Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434

Aravind Srinivas · 3h 02m · youtube.com ↗

20 korrents from this recording

Aravind Srinivas did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

0:17:27 · watch on youtube.com↗

Our reading
The way to make a dent in search is not to do what Google does better, but to do the thing Google does not want to do.

Their wordsWe would rather take a more dramatic position, that the best way to actually make a dent in the search space is to not try to do what Google does, but try to do something they don’t want to do. For them to do this for every single query is a lot of money to be spent, because their search volume is so much higher.

Interesting
8 min later
0:25:33 · watch on youtube.com↗

Our reading
A startup does not have to set out to build the greatest business in history; a merely good business is a legitimate thing to aim at.

Their wordsYou can set out to build a good business and it’s still fine. Maybe the long-term business model of Perplexity can make us profitable in a good company, but never as profitable in a cash cow as Google was. You have to remember that it’s still okay.

Interesting
1 min later
0:26:05 · watch on youtube.com↗

Our reading
Google's weakness is its own margins: it cannot go aggressively after anything less profitable than a link click, which is exactly where a challenger should attack.

Their wordsWhat is the weakness of Google is that any ad unit that’s less profitable than a link, or any ad unit that kind of disincentivizes the link click is not in their interest to go aggressive on, because it takes money away from something that’s higher margins.

Interesting
13 min later
0:39:20 · watch on youtube.com↗

Our reading
Prompt engineering is not a durable skill, because a good product is one that works for the user who never learned to phrase a question.

Their wordsAnd also this is where I believe the whole prompt engineering, trying to be a good prompt engineer is not going to be a long-term thing. I think you want to make products work where a user doesn’t even ask for something, but you know that they want it and you give it to them without them even asking for it.

Interesting
2 min later
0:41:07 · watch on youtube.com↗

Our reading
The obstacle to search is not the competition but the fact that people are not naturally good at turning curiosity into a question.

Their wordsThere is a skill to asking good questions, and everyone’s curious though. Curiosity is unbounded in this world. Every person in the world is curious, but not all of them are blessed to translate that curiosity into a well-articulated question.

Interesting
16 min later
0:57:04 · watch on youtube.com↗

Our reading
Open-sourcing capable models matters because it is the only thing standing between us and a world where two or three companies control the frontier.

Their wordsAnd it gives hope for a world where we can have more players instead of two or three companies controlling the most capable models. And that’s why I think it’s very important that he succeeds and that his success also enables the success of many others.

Interesting
9 min later
1:05:49 · watch on youtube.com↗

Our reading
The transformer was close to the last answer in architecture: nothing about the core design has changed since 2017.

Their wordsAnd that is the, I would say, it’s almost like the last answer. Nothing has changed since 2017 except maybe a few changes on what the nonlinearities are and how the square descaling should be done. Some of that has changed.

Interesting
6 min later
1:11:19 · watch on youtube.com↗

Our reading
Post-training is what turns a model into a product, but it can only do that on top of a pre-trained model large enough to already have common sense.

Their wordsAnd without good post-training, you’re not going to have a good product. But at the same time, without good pre-training, there’s not enough common sense to actually have the post-training have any effect.

Interesting
2 min later
1:12:57 · watch on youtube.com↗

Our reading
Spending most of the compute memorising the world in pre-training is brute force; what we want is a model that reasons the way a student sits an open-book exam.

Their wordsI think there’s an interesting thought experiment here that, we’ve been spending a lot of compute in the pre-training to acquire general common sense, but that seems brute force and inefficient. What you want is a system that can learn like an open book exam.

Interesting
11 min later
1:23:37 · watch on youtube.com↗

Our reading
Real reasoning will have arrived when spending more compute at inference time reliably buys a dramatically better answer.

Their wordsI think if we can achieve that amount of inference compute, where it leads to a dramatically better answer as you apply more inference compute, I think that will be the beginning of real reasoning breakthroughs.

Interesting
1:23:57 · watch on youtube.com↗

Our reading
Curiosity is the thing AI has not cracked: a model can research a question thoroughly, but it cannot decide which question is worth asking.

