Learning Library

← Back to Library

AI Safety Trends and France’s $100 B Fund

Key Points

  • Experts offered mixed opinions on AI safety over time, with some noting it’s becoming safer, especially due to growing open‑source initiatives.
  • This episode of *Mixture of Experts* will discuss test‑time scaling, Sam Altman’s latest blog post, and Anthropic’s new Economic Index.
  • The Paris AI Action Summit, organized by the French government, gathered civil society, industry, and policymakers to develop AI standards and guidelines.
  • President Macron announced a €100 billion AI fund—supplemented by private investors like Iliad’s €4 billion—to bolster sovereign European AI infrastructure.
  • Pleias, co‑founded by Anastasia Stasenko, will leverage this funding to help build and scale European AI capabilities.

Sections

Full Transcript

# AI Safety Trends and France’s $100 B Fund **Source:** [https://www.youtube.com/watch?v=KNNi-enb2B4](https://www.youtube.com/watch?v=KNNi-enb2B4) **Duration:** 00:39:45 ## Summary - Experts offered mixed opinions on AI safety over time, with some noting it’s becoming safer, especially due to growing open‑source initiatives. - This episode of *Mixture of Experts* will discuss test‑time scaling, Sam Altman’s latest blog post, and Anthropic’s new Economic Index. - The Paris AI Action Summit, organized by the French government, gathered civil society, industry, and policymakers to develop AI standards and guidelines. - President Macron announced a €100 billion AI fund—supplemented by private investors like Iliad’s €4 billion—to bolster sovereign European AI infrastructure. - Pleias, co‑founded by Anastasia Stasenko, will leverage this funding to help build and scale European AI capabilities. ## Sections - [00:00:00](https://www.youtube.com/watch?v=KNNi-enb2B4&t=0s) **AI Safety Trends and Perspectives** - Panelists quickly debate whether AI is getting safer over time, and the host previews upcoming topics—including test‑time scaling, Sam Altman’s blog, Anthropic’s Economic Index—and highlights the Paris AI Action Summit. - [00:03:09](https://www.youtube.com/watch?v=KNNi-enb2B4&t=189s) **GPU Limits, Open AI Funding** - The speakers discuss GPU shortages and environmental concerns while introducing the "Current AI" public‑good foundation, which aims to raise billions to support open‑data, open‑source AI models for broader, responsible applications. - [00:06:11](https://www.youtube.com/watch?v=KNNi-enb2B4&t=371s) **AI Safety, Democratization, and FOMO** - The speaker argues that leaders’ fear of missing out fuels rapid AI rollout, making safety increasingly complex as powerful models become more democratized, offering both advantages and unforeseen harms while regulatory frameworks lag. - [00:09:16](https://www.youtube.com/watch?v=KNNi-enb2B4&t=556s) **Building Foundations for Factual AI** - The speaker stresses prioritizing curated, multilingual databases and small retrieval‑augmented models to ensure safe, factual AI, noting current deficiencies in toxicity classifiers and the lack of AGI ambitions. - [00:12:20](https://www.youtube.com/watch?v=KNNi-enb2B4&t=740s) **Rapid Replication of Test‑Time Compute Trick** - The speaker notes that the swift, low‑cost recreation of a test‑time compute method—leveraging longer chain‑of‑thought generation, rejection sampling, and a single weight token—is anticipated rather than surprising. - [00:15:30](https://www.youtube.com/watch?v=KNNi-enb2B4&t=930s) **Prioritizing Quality Over Quantity** - The speakers contend that focusing on smaller, high‑quality, representative datasets—rather than massive noisy corpora—reduces computational waste and drives more effective model progress. - [00:18:36](https://www.youtube.com/watch?v=KNNi-enb2B4&t=1116s) **Legal Showdown of AI Models** - The speakers imagine a future courtroom where litigants use tiny versus gigantic AI models—favoring a pluralistic, capitalistic AI landscape over a single governing AGI—before transitioning to discuss Sam Altman’s recent “Three Observations” blog post. - [00:21:49](https://www.youtube.com/watch?v=KNNi-enb2B4&t=1309s) **Debating Hype vs. Realism** - A dialogue where one speaker questions exaggerated claims about a technology’s impact, while another expresses enthusiastic optimism about its world‑changing potential. - [00:24:52](https://www.youtube.com/watch?v=KNNi-enb2B4&t=1492s) **AI as Emancipatory Tool vs AGI Ambitions** - The speaker argues that AI assistants can democratize writing and coding for marginalized users, warning that the pursuit of AGI diverts focus from this empowerment and won’t solve looming ecological crises. - [00:28:00](https://www.youtube.com/watch?v=KNNi-enb2B4&t=1680s) **Rethinking AI Scaling and Energy** - The speaker argues that AI advancement should prioritize test‑time compute and consumer‑grade hardware, challenging the notion that massive data centers or nuclear power are required for scaling. - [00:31:08](https://www.youtube.com/watch?v=KNNi-enb2B4&t=1868s) **Anthropic Finds AI Use Still Tech‑Centric** - Anthropic’s analysis shows that about 36% of AI assistant usage remains concentrated in software development and technical writing, defying expectations of broader consumer applications. - [00:34:35](https://www.youtube.com/watch?v=KNNi-enb2B4&t=2075s) **AI’s Coding Dominance Over AGI** - The speakers argue that despite hype about AGI, current powerful AI systems like Claude are predominantly valued for coding assistance, making usage data appear skewed toward that niche. - [00:37:50](https://www.youtube.com/watch?v=KNNi-enb2B4&t=2270s) **AI as Platform, Not Just Tool** - The speaker argues that AI should evolve from isolated tools to integrated platform signals, emphasizing the need to discover more practical use‑cases and educate users before it becomes mainstream. ## Full Transcript
0:00Globally speaking, is AI getting 0:02more or less safe with time? 0:04Marina Danilevsky is a 0:05Senior Research Scientist. 0:07Marina, welcome back to the show. 0:08What do you think? 0:09Both. 0:10Okay, all right, we'll get into that. 0:12Uh, Chris Hay is a Distinguished Engineer 0:13and CTO of Customer Transformation. 0:16Uh, Chris, what do you think? 0:17More or less safe with time? 0:18I think it's getting more safe with time, 0:20so I'm, I'm pretty pleased with that. 0:21Um, and joining us for the very first 0:23time is Anastasia Stasenko, who is 0:25the CEO and co-founder of pleias. 