Dwarkesh Patel

I can recommend every Dwarkesh Patel pod and lecture with Sarah C. M. Paine. History is the best teacher, and that's her specialty. I learned a lot from each of the following: https://www.dwarkeshpatel.com/p/sarah-paine (WW2, Taiwan, Ukraine, & Maritime vs Continental Powers) https://www.dwarkeshpatel.com/p/sarah-paine-india https://www.dwarkeshpatel.com/p/sarah-paine-japan https://www.dwarkeshpatel.com/p/sarah-paine-china

He was on the Dwarkesh podcast last August to provide some lay person friendly synopsis and updates to “Who We Are.” Worth listening to even if you have read the book (in my mind at least). Warning, link has an auto play when I opened it (but don’t let that minor obnoxiousness dissuade you from listening). https://www.dwarkeshpatel.com/p/david-reich

Experimenting with creating semantic chunks of large podcasts. Got the following chunks, https://gist.github.com/nutanc/a9e6321649be5ea9806b4450b0bd6552 Dwarkesh has 18 splits. https://www.dwarkeshpatel.com/i/151435243/timestamps I got 171. So roughly 9 context discussions in one time stamp.

per Zuckerberg[0], ~half of their H100s were for Reels content recommendation: > I think it was because we were working on Reels. We always want to have enough capacity to build something that we can't quite see on the horizon yet. ... So let's order enough GPUs to do what we need to do on Reels and ranking content and feed. But let's also double that. So there's an immense capacity inside Meta, but the _whole_ fleet isn't available for LLM training. [0]: https://www.dwarkeshpatel.com/p/mark-zuckerberg?open=false#§transcript

Here is a great podcast interview with the post author on self-learning: https://www.dwarkeshpatel.com/p/andy-matuschak

I was reading an interesting interview[0] with Shane Legg of DeepMind where he mentioned reading Kurzweil's Age of Spiritual machines in 2001. I also read ASM around 2003-4 and just decided to give it a quick re-read to see how the predictions feel from 2023. As part of this re-read I'm checking in on things mentioned in the book, and I found this. If you haven't read ASM, or if it's been several years, I do recommend giving it a look. [0]: Shane Legg - 2028 AGI, New Architectures, Aligning Superhuman Models: https://www.dwarkeshpatel.com/p/shane-legg

>On the plus side, some software engineers save a couple minutes here and there with a slightly nicer autocomplete, white collar workers can get some advice on email style, and maybe it helps with sifting through millions of pages of legal discovery. Lol get a grip. Jeff Dean just said 25% of the code being checked into Google's internal codebase is now AI generated. If you think it's just kids using it for homework and engineers saving a few minutes here and there you're kidding yourself. https://www.dwarkeshpatel.com/p/jeff-dean-and-noam-shazeer >I am hopeful for some of the more advanced applications, e.g. proof assistance in math, "I expect, say, 2026-level AI, when used properly, will be a trustworthy co-author in mathematical research, and in many other fields as well." You just aren't in the same universe as what the top experts are saying about this tech. https://unlocked.microsoft.com/ai-anthology/terence-tao/ You seem like one of those people that used GPT3.5 once and hasn't updated at all since then. You're criticizing the wright flyer and we're already at turboprops.

> [T]he mythos of people having popped out from the same area or region that they their most recent ancestors have lived […] >> Genetic studies have put an end to that kind of speculation. Only crackpots and religious nuts are supporting alternatives. David Reich (author of "Who We Are and How We Got Here: Ancient DNA and the new science of the human past") says[^0]: " The modern human lineage, leading to the great majority of the ancestors of people today, was probably in sub-Saharan Africa for the last 500,000 years at least. It might be much more. Certainly our main lineage was in Africa, probably 3-7 million years ago. But in a period between about 2 million to 500,000 years ago, it's not at all clear where the main ancestors leading to modern humans were. There were humans throughout many parts of Eurasia and Africa with a parallel increase in brain size and not obviously closer ancestrality to modern humans in one place than in the other. It's not clear where the main lineages were. Maybe they were in both places and mixed to form the lineages that gave rise to people today. There's been an assumption where Africa's been at the center of everything for many millions of years. Certainly it's been absolutely central at many periods in human history. But in this key period when a lot of important changes happen—when modern humans develop from Homo habilis and Homo erectus all the way to Homo heidelbergensis and the shared ancestor of Neanderthals, modern humans, and…

I hope you'd at least concede without needing bulletproof evidence that it's a more reasonable possibility than the second coming of Jesus. The best evidence I can offer is the continued saturation of benchmarks, the continued effectiveness of scaling laws with respect to improvements in general reasoning, the massive improvements in algorithmic capabilities which lead to better use of existing models and inference-time compute, and the overall lack of fundamental blockers articulated that differentiate AI and human reasoning that can't be fixed via these developments. Every LLM of a given size is invariably at or near the state of the art with respect to all benchmarks available to test general reasoning. GPT-4 was the first model in the >1T parameter range, and every model that emerged around that level (Claude 3 Opus, Gemini Ultra) reached that level of performance in MMLU, MATH, etc. All subsequent improvements in models have been distillations and algorithmic improvements, synthetic data, etc. but so far a ~10T model has not been attempted, though it is likely around the scale for the upcoming GPT-5, Claude 4 and Gemini 2 models. Sources: https://www.dwarkeshpatel.com/p/will-scaling-work https://www.lesswrong.com/posts/arPCiWYdqNCaE7AQv/superintelligent-ai-… https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1…

Is this the transcript of the interview (podcast) with Dwarkesh? https://www.dwarkeshpatel.com/p/satya-nadella Because if so, > He doesn’t necessarily believe in fast takeoff AGI (just yet) the term "fast takeoff AGI" does not appear in the transcript.

