[HN Gopher] Interview with Terence Tao in Barcelona
___________________________________________________________________
Interview with Terence Tao in Barcelona
Author : belter
Score : 50 points
Date : 2024-10-12 20:56 UTC (3 days ago)
(HTM) web link (english.elpais.com)
(TXT) w3m dump (english.elpais.com)
| blackeyeblitzar wrote:
| 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.
| koolala wrote:
| >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?
| solveit wrote:
| 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.
| ashater wrote:
| 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.
| raincole wrote:
| 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.
| energy123 wrote:
| 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.
| sahbasanai wrote:
| 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.
| arnogau wrote:
| 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.
| ithkuil wrote:
| 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)
| raincole wrote:
| 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.
| madcaptenor wrote:
| (In English, if you're wondering.)
| anthk wrote:
| 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.
___________________________________________________________________
(page generated 2024-10-15 23:01 UTC)