0:28Uh, Anastasia, more or less safe? 0:30Uh, it's definitely getting safer 0:32because it's getting more open source. 0:34All right, great. 0:34More to dive into. 0:36All that and more on today's Mixture of Experts. 0:44Greetings from Paris. 0:45I'm Tim Hwang, and welcome 0:46to Mixture of Experts. 0:47Each week, MoE is the place to tune 0:49into to hear how leading researchers, 0:51engineers, entrepreneurs, technologists, 0:53and many more are thinking about the 0:55latest trends in artificial intelligence. 0:57As always, we have a lot to 0:59cover, way too much to cover. 1:01Um, we're going to talk 1:02about, uh, test time scaling. 1:03We're going to talk about a 1:04new blog post from Sam Altman. 1:06We'll talk about Anthropic's new Economic Index. 1:08But first, uh, given that I'm in Paris 1:10and Anastasia is in Paris as well, we want, 1:12to talk about the Paris AI Action Summit. 1:15So this is the latest of addition of a series 1:18of summits that governments been, have been 1:19holding, uh, around AI in the last few years. 1:22Uh, this year is hosted 1:23by the French government. 1:25Um, and it collects representatives 1:27from civil society, governments, 1:30companies and, and more, to talk 1:32about and set sort of standards and 1:34guidelines around the development of AI. 1:36And there are some really big 1:37announcements that we wanna get into. 1:39Anastasia, it's really 1:40exciting to have you on the show. 1:42In part because you were sort of 1:43directly involved in some of these events. 1:45I understand basically that 1:47Macron has announced a enormous, I 1:50think it's like a 100 billion 1:51dollars fund, to support AI and 1:54support AI specifically in France. 1:56And, I know pleias is, is part of 1:58that, but if you want to tell us just a 1:59little bit more about, how you got 2:01involved in and what you're going to 2:02be doing with this new fund. 2:05Yes, of course. 2:06Well, so first of all, there actually have 2:09been multiple announcements about investment. 2:12We do love announcing investment in France. 2:16We do hope that the real action will follow. 2:18However, what's important to say is 2:21that 109 billion investment fund is 2:26actually not only of France investing. 2:29This is, an international 2:32and, have quite, there are 2:35some private companies involved. 2:37For example, Iliad Group, which is a 2:40free, mobile company, basically, 2:44giving over 4 billion, et cetera. 2:47And this fund, has the objective 2:50to really focus on sovereign 2:53European AI infrastructure, right? 2:55So this is one part, of this and 2:59it's true that we have been, we have 3:02been listening, we have been hearing for 3:04a long time that Europe is lagging behind 3:07in terms of AI infrastructure, right? 3:09We don't have, enough GPUs to basically 3:13neither train, frontier models, 3:16nor to then run inference, for 3:21actually scaled AI applications, right? 3:24I'm not sure that this is entirely true, and 3:26I'm not even sure that we should be really going 3:29into scaling the AI infrastructure in the world 3:32where we actually have ecological imperatives 3:36kind of haunting us at the same time. 3:38another big one was the announcement 3:41of the Current AI, which is, 3:45actually AI for public good foundation. 3:49Yeah, I know and this is like a very 3:50particular part of pleais' work, right? 3:52Because I know you guys are specifically 3:54working on sort of open source 3:56and open data models, ultimately. 3:58Yes, totally and we have been, we have 4:02actually trained the world first models, 4:04on exclusively open data, open in the strong 4:07sense of this word without copyrighted 4:10material with permissive licenses only and 4:12it's true that opening data and creating 4:16this open data infrastructure is something 4:19that is important to us, but actually is 4:21important to larger AI communities, which 4:24is of today actually cannot advance as 4:27fast or work with as many applications for 4:31the good of the communities, which are 4:33Even think about more resource languages, 4:35think about more specific applications 4:38without actually being supported by 4:40the initiative, such as Current AI. 4:42So this foundation with already 400 4:46million secured for the first year 4:48aims to actually raise over $2 billion 4:53for their five year, five year run. 4:56At least as of now. 4:57And yeah, we are very happy to be 4:59part of this from the very beginning. 5:01We have signed the open letter with 10 5:04other industry leaders, such as Mistral, 5:08ALEPH ALPHA, Hugging Face, et cetera, et cetera. 5:11So it's all very exciting. 5:13And I'm most particularly excited about 5:15data finally not being like kind of 5:18the forgotten piece, of this AI, 5:21AI hype and like not only hype, but 5:24like, AI development in general. 5:26For sure. 5:27Yeah. 5:27And I think it's something that we talk a 5:28lot about on the show is just how much like 5:30the data piece often gets lost, even though. 5:33Arguably, it's like the most 5:34important part of the whole thing. 5:36One of the themes I did want to 5:38pick up on,you know, Politico 5:39did this interesting article. 5:41The headline I'll just read here was, 5:42quote, "How the world stopped worrying 5:44and loved to, learned to love AI."" 5:46And sort of the argument of the article 5:47was that, you know, the last few 5:49summits had really focused on, like, 5:50safety and security around the model. 5:53But in contrast, I think this year, 5:55there was just a lot more kind of lean 5:56forward, like, we just need to deploy 5:58this faster and better and bigger. 6:00Than ever before and I guess Marina, 6:02maybe I'll call on you because I know in 6:03response to the first question you said, 6:05well, is it getting safer or less safe? 6:07Well, maybe a little bit of both. 6:08I'm curious about kind of what you meant 6:09by that and how you think about it. 6:11Um, particularly in this context where it 6:12feels like at least kind of world leaders 6:14are like rah, rah, in a way that maybe 6:16they haven't been so much in the past. 6:17World leaders have FOMO. 6:18So they're saying, well, 6:19safety is all well and good. 6:21Everybody else is working on this, so 6:22we better go ahead and work on this too. 6:24And why I say both is because there are 6:26aspects in which the safety is getting better. 6:28There are aspects in which the power of 6:30the models makes it easier, again, to do 6:33potentially more deeply nuanced and misleading. 