# 182 - Alexa 2.0, MiniMax, Surskever raises $1B, SB 1047 approved

We try to record on Friday and release on Sunday, but, uh, you know, sometimes Things get in the way of that. Also, as you have been hearing, if you've been listening to the last couple of episodes, more likely than not, for the next several episodes, Jeremy will not be co hosting. He will be having, I suppose, a lot to do with a new baby underway. So, you can expect that. Hopefully, you will still like the episodes with the guest co host.

And just one last thing before we get going, we haven't had sponsors in a while, but we are going back to that and we have a couple of things to mention in this episode and possibly in future episodes. Don't worry, we won't be promoting any vitamins, we're gonna do our best to promote only things that are relevant to those interested in AI and tech. And that is definitely the case this week. The first thing I will mention is agent. ai, which is a service where you can actually hire AI agents.

And here's the text they would like me to read. Agent. ai is where the world discovers, builds, and hires AI agents to help them do more. Some of the tasks AI agents are doing for users include supercharging workflows, automating routine tasks, debugging complex issues in your codebase, Creating memes for your tech blog or social media and conducting in depth research on companies or technologies. Agent. ai is built for builders.

In fact, it was built by Dharma Shah, co founder and CTO of HubSpot. Agent. ai is your best global marketplace for agents and the humans who build them. It's a platform where you can find and connect with AI agents that will help you get more stuff done. With Agent. ai, you're not just using AI, You're at the forefront of its growth. Join us. We are building the future of AI together. Sign up for agent. ai at agent. ai today and get 100 free credits so you can start building right away.

So there you go. That's an interesting new effort that I suppose probably some of our listeners would find interesting. And just one more thing to mention before we get going. Another thing that we're actually having a cross promotion with another podcast that you may be interested in is Pioneers of AI, and here's the quick description of that. What will the AI revolution mean for jobs, for our kids, for our relationships, and daily lives?

A new podcast, Pioneers of AI, is your trusted guide to this emerging technology. Host Rana L. Kalyubi is an AI scientist, entrepreneur, author, and investor exploring all the opportunities and questions AI brings into our lives. Listen to Pioneers of AI with new episodes every Wednesday, wherever you tune in. So, once again, always fun to do a cross promotion with another AI. Focused podcast and if that does sound intriguing to you, then go check it out. And that is it for this prologue.

Please enjoy the actual episode. Hello and welcome to the latest episode of Last Week in AI, where you will hear us chat about what's going on with AI and of course specifically what's been AI this past week. And the most interesting and impactful news that happened. As always, I will mention, you can also go to lastweekin. ai for the text newsletter, where we'll have all the other stories that happened this week that we are not covering.

And also, by the way, if you don't know, if you subscribe to a newsletter, you also will get an email for each podcast episode and that email will have all the stories, all the links, so that's another reason to check it out. Alrighty, and that being said, I am one of your hosts, Andrey Kurenkov. I finished my PhD at Stanford last year and I now work at a generative AI startup. And I'm

I guess people did listen to me begging for a few more for a 200, but, uh, yeah, one of them, uh, said there was this review said great show every episode devout fail, Andre announces onto a lightning round only to dive into a segment that's just as detailed as the rest and that's been true and that's why the episodes have been two hours long, but we will see maybe this one that will be a bit shorter, uh, we shall see.

A couple more shout outs on the reviews front, a few more, uh, comments saying appreciation for the geopolitics and for your perspective on hardware, Jeremy. So glad to hear people are into it. I think our policy is we will keep covering it, but we will try not to like inject too much China coverage. In research segments and whatnot, where we can focus more on technical details. Uh, and you did get a congratulations on, uh, becoming a dad soon, Jeremy. So there you go. Wow. That's super sweet.

Oh my God. Uh, you,

So, uh, you know, uh, tracking and, uh, I will, I will be sure not to be single handedly preoccupied with stopping an AI apocalypse. You got to get some

This is set to be released in October and will actually Have a monthly subscription price between five to 10 per month. And interestingly here, Amazon did say that they are using Claude to power this new version of Alexa, which is interesting because they could have tried and created their own in house technology, presumably they will have to pay. And tropic for the use of cloud in this case.

Uh, but Amazon also has invested in partner drone tropic to have cloud on their platform to let people use the cloud through AWS. So in a sense, this is a continuation of our partnership. And it does, I think, show also that it really is very hard to even try to compete at the frontier model level, unless you're already a company like Google meta. Um, OpenAI on Froplic that has a GPT 4 ish level model.