Their wordsWhat makes humans special though, is our curiosity. Even if AI’s cracked this, it’s us still asking them to go explore something. And one thing that I feel like AI’s haven’t cracked yet, is being naturally curious and coming up with interesting questions to understand the world and going and digging deeper about them.

Interesting
4 min later
1:27:47 · watch on youtube.com↗

Our reading
The dangerous concentration of power in AI comes from who can afford the compute, not from who holds the model weights.

Their wordsIt’s less about access to a model’s weights, it’s more access to compute that is putting the world in more concentration of power and few individuals. Because not everyone’s going to be able to afford this much amount of compute to answer the hardest questions.

Interesting
29 min later
1:56:44 · watch on youtube.com↗

Our reading
The way to stop a chatbot inventing things is to forbid it from saying anything it did not retrieve, not merely to hand it extra context.

Their wordsThe principle in Perplexity is you’re not supposed to say anything that you don’t retrieve, which is even more powerful than RAG because RAG just says, “Okay, use this additional context and write an answer.” But we say, “Don’t use anything more than that too.” That way we ensure a factual grounding.

Interesting
9 min later
2:05:38 · watch on youtube.com↗

Our reading
Vector embeddings will not solve search: a decades-old term-frequency algorithm still beats most of them at ranking.

Their wordsThere is an algorithm called BM25 precisely for this, which is a more sophisticated version of TF-IDF. TF-IDF is term frequency times inverse document frequency, a very old-school information retrieval system that just works actually really well even today. And BM25 is a more sophisticated version of that, that is still beating most embeddings on ranking.

Interesting
3 min later
2:08:56 · watch on youtube.com↗

Our reading
Once a language model reads the results, search can trade precision for recall, because the model does not care that the right link came ninth.

Their wordsIn Google, even though we call it 10 blue links, you get annoyed if you don’t even have the right link in the first three or four. The eye is so tuned to getting it right. LLMs are fine. You get the right link maybe in the 10th or ninth. You feed it in the model. It can still know that that was more relevant than the first.

Interesting
3 min later
2:11:51 · watch on youtube.com↗

Our reading
Owning the dominant model is not worth chasing, because no user has ever cared which model produced a good answer.

Their wordsDoes the user care if Perplexity has the most dominant model in order to come and use the product? No. Does the user care about a good answer? Yes. So whatever model is providing us the best answer, whether we fine-tuned it from somebody else’s base model or a model we host ourselves, it’s okay.

Interesting
9 min later
2:20:32 · watch on youtube.com↗

Our reading
If reasoning can be separated from facts, the frontier labs lose their advantage, because nobody needs a million-GPU cluster any more.

Their wordsBut any breakthrough can disrupt that. If you can decouple reasoning and facts and end up with much smaller models that can reason really well, you don’t need a million H100 equivalent cluster.

Interesting
1 min later
2:21:29 · watch on youtube.com↗

Our reading
A founder who picks an idea because the market wants it will give up, because nobody can persist at something that was never important to them.

Their wordsI think the thing that most people get wrong after they’ve decided to start a company is work on things they think the market wants. Not being passionate about any idea but thinking, okay, look, this is what will get me venture funding. This is what will get me revenue or customers. That’s what will get me venture funding. If you work from that perspective, I think you’ll give up beyond the point because it’s very hard to work towards something that was not truly important to you.

Interesting
27 min later
2:48:23 · watch on youtube.com↗

Our reading
Every announcement of a longer context window hides the cost, which is that the model gets worse at following instructions.

Their wordsSo most people, when they advertise new context window increase, they talk a lot about finding the needle in the haystack sort of evaluation metrics and less about whether there’s any degradation in the instruction following performance. So I think that’s where you need to make sure that throwing more information at a model doesn’t actually make it more confused.

Interesting
3 min later
2:51:45 · watch on youtube.com↗

Our reading
We will spend more of our working hours talking to AIs than to other people, but AI companionship should not be the answer to loneliness.

Their wordsLoneliness is one of the major problems in people. That said, I don’t want that to be the solution for humans seeking relationships and connections. I do see a world where we spend more time talking to AIs than other humans, at least for our work time.

Interesting