6:36Tasks. And so they're, the safety is 6:39just sort of getting, I think, more complicated. 6:41So less the model is going to teach you to build 6:43a bomb and more the model can still potentially 6:46over time, depending on how it's used, 6:48have, maybe not effects that you intend. 6:52I mean, always when you have 6:52technology, you let it out and you 6:55can't always control where it goes. 6:57It's going to go where it goes. 6:58So this has always been to me, the 7:00flip coin of the democratization of AI. 7:03Having it in more hands is going 7:04to be always at the same time. 7:05Good. 7:06And bad. 7:06And I guess, Chris, this is a good 7:07chance to kind of bring you in as well. 7:08I know you were playing the voice of kind 7:10of optimism here, saying, well, overall, 7:12things are getting, in fact, safer, 7:14and you know, I guess I had maybe one 7:16way of nuancing this question a little 7:18bit is to talk about kind of like at what 7:19layers it's getting more or less safe. 7:21It seems to me at least kind of on the 7:23regulatory side, there's been more of a push 7:26certainly to kind of say, well, look, we 7:28don't want to restrict to this technology. 7:30Let's leave it more open. 7:31Which I think certainly 7:32some people think is like, 7:34less safe, right? 7:35But I think there's also people 7:36saying, look, our techniques around 7:37safety are getting a lot stronger as well. 7:39Is that kind of how you think about it? 7:41Like the argument for why it's safer? 7:42I think so. 7:43I mean, if I think of the models that, if we even 7:45just go back two years ago, right, think of like 7:48the GPT, version 3 at that point, or if 7:52you think of the early open source models, like. 7:54The kind of remember LlaMA one, et cetera. 7:56Now, if we started to think about that, 7:58I mean, come on, it's like those models 8:01were terrible in comparison to today's 8:04models, today's models are much more 8:05safer, they come back with better answers. 8:08They hallucinate less. 8:09And then if we think about things from a stack 8:11perspective, we've now got the guard models, 8:13we were talking about that last week, you 8:15know, the idea of being able to reduce bias, 8:17how we train the models are a lot better. 8:19So I think in general, we're thinking about 8:21this a lot more now, that's not, that 8:25we are not going to be much safer in the 8:28future and we haven't got a long way to go. 8:30And we can't do bad things with models today. 8:32Of course you can, but then the way I like 8:34to think about it is, my friend used 8:37to write test code for missiles and he 8:40was like the worst programmer I ever met. 8:42So I'm like, would I even want an old Llama 8:46model doing that versus him writing that code? I, you know what? 8:49I'm like, I think actually maybe it's safer. 8:51Yeah, it's fine. 8:52It's fine. 8:55Anastasia, as, as kind of a 8:57model developer in the space, how do, 8:59how's play us thinking about safety? 9:01Do you feel it's kind of like distinctive 9:02a little bit from what you see elsewhere? 9:04I know kind of philosophically, you're 9:05very, you know, kind of focused on 9:07open, but kind of curious about 9:08your, if you're also trying to kind of, 9:10blaze a new trail in the safety space as well. 9:12For us, we, we don't develop, 9:14conversational models. 9:16We don't develop, chatbots. 9:19We do really specialize, in the models for 9:22both data processing and data processing until 9:26it goes to the retrieval augmented generation. 9:28So basically for us, the most important 9:30part when it comes to safety is actually 9:33the development of the curated and vetted 9:37databases, which are prepared well, 9:39for them to be used for factual AI. 9:44And, this is where actually not that 9:47much work is done, nowadays, 9:51because you still, for example, do not have 9:54Good multilingual, and I insist on 9:57the word multilingual, classifiers, 10:00even for sentiment analysis. 10:01You don't have good toxicity classifiers, 10:05which you can actually understand 10:07what data they have been trained on, and 10:09why we say this is toxic and this is not. 10:12I mean, there is so much work to be done 10:14to actually prepare good data foundations 10:17for factual AI, and where you can actually 10:20have more bound models to the data, to 10:23the even proprietary data, which are the 10:25open data that you use in your stack. 10:28And this is where we concentrate our efforts. 10:30We are not building AGI. 10:32We don't have resources for, and I'm not even 10:34sure we do need to, for multiple reasons. 10:37But we do need this working horses, these 10:40small models, which allow to work through 10:43data, which is not good now, but can be 10:46brought to the, to the quality, where 10:50this kind of RAG applications or whatever 10:52technology will be, brought afterwards. 10:54So this kind of live actual 10:56AI, would be deployed. 10:58I was just going to say, I know a guy who's 11:00got 109 billion dollars that he might be 11:02able to help you build AGI, Anastasia. 11:04You might want to tap him up for it. 11:06We will get to that gentleman 11:08a little bit later in the show. 11:14One of the things I do want to talk about 11:15is that there's, the paper of the week, 11:17the flavor of the week, was this, 11:20Simple test-time scaling paper that I feel 11:22like a lot of people have been talking about. 11:25I'll kind of sketch the overview, if 11:27you have been watching, of course, o1 preview 11:29came out and one of OpenAI's kind 11:33of stated sort of advancements in the model, 11:36was really the idea of test time compute. 11:38Kind of the idea that you'll take 11:39a model, you'll get it to basically 11:40think harder, and it's able to achieve 11:43much better results as, as a result. 11:45And this paper came out, saying, 11:48look, we are trying out this technique that 11:50we call s1, where in order to try to replicate 11:54sort of o1 preview's reasoning ability. 11:56We collected about a thousand questions 11:59and they're kind of reasoning traces, 12:01and we use a couple of different hacks I 12:03think one of my favorite hacks is one where 12:05they just insert the token weight to 12:07get the model to keep, you know, thinking 12:09about a problem, rather than stopping. 12:11And they say look with all these kind 12:12of pretty cheap like rough and ready hacks, 12:15we're able to get a model that's right up 12:17there and in competitive with o1 preview. 