Uh, it's probably easier to just partner with one of the existing ones rather than build your own version of it. Yeah, we've even

In kind of model architecture, design, um, you know, optimization schemes and all that. And also just finding out ways to make the hardware kind of sing together really effectively. That's still a challenge. Like you can see, it's not trivial. Even a company like Amazon that has tons. And I mean, Amazon at one time at least was like the, you know, cluster company. They, they were the ones who did cloud best. And, um, you know, that's arguably shifted actually.

Clearly shifted for AI at least, but, but still, you know, they, they had that, uh, that mentality and now no longer quite, uh, quite being able to stand on their own two feet. The partnership with Anthropic makes all the more sense through that lens. And it is something they're using as well to like kind of bolster their own, um, hardware fine tuning efforts so they can partner with Anthropic on that. Um, it is also partly a consequence of Anthropic's, uh, ROI oriented culture.

Um, so if you know anybody from, from, um, I'm sorry, I meant Amazon. I don't know if I said Anthropic too many A's. Um, Amazon has a very ROI oriented culture. You know, they always ask themselves if you know any kind of Amazon alums or people who work there, they're always kind of trying to figure out what is the payoff in the near term. What do customers actually want? So not building towards hypothetical things like AGI so much that they do not have an AGI team.

Um, but, but more like what are the tangible kind of value ads we can create today. And that's been part of what's led them to this, uh, position in the AI race where they're sort of seen as falling behind. And I think quite accurately. So, um, so anyway, yeah, Alexa has been a sort of struggling product line for some time and having a little bit of anthropic injected in there. You know, might, might actually help to liven it up a bit.

So, uh, monthly fee of apparently five to 10, by the way, for this new version of Alexa, that'll be powered. It seems by, uh, by enthral. Right. I

You could talk to it as you could to JGPT or Claude, you can ask it shopping advice, uh, and apparently can also fulfill complex queries like ordering food or drafting emails. So it will be interesting to see how many people do opt in. There's some analysis in this article saying that there are approximately a hundred million active Alexa users. And so if even 10 percent of those opt in, that could be hundreds of millions of dollars for Amazon.

So, yeah, I think this is, you know, this and Siri are both going to be interesting, uh, demonstrations of whether everyone wants to have, you know, the smart chat bot functionality in their assistant, or if people don't need it that much. And onto our only Lighting Round story, because we really are keeping this section short. It's one way to do quick Lighting Rounds, I guess. Yeah, yeah. This one is about Minimax, a new realistic AI video generator. So this is another one coming from China.

Then this was backed by Alibaba and Tencent. So those are huge, uh, You know, players in the space of AI over at China. And so we got a demo and a trailer showing various AI generation tools. Not clear it's quite up to par with the latest generation of Runway Gen 3 Dream Machine Kling. This is, after all, coming from a startup. Not, you know, hard to get to that level, but still another new model that is pretty impressive when you do look at the videos.

And as with the last video posted to YouTube, I will try to actually do a bit more editing and include Video clips on top of our talking about this.

It does seem like it's a, maybe a little lackluster, um, up to six seconds in video generation length. So, you know, compares sort of favor, uh, unfavorably rather to a bunch of other current industry leaders that could do 10%. The resolution is pretty good, 1280 by 720. Um, so, you know, reasonably good stuff, but it's, these are short videos. Um, so yeah, I mean, it's got, uh, one of the differentiators apparently that it has is it does seem to have captured generating human movement.

Well, so it does well with hands. It does well with walking and things like that. So that's, you know, kind of interesting, but even within China, right. You've got so many different competitors who are putting out text to video that it's like, I don't know, it's, it's already looking like a crowded market. It seems like all of a sudden this just kind of came out of nowhere, but. Minimax is interesting as a company itself. It's, it is one of, um, China's AI tigers.

Uh, so these sort of famous companies that are unicorns that are building generative AI products. Jerpoo, by the way, which we talk about a lot on podcast is, uh, is among those as well. So there you have it. Um, we'll, we'll see it. I think it's touted as, as the first foray that minimax is first foray into this space. So, you know, presumably we'll see the quality go up over time as well.

And so ever since the big drama over at OpenAI last year, when Semaatman was briefly kicked out, SysCover, let's say, took the wrong side on that and, uh, wound up leaving OpenAI after a little while. It was announced that he will be funding this startup, creating the startup safe superintelligence with some co founders, including Daniel Gross.

And I think it was maybe a few weeks or it's been a little bit of time, but now we have this upcoming news of the only additional bit of news about SSI since the announcement, which is that they have raised 1 billion. 1 billion from a bunch of big name tech investors, A16Z, Sequoia, and the firm that Daniel Gross, one of the, the CEO of SSI, Also a big investor and part of the, uh, group that he is funding this company.