Rhymes with another story (https://www.dwarkeshpatel.com/p/notes-on-china) on the front page > It’s funny how China has basically the inverse problem as America. We subsidize demand and restrict supply. They subsidize supply and restrict demand. We really should reconsider how we structure our policies

aha, found the https://www.dwarkeshpatel.com/i/151435243/transcript ; currently queued after vM but maybe I'll bump it to before?

Not a big fan of watching talking heads, personally, but Gwern is likely to have some interesting things to say. Patel's page for the episode: https://www.dwarkeshpatel.com/p/gwern-branwen there's a transcript, and links to Spotify and Apple (the Apple link doesn't work yet, though)

Likely submitted due to patio11 being on Dwarkesh: https://www.dwarkeshpatel.com/p/patrick-mckenzie One big takeaway, I think, is that expertise and ability are highly unevenly distributed, including within institutions: https://jakeseliger.com/2024/01/11/on-not-being-a-radical-medicine-ske…

As far as I (ex-ML researcher) know, the main technological case that LLM performance will hit a limit is due to the amount of text data available to train on is limited. The ways these scaling laws work is they require 10x or 100x quantity of data to see major improvements. This isn't necessarily going to limit it though. It's possible there are clever approaches to leverage much more data. This could either be through AI-generated data, other modalities (e.g. video) or another approach altogether. This is quite a good accessible post on both sides of this discussion: https://www.dwarkeshpatel.com/p/will-scaling-work

Zuckerberg said the reason he bought so many GPUs was to build Reels to compete with TikTok. Is that the big claim you're asking about? https://www.dwarkeshpatel.com/p/mark-zuckerberg

There are people who strongly believe data is a non-blocker possibility due to synthetic data that is of high quality. Dario Amodei from Anthropic. https://www.dwarkeshpatel.com/p/will-scaling-work

> I like science fiction too, but all of these potential scenarios seem so far removed from the low level realities of how these systems work. Today, yes. Nobody is saying GPT-3 or 4 or even 5 will cause this. None of the chatbots we have today will evolve to be the AGI that everyone is fearing. But when you go beyond that, it becomes difficult to ignore trend lines. Here's a detailed scenario breakdown of how it might come to be –https://www.dwarkeshpatel.com/p/carl-shulman

> I regard a physics education as a modern-day ‘liberal arts’ technical degree. Some anecdotal evidence for this: https://www.dwarkeshpatel.com/p/dario-amodei#details > Dario Amodei: We have generally found that if we hire someone who is a Physics PhD or something, that they can learn ML and contribute just very quickly in most cases.

Don't entirely disagree but I think "mad" is an unfair characterization and simplifies how the fascinating competitive dynamics have played out that resulted in the Llama models. Meta as well as Amazon and MS have been strategically constrained by not owning the mobile platform and they've been pretty consistently clear about that. https://x.com/ianand/status/1753425306181116394 Like it or not, Apple's ATT move (i.e. "Ask App Not to Track") hit their revenue in real dollars and stock price significantly. Not only that, the ensuing chaos helped pave an opening for the rise of competitor TikTok: > Meta, formerly known as Facebook, said that one setting alone cost the company an estimated $10 billion. Its stock value has plunged 70% this year. But ATT had another side effect, one that got far less attention than Meta’s troubles. Apple’s iPhone privacy setting gave TikTok a significant leg up in its fight for social media dominance. (from https://gizmodo.com/how-apple-s-ask-app-not-to-track-prompt-helped-tik… Their savior to improve monetization in the face of ATT turned out to be AI: > Further integration of AI helped drive Meta’s first revenue increase in three quarters, the company said on Wednesday. Reels monetization is up over 30% on Instagram and over 40% on Facebook on a quarterly basis as AI plays a larger role in the platforms. (from…

Self-imposed limitations are key to being effective. I was referring to pointless self imposed limitations. I was reading an interview with Andy Matuschak[1]. > One of the things that I think is kind of weird about this memory system stuff, or like memory champions, or something like that is “Oh, if you do these things, will you start to forget other normal human stuff?” And what's weird is, no. I've been doing this memory system stuff for years and I just know more stuff now. This is aligned with the experimental literature, which seems to suggest that, there's probably upper bounds but we're not close to them. Some of these memory champions have memorized maybe two orders of magnitude more things than I have practiced. Certainly people who are multi-lingual have really, really absurd numbers of things memorized. So there isn't a resource management argument. The notion that stuffing your brain with trivia will be damaging is merely an unwarranted fear. [1] https://www.dwarkeshpatel.com/p/andy-matuschak