12:20So kind of a shocking result, I guess 12:23in some sense, but maybe actually Chris, 12:25I see you're already going off mute. 12:26Like, is it that shocking that 12:27people can just do this replication? 12:29Because I don't, this is going 12:30to be the new thing, right? 12:31Like test time computes 12:31going to change everything. 12:33It's going to be the new whiz bang technology. 12:34And then these people, these researchers 12:37have just replicated it in basically no 12:39time and at much lower cost, apparently. 12:41I think it's the "How I Met Your 12:43Mother," Barney Stinson method. 12:45What's 25 + 2? 12:46Wait for it, wait for it, 27. 12:49You know, and that, that's the 12:50basic technique, as you said there. 12:51And, no, it's not a surprise, right? 12:54We kind of already know there from the model 12:57that the longer that it spends thinking about 12:59it and being able to generate more tokens, then 13:01it's going to have the opportunity to reflect. 13:03And we saw that in the DeepSeek paper, right? 13:05I mean, ultimately, the trick underneath 13:06that was to create multiple samples, take 13:09longer, get longer chain of thoughts. 13:10And once you have longer chain of thoughts, then 13:12the model is more likely to get to the answer. 13:14And that's effectively what they're doing there. 13:16And then they're essentially rejection 13:18sampling, anything that has sort 13:20of bad chain of thoughts there. 13:21Get rid of that. 13:22And, and therefore you're going 13:23to end up with a quality day. 13:24So I don't think this is a surprise, 13:27but it's really cool that it 13:28works with just one token, right? 13:30Which is the weight token. 13:31And, and then it generates 13:33that chain of thought. 13:33I didn't, I think if I'm honest, in 13:37the same way as the kind of step by 13:39step thing in about two years time. 13:42Probably a lesson that we're, we're not 13:43going to do these little hacks anymore, 13:45because what we're going to do is we are 13:47going to have a good set of kind of, 13:50cold start, chain of thought data set to 13:54be able to bootstrap the model with anyway, 13:56and therefore going along and saying, wait, 13:58wait, wait, or whatever, isn't going to have 14:01that effect because the model's going to be 14:03producing the correct chain of thoughts 14:04in the first place, but I think from 14:06this kind of starting with a kind of 14:09relatively simple and small base model. 14:11I think it was the coin two five base model they 14:13use but be able to sort of generate, uh, those 14:16chain of thoughts very, very quickly and get to 14:20sort of decent performance on that domain. 14:22I think great job, right? 14:24But it's really just building upon the work 14:26that kind of everybody's seen with DeepSeek. 14:28So yeah, great job from them 14:30Yeah, for sure. 14:31I think that's like, like what I like 14:32about the solution of the wait, wait, wait, 14:33it's like, it's just like classy solution. 14:35It's like very simple, but really nails it. 14:37Marina, one of the questions I have 14:39kind of like reading this paper is just 14:42like, how far can the test time compute go? 14:45Cause it kind of feels like one of 14:46the remarkable things is you take 14:48arguably less sophisticated models and 14:50you just get way better performance. 14:53Um, and you know, I guess there's kind of a 14:55question just like how far that can go or if 14:57like basically, you know, your base level model 15:00just sets a ceiling at some point on how far 15:02you can kind of reason up and be kind of at 15:04parity with much more sophisticated models. 15:06Are you sort of optimistic that like essentially 15:09test time compute will take us very, very far 15:10or is this kind of just like it's sort of a 15:12hack at the margin for some of these things? 15:14So first I'll say that yeah, it 15:16might be a less sophisticated model. 15:17It's more sophisticated data. 15:19So they didn't take a thousand data points. 15:21They took fifty nine thousand data points and 15:23there was a whole bunch of different filtering 15:26and qualifying and stratifying and the rest 15:28of it that got them down to that thousand. 15:30They spent some time talking about 15:31quality, difficulty, diversity. 15:33Listen, couldn't agree more. 15:34Okay, because everything at that point in 15:36time, the work is going to be done somewhere. 15:37If it's not going to be done in the 15:39model, having to deal with the noise of a 15:40whole bunch of data points, instead it's 15:42going to be done with these are really, 15:44really good representative data points. 15:46And, like Chris said, wait sounds 15:48like, oh, let's think step by step 15:49when that little trick was introduced. 15:51There's other ways to do this. 15:52You can have, again, examples that 15:54where you take through the thinking 15:55of, well, as you go through this model, 15:56you could say, oh, we tried this way. 15:58It didn't, that didn't work. 15:59Let's back up. 16:00This consistently reminds me of taking my 16:03nine year old through his math problems, 16:05where it's like, first you try this. 16:06Oh, it seems like this is not working. 16:08Why don't you back up? 16:08Like, these are the kind of tokens 16:10that you go through and go through. 16:11And the important thing again 16:12here is what data is being used. 16:15It's good that people are 16:16trying to get the compute down. 16:17This is a positive thing because I 16:18think right now we're still in a land 16:20of a really huge amount of waste. 16:22You do not need things that big with 16:25data that is that much and that noisy. 16:27Quality goes a really long way. 16:30So I think the more we continue to focus on the 16:32quality and the type, of data that's being used 16:35here, the more progress we're going to make. 16:37This is all very interesting in the context 16:38of like openness versus closeness in data. 16:41Cause I feel like one of the arguments 16:43that I've heard from some people as well, we 16:45just need so much data that it's impossible 16:48for us to figure out what's open and closed 16:50and we just need to be able to like use it. 16:52Totally. 16:52And you don't need that. 16:55You don't need the vast amount. 