Uh, there's some inside baseball, uh, regarding like why combinator, even this is a competitor to open AI, right? Very clearly, this company is trying to achieve the same thing. Open AI is trying to achieve maybe in a different way, maybe with more emphasis on safety, but certainly that's the play here, Daniel gross. Is a former partner at Y Combinator. Actually, he was one of the people who interviewed me when I went in, uh, to apply to YC back in the day.

And so he worked closely with Sam, Sam Altman back when Sam Altman was president of Y Combinator. Now, SA Sam is off, uh, obviously the CEO of OpenAI. So there's a lot of like YC kind of in the background relationships here. And, uh, and Daniel Gross famously headed up, uh, AI at Apple

A billion dollars sounds like a lot of money, it is. Um, but when you're talking about competing with, Companies like OpenAI, companies like Microsoft, the next beat, the next generation of clusters is going to cost 100, 125 billion dollars to throw down, right? These are the 2027, 2028 era clusters. Raising a billion dollars is not enough. If you are going to follow the same path that OpenAI is going to follow, it's simply not enough.

We talked about this before, you know, when safe superintelligence was first announced, one of the questions that we raised here, right, was, Uh, what's the differentiator, like how are these guys going to compete with the giant ball of capital that is Microsoft OpenAI or Google in a context where capital is what you need to buy the outrageous levels of scale that you need to pull off these training runs.

Some way, somehow, Ilya has got to come up with a strategy that allows him to do scaling better than OpenAI. That allows him to maybe bypass the need for scaling or more realistically, as he puts it, kind of scale something that's different because there's this question of what are you scaling right? As you train, for example, language model, you're training them at ridiculous scale to be incredibly good at doing.

Text autocomplete really autoregressive language model right and so he's dropping some hints that we don't really know but it seems like what he's suggesting is, you know, we're we're we're rethinking what we're optimizing for rethinking how we're training the model there's something more than just kind of vanilla scale it to blazes strategy going on here beyond that. You know, very, very kind of thin on the details. It's clear they want to emphasize recruitment very heavily.

Like there's a, you know, really a lot of emphasis on culture and all that, but we don't know what the nuts and bolts are, the meat and potatoes. There is one last detail here, which is that, uh, safe super intelligence is going to be split between, uh, Palo Alto in obviously California and Tel Aviv in Israel. And that kind of. Okay. Thank you. Is it interesting in its own right? Maybe it's a recruitment play. There is a lot of great talent in Israel.

Um, there's also questions, you know, maybe it's a bit of a hedge against potential AI regulation in any single jurisdiction, really hard to know. Um, but, uh, that just is an unusual and I think fairly noteworthy aspect of this. So we may learn more about the deal. We may not. Uh, but a billion dollars right now does not seem like enough. You know, if, again, if you're just going to go the traditional scaling route, you're competing with.

Real monsters who are going to try to, you know, build these hundred billion, 200 billion, 500 billion clusters. You gotta be able to compete somehow with that. So presumably the competition is going to come from what's between Ilya's ears. And, uh, he, you know, he needs to have some good ideas.

So it would seem to me that they would have to a build upon something like LamaFree because you have to have a model trained. On all this data, even if you try to do something clever and smart from a research perspective, there's no getting around scaling and the need for scale. So it'll be very intriguing to see for sure if they do find a way to utilize this amount of money to somehow make a lot of progress without.

You know, using the recipe that has been the key to progress, basically for the history of open AI, right?

And that's about the details we have here, uh, the 16 process is still in early stages of the mass production is begin expected to begin later in the year or two to me, Jeremy, you presumably have more background on this, like my reaction to this was, wait, is OpenAI planning to like manufacture hardware of their own? Is that what this is saying? Cause that is kind of a big deal.

And the fabrication process is arguably the single most technologically involved thing that human beings on planet Earth have figured out how to do, right? It's this whole thing about being able to, to, to write chips with like seven, five and three nanometer resolution. It's a ridiculously hard thing to do. Um, These fabs are fucking hard to build. All right, for context, if you wanna build a new semiconductor fab, you need to put down an aircraft carrier worth of risk capital.

We're talking like $50 billion. That's risk capital. What does that mean? It means you're not guaranteed to even get worthwhile outputs. At the end of the day, you could spend $50 billion and, and basically lose it all. So incredibly, incredibly hard to do. Um, TSMC has had trouble kind of setting up fabs. Uh, in North America, you better believe I mean, they're based in Taiwan. They want to get fabrication capacity off Taiwan in the event that China invades, but it's just really hard to do.

