held-by-one-or-two-companies.html
15 comments
Terrence Tao has a very healthy view of what AI can and cannot achieve shorter term. It is refreshing to hear more grounded views from a top mathematician who is clearly well versed in the topic.
The interview looks translated from Spanish to English, as the prosody looks closer to Romance languages. Altough who knows, as we Europeans prefer to use lots of Latin/Greek/French derived English words because of convenience, as most educated speakers in Academia will know the meaning. As Europe it's a hub on Romances/English on scientific environments, native English speakers might get these 'oddities' back.
But, as a con, we might get more difficulties on Germanic words which are pretty much a piece of cake for any English speaker as they are used to them far more than to "technical" Romance words.
Paradoxically we might get along better on highly educated English native speakers than having a random chat with the average guy using mundane words. In my case I have to do huge efforts in order to 'simplify' my English.
But, as a con, we might get more difficulties on Germanic words which are pretty much a piece of cake for any English speaker as they are used to them far more than to "technical" Romance words.
Paradoxically we might get along better on highly educated English native speakers than having a random chat with the average guy using mundane words. In my case I have to do huge efforts in order to 'simplify' my English.
The author is the science writer for El Pais and many of his articles come out in both English and Spanish (no translator credited? Maybe that's normal, or maybe he writes them in both languages at once). So I would expect him to write with a very Romance prosody in English.
This article is not in the Spanish edition (yet) from what I can tell, probably because the interview is in English, so you would need a full professional translation rather than an author translation.
This article is not in the Spanish edition (yet) from what I can tell, probably because the interview is in English, so you would need a full professional translation rather than an author translation.
This interview covers a few different topics, like detecting flaws in the Venezuelan elections, to dangerous applications of AI, to monopolistic control of AI (towards the end of the interview).
Monopolistic control is clearly a problem, and all the barriers (money for GPUs, regulations that pull up the ladder, hoarding of user data, ability for big companies to copy smaller ones, etc) make this AI transition problematic unless we also come up with new easy to enforce antitrust regulations.
But I think an even bigger issue is that these monopolists are all closed source. Open source in AI means we should be able to reproduce their exact model. Everything from training data to censorship filters needs to be transparent. Otherwise we really have no idea how we might be getting manipulated through these systems, whether by the company providing the models or those who can regulate or pressure or otherwise control those companies.
Monopolistic control is clearly a problem, and all the barriers (money for GPUs, regulations that pull up the ladder, hoarding of user data, ability for big companies to copy smaller ones, etc) make this AI transition problematic unless we also come up with new easy to enforce antitrust regulations.
But I think an even bigger issue is that these monopolists are all closed source. Open source in AI means we should be able to reproduce their exact model. Everything from training data to censorship filters needs to be transparent. Otherwise we really have no idea how we might be getting manipulated through these systems, whether by the company providing the models or those who can regulate or pressure or otherwise control those companies.
>Open source in AI means we should be able to reproduce their exact model.
like: Use this specific RNG seed and train for 10 trillion cycles in this data to get X duplicate model?
like: Use this specific RNG seed and train for 10 trillion cycles in this data to get X duplicate model?
Practicality of reproducing training runs that cost tens of millions aside, it's hopeless. Determinism is hard enough with a single GPU, fixing a seed isn't going to be much help when training is distributed across hundreds of GPUs.
(In English, if you're wondering.)
I don't get his reasoning about the Venezuela election.
In theory, the chances one candidate get exactly 56.1% votes and 56.108174% votes are the same, right?
We will need the prior probabilities of
1. Assuming we don't know the vote count, how likely it's rigged
2. Assuming it's rigged, how likely it's rigged in a way that the percentage is precisely XX.X%
To calculate
3. Given the vote count we know now, how likely it's rigged
Of course I practically know nothing about math compared to Tao, so I might completely miss his point.
In theory, the chances one candidate get exactly 56.1% votes and 56.108174% votes are the same, right?
We will need the prior probabilities of
1. Assuming we don't know the vote count, how likely it's rigged
2. Assuming it's rigged, how likely it's rigged in a way that the percentage is precisely XX.X%
To calculate
3. Given the vote count we know now, how likely it's rigged
Of course I practically know nothing about math compared to Tao, so I might completely miss his point.
Yes, those chances are the same. But they’re also the same as the chances of the result being 56.108175% or 56.1081756%. So, using your number of decimal places (choosing this number of decimal places is a slight simplification but it doesn’t affect the point), there are 10 million numbers for which the probability was the same. It’s like saying “draw a random number from 0 to 10 million” and drawing exactly 1 million. He says that this unlikely event happened with multiple results in the election, and this would likely mean there is anomaly.
I think he's defining a round number as, for some precision Y, X number of consecutive most significant digits are 0.
e.g. for a precision = 8 number, its "round" when the last 5 digits are 0, so like 32.300000 would be round.
The probabilities of round vs non round are pretty skewed. There are far fewer round numbers (for reasonable definitions) than there are non round numbers.
e.g. for a precision = 8 number, its "round" when the last 5 digits are 0, so like 32.300000 would be round.
The probabilities of round vs non round are pretty skewed. There are far fewer round numbers (for reasonable definitions) than there are non round numbers.
Not all numbers come out equally as likely. Round numbers are more likely to be picked by humans (or at least that's the argument)
That's why I said we need to know:
> 2. Assuming it's rigged, how likely it's rigged in a way that the percentage is precisely XX.X%
But I didn't realize he wrote a blog post about it. He actually addressed everything I listed (and more) in detail:
https://terrytao.wordpress.com/2024/08/02/what-are-the-odds-...
TL;DR: Yes, we need to know all these priori probabilities to get the exact answer. He made an assumption about them.
> 2. Assuming it's rigged, how likely it's rigged in a way that the percentage is precisely XX.X%
But I didn't realize he wrote a blog post about it. He actually addressed everything I listed (and more) in detail:
https://terrytao.wordpress.com/2024/08/02/what-are-the-odds-...
TL;DR: Yes, we need to know all these priori probabilities to get the exact answer. He made an assumption about them.
The chances are only the same if it wasn't rigged. If it was rigged, the chances aren't the same because humans tend to pick round numbers more frequently. Tao explains this in his longer blog post.
JoeyBananas(1)