16:58You need vast amounts of data, but you 17:00first of all, you need good quality data. 17:03Reason in rich data. 17:04However, what is really interesting in 17:07this moment with well after post DeepSeek 17:10moment and test time, compute, et cetera, is 17:13that we are seeing that we can actually boost 17:15smaller models, which will have smaller impact, 17:20actually, in terms of energy, et cetera. 17:22and we can also boost them for specific 17:26domain reasoning, and it has 17:29been has been happening specifically 17:30for math and coding. And at pleias 17:33as of now, we are actually start 17:35we have started to work on 17:37this for legal reasoning. 17:39And these are the domains 17:42where you actually do have truth. 17:44You can create a chain of thought. 17:46You can create the verifiers, et cetera. 17:49But those data sets are a little 17:52bit more complicated to create 17:54than the coding and the math ones. 17:56And we have been experimenting with 17:58this for legal reasoning, reasoning over 18:02administrative documents, as well as even 18:05sociological reasoning, because you actually 18:07have, you can have quite clear guidelines 18:11that you can depart from, but all of these 18:13things are really, let's say, emerging at 18:16this point, and I'm really look, I'm very 18:18much interested in how it will help to boost 18:20smaller, specialized model for, let's say 18:24industry and for specific domains outside of 18:27general reason and capabilities, uh, which are 18:30tested on traditional benchmarks, uh, basically 18:34mass coding on all these kinds of things. 18:36And 18:36I think it's like the future is, 18:38is really in these small models. 18:40Yeah, I was going to say, I've now imagined 18:42this new future where people will be 18:44like, Oh my goodness, they're going to 18:46court and they're only bringing their 18:47two and a half billion parameter model. 18:49That model is up against a lawyer 18:51with a 70 billion parameter model. 18:53And then there'll be like, "Oh no, the 18:54judges got a 405 billion parameter model."" 18:57And then the witnesses, "Oh no, that 18:59is a 3 billion parameter witness." 19:00Is this our legal future where 19:02we're going to be having small 19:04models versus large models in court? 19:06That I find that future interesting. 19:07Actually for me, this is a more desirable 19:09future than being governed by one AGI. 19:12I mean, this is like at least, 19:14you know, you have to— 19:16it's kind of more capitalistic 19:18future that we are discussing here. 19:20Smaller models accompanying some, 19:23bigger models accompanying others. 19:25I'm, I'm not sure that I want to live in the 19:28one AGI world, but, that's probably just me. 19:31Maybe that'll be our next like hot take 19:32question at the top of next episode. 19:39The next topic we will bring up. 19:41is a blog post that came out 19:43from Sam Altman this week, 19:44simply entitled "Three Observations". 19:47It's a provocative blog post, and 19:49I figured it would just be sort of 19:50interesting to raise and talk about. 19:52the effect of the blog post, I think, 19:54is to talk a little bit about the economic 19:56impact, that we expect to see as 19:58AI systems get more and more powerful. 20:00And in effect, I think Sam makes 20:02these kind of two big arguments. 20:03You know, the first one is that 20:05we see model performance scaling. 20:07Right, so the bigger we get with models, 20:09the better they are, which has kind of 20:10been a longstanding kind of article of 20:12faith in the machine learning community. 20:14And then the second bit is that, like, as the 20:16costs of delivering these models drop, we're 20:19just seeing sort of demand keep increasing. 20:22and sort of his ultimate argument is, 20:24look, we should keep scaling, things will 20:27get cheaper, and that will ultimately have a 20:29sort of gigantic impact, on the economy. 20:31And, you know, in effect, it's kind of 20:33a case for why people should continue 20:35believing in open AI, in some ways, 20:37because I think this is really, at the 20:39core, their sort of value proposition. 20:42and I want to get this kind of 20:43group's take as kind of both folks who 20:45really do believe these technologies are 20:47going to get a lot better in the coming 20:48years, but also I think has tended to be. 20:51You know, I would say AGI skeptics 20:52overall about where all this is going. 20:54and I guess maybe, Marina, 20:56I'll pick on you first. 20:57Curious about kind of what you 20:58thought about the blog post. 20:58Do you sort of agree with the argument? 21:00If you got quibbles with it, kind of 21:01curious about where you felt there 21:02was like problems in the logic. 21:04And maybe that sigh tells us 21:05everything we need to know. 21:06Sam's, 21:07main, the main point that I 21:09cottoned onto, there was a couple. 21:10So first of all, his point three. 21:12"The socioeconomic value of linearly increasing 21:15intelligence is super exponential in nature. 21:17We see no reason for exponentially increasing 21:20investment to stop in the near future." 21:22Give me more money. 21:23Money to me. 21:24Give it. 21:25Give it now. 21:26More to me money. 21:27Seems to be most of the message here. 21:29and then also, I don't think that he, 21:33along with very often a lot of other folks 21:35in Silicon Valley, live in the real world. 21:37Very often, because when they're making 21:39statements like, "Hey, in 10 years, 21:41everybody is gonna want to and be able 21:43to and get benefit from accessing the 21:45AI that only some people can access now." 21:48No. 21:49Things don't go in that kind of a scale, and 21:51that is not the kind of a need that people have 21:54and the benefit that people are going to have. 21:56There's a real specific perspective that he 21:58has, and again, I think that that is more 22:01narrow and more limited, and in some ways, a 22:04little bit, off putting, at least to me. 22:07Again, I'd like people to understand what 22:09the benefits and the use of this technology 22:11are without making statements like this, 22:14which I feel like undermine the work that 22:15a lot of us actually do in the field. 22:18Yeah, and so if I have it right, it 22:19seems like part of the critique is 22:20just that it's like an overstatement. 22:22Yeah, I think so. 22:22I mean, Chris, what do you think? 22:24I love it. 22:25I love Sam Altman. 22:27Go for it, Sam. 22:28It's like, it's gonna change the world. 22:30It's, it's great. 22:32You know, you know, I think, I think 22:36everybody's right I think you gotta have a 22:38super positive attitude in this sense, right? 22:40Which is this is going to be a world changing 22:42technology and and we can see that 22:45from how good things have improved over 22:48time. So it's gonna have an impact. 22:50Of course, it's gonna have an impact, right? 22:51But then new value creation is going to happen. 22:54We're going to do new ways of doing things. 22:56I think that's great. So I think it's 22:58healthy to talk about, you know, 23:01what the impact is going to be. 23:03Is it likely to be somewhere in between? 23:05Maybe because you know what, as soon as 23:06we get something that's super cool, we 23:08then just take it for granted as normal. 23:10And then we move on from there. 23:12I mean, I can guarantee I 23:15I've said this for a while. 23:16I think as soon as we get AGI and we can, 23:18we can define what an AGI is later, but 23:21I think, but I, but I think the first 23:23thing that's going to happen is it's 23:24going to be put in a box and then there's 23:26going to be a big museum open somewhere. 23:28And we're all going to walk in 23:29and go, you can chat with the AGI. 23:31And we'll be like,"" Ooh, there's the AGI."" 23:33So that thing's going to 23:34be in a box for a while. 23:35That's, that's his future, 23:37like a kind of circus. 23:38And, and, but, but honestly, I, 23:40I am super positive about things. 23:42I think that this is a 23:44world changing technology. 23:45Is it going to be like a 23:46conscious thing around that? 23:47No. 23:48But if we look at things 23:49like coding, the reality is 23:52that, even if you're using kind of like the 23:54o1 models or the o3 mini models 23:56today, they are better for like a lot of 24:00tasks, you know, to be able to turn out that 24:03code really quickly at a super high quality. 24:05And, and that is a reality. 24:07And if we think it's going to stay in that 24:08domain, I think we're kidding ourselves, right? 24:11As, as the cost comes down, more and 24:13more people are going to use this. 24:14Yeah. And this is kind of, I'm trying to 24:15parse optimisms, I guess, in some sense. 24:17Right. Cause I think like on one hand, Marina, 24:19I agree with you, which is like the 24:21blog post is very frustrating, like 24:22the tone is very frustrating to me. 24:24On the other hand, it's like, okay, 24:26Tim, but do you believe the technology 24:27will have like a really big impact? 24:29I'm like, oh, yeah, for sure. 24:30I think it will have a really big impact. 24:31And it's like, kind of like parsing, 24:33like, how do you articulate a way of being 24:35optimistic about this stuff that doesn't 24:37kind of fall into the usual valley tropes 24:39seems to be like part of the problem. 24:41Not being in the valley. 24:42I, I mean, for me, just, just 24:45to probably continue with, 24:47with the optimism. 24:49Uh, the ChatGPT moment was 24:50like a huge moment of liberation. 24:52I am. 24:53I don't like to write like, I mean, and 24:57we don't sometimes understand the, the, 25:00the emancipation liberating power of this 25:03technologies for people who actually like 25:06who had to really struggle with, let's say, 25:10writing, with coding, who didn't have access to 25:12this, who had less this intelligence, who use 25:15this as tools now to actually do something 25:18they weren't able to use before to do before. 25:22I, however, this being said, I'm not sure 25:26that, this emancipating power is really 25:30well understood and put in the place where 25:33it should be put because we are kind of going 25:35and hunting, the AGI where actually we 25:39do need like this assitants helping us 25:42to be like to be what we actually can be, 25:47in the society with less constraints on like 25:51how you can learn code, to code, how you can 25:53learn to write well, and how not to be judged 25:56by the society and by the economic system on the 25:59on the sheer amount of autographic errors that 26:02you are making, you know, for me, this is 26:05like more important than like hunting the AGI. 26:08However, I think that and Marina, 26:11you actually brought up this 26:13third point of exponential growth. 26:16The planet is limited. 26:19We do have limited resources and I'm very 26:23much surprised to to to read such takes 26:26because AGI won't resolve the issues of 26:31ecological crisis, which we are going through. 26:34We cannot and we won't have time to build enough 26:39nuclear power plants to actually 26:42run everything on nuclear energy. 26:45And even with nuclear energy, it won't help 26:48us to to resolve the ecological crisis. 26:50So I'm sorry to say this. 26:52We it's for me like and probably this is 26:55a bit of like European perspective, an old 26:58fashioned way of saying this, we cannot 27:02just forget everything we were saying just 27:05two years before, two years before, 27:09and like, okay, AGI will solve everything. 27:13It won't solve everything just like 27:14Internet didn't solve everything. 27:16You still have lots of places in multiple 27:19continents without Internet access 27:21and we still haven't resolved multiple 27:23issues which won't be resolved with AGI. 27:25So, I mean, we, I really do feel that we need 27:29to be more pragmatic and this is also important 27:33because Sam does this, and I actually heard him 27:37do the same speech at the Élysée Palace in front of 27:41Emmanuel Macron on Tuesday, where he basically 27:44said the same and saying, we need to invest. 27:46Let's invest in, in data centers. 27:50Let's, and I mean, we need to be careful 27:52about this calls for investment and how 27:55they will actually really impact the society 27:58without saying we don't need this technology. 28:00We do need it. 28:01We need to develop it, but in a reasonable way. 28:04Which is not building a data center 28:07near every school, you know, so 28:09I think I agree but disagree a lot and I think 28:13the reason is that I don't think we need a 28:17nuclear power stations as you say, I think I 28:20know Macron's suffering is nuclear power stations 28:22for everybody as part of his 109 billion but 28:25but but the reality is with test time compute. 28:28We should be spending less time focusing 28:30on the pre train, pre training stage. 28:31Now that's not to say that we're not going 28:33to be pre training, but it, it shouldn't 28:35be like pre train, pre train, pre train. 28:36With test time compute, you can get very, 28:39very far as we were just talking earlier in 28:41the podcast with the, like the s1 models. 28:43By just using high quality data sets. 28:46So, but being able to push that to longer 28:48chain of thoughts, being able to push that onto 28:50consumer grade hardware, I think we are already 28:53proving that scaling can occur at a lower cost. 28:56So I think if we're needing nuclear power 28:58stations to be able to, to, scale AI, 29:03then I think we're on the wrong track there. 29:06But the question is inference, 29:07not pre training only. 29:08So, it's a much bigger, 29:11bigger chain, for energy. 29:14And this is, yeah... 29:15But I can run inference 29:16I can run inference on my laptop, 29:18right? With my Apple silicon chips. 29:20It's fine. 29:21The cost of the cost of inference is much, 29:23much lower than, than the cost of training. 29:27But once we do scale AI applications, 29:29it will also be a question, however. 29:31But let's agree to disagree. 29:33That's not how this podcast works. 29:34Yes, 29:35it's not. 29:37I'll say something that might be funny 29:40coming from my background as a language 29:42primary study person, which is we should not 29:45forget the power of these models in other 29:49non language domains, multimodality, sensors, 29:53time series, all sorts of kinds of things. 29:54There is so much use to be gotten out 29:56of these, not only in helping you write, 29:58which no, I love it too, I hate writing 30:00emails, not only in helping you write 30:02code, but also imagine all of the ways in 30:05which you could improve, factory work. 30:07Imagine all the ways in which you could 30:08improve tracking the sensors that we 30:10put on, migratory animals to see if 30:12we can help, you know, their habitats 30:14out, all sorts of things like that. 30:16If we really actually broaden where this 30:18technology could be thought of being used, 30:20because again, remember, this technology 30:22actually has nothing to do with language. 30:23Let's just remember that. 30:24It doesn't. 30:25And then I think that we could actually go 30:28in places that are much more interesting, much 30:30more pragmatic, much more practical, and 30:33yeah, maybe not only focus on, on the parts that 30:36are language and certainly not on chasing AGI. 30:44Last item I really wanted to touch 30:46on was just a kind of fun sort of 30:48data set, that Anthropic put out. 30:50they're calling it the 30:51Anthropic Economic Index. 30:53And what I like about it so much is, if 30:55you remember from the early days of Google, 30:57they had a project they did called Google 30:58Flu, which is basically using people's, 31:01search results to try to identify where people 31:03were getting sick and creating basically 31:05like a live heat map of like where illness 31:07was kind of spreading around the population. 31:09And this in some ways is kind of like a weird 31:11update of it in some ways, only it's really 31:13looking rather than at people getting sick. 31:15What people are even using AI for. 31:17And so Anthropic basically looked across 31:20a sampling of all of its conversations, 31:22and using an anonymized set of them said, 31:24okay, well, what can we learn about how AI 31:27is kind of spreading across the economy? 31:29And they reached some 31:30pretty interesting results. 31:31I recommend going on the website and looking 31:33Um, the blog posts that they did. 31:36The one that I kind of wanted to touch on, 31:37because we don't have too much time left in the 31:39episode, is really this finding that right now. 31:42about 36%, of usage of AI 31:47assistants is really still in software 31:49development and technical writing tasks. 31:51And I think this is almost a 31:53kind of Marina's point, which is that 31:54we've kind of sold this technology. 31:55I thought about this technology as this 31:57kind of economy spanning thing, but one of 31:59Anthropic's findings is that it still ends up 32:01being quite concentrated what these people, 32:02what people are actually using it, for. 32:05And outside of that, it's 32:07kind of a very, very thing. 32:09And so we just want to first kind of 32:10talk a little bit about that result 32:12and happy to talk about any of the 32:13other data that they kind of mentioned. 32:15But I think that finding was just so 32:16interesting because it, at least for me, 32:17was kind of a violation of my expectations. 32:19I was like, oh, people are using it for 32:21writing emails and composing poems and 32:23essays and, you know, all this other stuff. 32:26But still, it's ultimately very 32:28much in It's like a software tool. 32:31I guess, Anastasia, do you 32:32think we should be surprised or am 32:34I just like, kind of not with it? 32:36It's actually, well, this study 32:39corroborates what we have been also seen, 32:42in other studies as well as, well, the 32:44one that we have conducted for the French 32:47government based on the, well, the 32:49usages, that have been done with the, 32:51with the copilot we developed for them. 32:53I think that. 32:56One of the reasons is when I actually, for 32:59example, see how far from software development 33:03and even from, let's say, application, 33:06you tasks and marketing, et cetera, for 33:09example, financial analysis and things like 33:12this are, I really do think that, once 33:16again, this, the tools that we're, that we 33:19have now, this chatbots, they're not really, 33:22they're not really adapted to the, to the 33:25exact knowledge work, that, most that 33:30people in other industries do have, for 33:33the software development, I wasn't surprised. 33:35And it's true that also Anthropic has been 33:37largely marketed as the state of the art 33:39coding tool. 33:41So it's kind of normal to see this. 33:43I think it's and they say 33:44it in the paper themselves. 33:45So this could be a little 33:47bit biased because of this. 33:49However, I really do think that we this 33:53study shows how actually far we are from a 33:57wider adoption of, large language models 34:02as everyday tools at work, which is 34:06still surprising I mean we could think that 34:09yeah, we all use it now we actually don't 34:12and there are some parts of the of the 34:15population that are much more exposed to this 34:18and It will require also quite an important 34:22part of education on workplaces, actually 34:26to the people who are using it now less, as 34:29well as UX and like other product adaptations 34:33from the model providers, of course, as well. 34:35Chris, I guess this almost builds a little bit 34:36on, I think you were making a joke earlier, 34:38which is like, we're going to invent the AGI and 34:40it will just kind of live in the zoo for a bit. 34:42Like this is literally it, right? 34:43Like not AGI, but like we have like 34:45really powerful AI systems, but it's like 34:47still largely like a technical industry phenomenon. 34:51I'm sure, I mean, Anthropic would look at this 34:52data and say, we have so much more market. 34:54We can grow so much more. 34:56I guess the pessimist view is like, 34:58well, maybe ultimately this thing is most 35:00useful and will continue to be most useful 35:02in stuff like coding. 35:04Is that a concern that you 35:05think for the AI industry? 35:06No, I think I'm gonna agree with Anastasia 35:10on this one which is that I think if you 35:14asked Moët & Chandon, what the primary 35:18use of glass bottles are, I'm pretty sure 30 35:21percent of them is gonna say champagne, right? 35:23And I think that is the reality for 35:26Anthropic as Anastasia said, right? 35:29What Claude has marketed is 35:30this, the best at coding. 35:33If you think of the ecosystem, so if you think 35:35of things like Klein, et cetera, the default 35:37models that they put in there is Claude, right? 35:40So, you know, so anybody in that industry 35:42knows that, you know, you know, up until the 35:45reasoning models point of view, you know, you 35:48would typically go to Claude for code intel. 35:50So I really think that data 35:53is skewed, as you say there. 35:54So I, I think that's probably where it is. 35:59So if we asked OpenAI what the primary uses are, 36:03I think you would get a different result set. 36:07I think it would be a different 36:08variation, just because that's kind of 36:10more aimed at a wider consumer base. 36:12And I would even, I would even 36:15probably argue that the, GPT-4o mini 36:18versus the o1-mini 36:21versus the o3-mini 36:22the o1 36:24pro would have a completely different, 36:27usage set of data there as well. 36:30So I, I, I think it's interesting that 36:33they've came back with it, but I, I, I just 36:35don't know how to read into those answers 36:38just because it's, it's really a very 36:40thin vertical slice, in my opinion. 36:42Yeah. All it's telling you about is like 36:44what Anthropic's being used for. 36:46Yeah, exactly. 36:47I don't, I don't think it's 36:49representative of the world. 36:51I don't think it's representative of America. 36:54I don't think it's representative of, of, of UK. 36:56It, to your point, Tim, it's representative 36:59of how Claude and Anthropic is used. 37:02You know, which is super interesting, 37:04but yeah, that's what it is. 37:06I mean, they explicitly say that, they 37:09freely admit it and explicitly say it. 37:10They say that one of the biggest points of 37:11releasing this dataset is to release this 37:13dataset and to hope that we can get something 37:16from other people that's kind of similar. 37:17It's like releasing search logs. 37:19You're never going to get all of 37:19them, but it's nice to get something. 37:22I don't know. 37:22I liked the economic perspective. 37:25They were very careful with stating 37:27their limitations or their assumptions 37:29that you shouldn't read into it. 37:30And it's more about the process of the analysis 37:32than looking at the results of the analysis. 37:34I agree with that. 37:34Yeah. And I think that's actually, Marina, if I could 37:37kind of follow up question there, I think one 37:38of the ones that I had to kind of talk about 37:40with the panel was, this is cool because I think 37:42to date we really haven't seen these companies 37:44say based on all the aggregate data we have, 37:48there's useful things we can build on top of it. 37:50and I think this is kind of is like 37:53a sort of new start in some ways, right? 37:55Rather than them simply saying, oh, 37:56we provide an AI tool you can use. 37:58This is like our platform now gives 37:59a signal about the world at large. 38:01I guess I had a question for you 38:03is like, do you think companies are 38:04going to do more of that going forwards? 38:06Or if this is kind of like, well, more of 38:07a demo project more than anything else? 38:09I don't know. 38:09I think this technology is still a 38:10hammer in search of a lot more nails. 38:12We've gotten a couple of nails, but there's 38:14a lot more out there and given the investment 38:15that people have put in and are apparently 38:17wanting to continue to put in, it'd be nice 38:19to find some more nails, um, and go out and 38:22try to figure out from people like, hey, do 38:24you even know how to use this technology 38:26or do you not know? 38:27Because I don't think yet we're at the point 38:29where people have an knowledge of how to use it. 38:31When we first started with the internet, people 38:33didn't quite know how to use it correctly. 38:35People got better and you can 38:36make endless examples here. 38:37So we haven't gotten there yet. 38:39I think it's actually to company's benefit 38:41to continue to get people more comfortable 38:44having a broader view of this and all the rest 38:46of it and not just makes this seem like, oh 38:48cool, this is by tech folks, for tech folks. 38:51written about, you know, by tech 38:53folks like, you're going to run 38:55a market as you said correctly. 38:57Yeah, and I think that will be a 38:58really long process of just like 39:00understanding how to, how to use it. 39:01Like I think about those, um, if you've 39:03ever seen those early films of people, just 39:05as film cameras were coming out and, um, 39:08you know, you basically people see people 39:09line up and they kind of like pose it 39:11completely still assuming that it was like 39:13a photograph versus like a film camera. 39:15And, and it's like, it took a while for us 39:17to be like, Oh, you can do movies with this. 39:18It's not just like a camera. 39:20Yeah. So, maybe we'll see. 39:21We're at that early stage with, with AI as well. 39:24well that's all the time we have for today. 39:26Chris, Marina, thanks for joining us again. 39:28Appreciate you doing double 39:29duty for two episodes in a row. 39:31Anastasia, it was great 39:32having you on the show. 39:33We'll have to have you back at some other time. 39:34And, thanks to all you 39:35listeners for joining us as always. 39:37If you enjoyed what you heard, you 39:38can get us on Apple podcasts, Spotify, 39:41and podcast platforms everywhere. 39:42And we will see you next 39:43week on Mixture of Experts.