Like it is that difficult. Um, there's a, an old famous, um, sort of story that I think this is about Intel back in the day when they were kind of bleeding on, on a lot of fab stuff. When they would build a new fab, they would try to replicate down to the last detail. Every detail of the construction of the old fabs that were actually working because no one knew why they worked. It's like a fab is like a machine with like 500 dials on it.

Nobody knows exactly like you through trial and error kind of get a million PhDs were incredibly well paid to try to tune those dials. Right? So what they would do is they would do literally, they would paint the bathrooms the same color because nobody had a clue. Like maybe this is the detail that makes the difference. It's so finicky. So the idea that open AI was even considering, you know, Building a dedicated fab is insane. That is like, that should leap out to you.

It's like, wow, holy shit. You know, this is a really like that, that's an ambitious company. They now seem to be stepping back from that and they're saying, okay, okay, we're not going to try to like build a freaking fab ourselves instead of being the next TSMC. We're going to be the next NVIDIA. Maybe what we're going to do. As we're going to partner with companies like Broadcom and Marvel. So these are companies, you can think of them kind of like NVIDIA.

So NVIDIA designs GPUs, but Broadcom and Marvel, they also design chips. They design though custom ASICs or like, um, basically these, these more kind of application specific, uh, application specific, uh, like kind of integrated circuits. So these are more use case tailored chips. You can think for example, as an example of an ASIC, you can think of Google's TPU. So that is a custom chip specifically for training, you know, large deep learning models, right? That is what it's about.

So opening eye essentially is trying to make their own TPU. That's one way you can think about the story. That's a big, big deal. Like this is a massive, massive investment of time and energy. They got to do this. Um, they're going to be doing this. It looks like on this a 16 process. So this is where you're, you're, it's a 16 angstrom process. So, so right now with the. 16 angstroms, like 1. 6 nanometers.

So right now we're at three nanometers and, and so the next couple of beats, we're going to get two nanometers and then we'll eventually get down to, uh, to the, um, uh, yeah, 1. 6 nanometers and that'll be that. So it's a couple of beats away. Um, but essentially, yeah, they're, they're partnering with Broadcom Marvel to set this up, uh, it's. An incredibly capital intensive project. And really what this is, it's a really interesting strategic partnership too.

Right now, Broadcom and Marvel, because they design ASICs, right? They don't design GPUs. They're not directly competing with NVIDIA. But through a partnership like this, OpenAI is basically asking them to design chips that would compete with NVIDIA GPUs, and that is pushing Broadcom and Marvel into closer competition with NVIDIA. So it's a really, really interesting play, and it has all kinds of implications for the, for the industry of chip design itself.

So I think this is a story we're definitely going to be hearing more about as it unfolds. Um, but, uh, but yeah, this, this node, the A16 node is going to be a very, very Loaded node, uh, from the standpoint of, um, of competition on chip design.

So one of the big advantages that Google has over competitors is that they have been working on TPUs since, I think the first one was 2015. Now they're in their fifth or sixth generation. So designing such a chip is not easy. Broadcom has shown that they can do it. And they are. You know, raking in billions of dollars because of our partnership with Google, Google actually wants to move away and be able to do it themselves.

But, um, yeah, we've seen that precedent and it would be interesting to see if OpenAI is able to essentially replicate what, uh, Google has done with our own custom hardware, which would then kind of shift to the Playing space in several ways, for instance, reliance on external compute on Azure from Microsoft. Openly, I would imagine still can't do their own training runs in house, so to speak. So having. They're your own custom chips.

It's all very intriguing in a very nerdy kind of way on to the lightning round. And the first story is again on open AI. It is that open AI is weighing changes to its corporate structure amid the latest So as we've noted, I believe last week or two weeks ago, OpenAI is looking to raise more money. They're looking to be valued at over a hundred and billion dollars. Uh, they have some big names like Apple, NVIDIA, talking to participate in this round alongside Microsoft and existing investors.

If this would be successful, That would make OpenAI one of the most valuable technology startups in Silicon Valley history. That would be like above Stripe. Uh, which raised a 95 billion, uh, valuation while doing private fundraising. So. Yeah. Add to that kind of price, a hundred billion valuation wouldn't be surprised if investors are hoping for a slightly less weird corporate structure.

The board structure that like, it's weird for sure, but they created it for a very explicit reason back in the day they were concerned about a single company monopolizing the revenues that would come from the success of AGI, right? And so the way it was set up initially was you had this like non profit entity, this non profit board that, um, kind of runs, uh, it owns essentially a for profit entity, like OpenAI itself. For the for profit entity.

Um, and essentially if you invest in the for profit entity, your returns are capped. And this is the capped for profit structure. You could only make, I think it was a hundred fold, a hundred X return. So in Microsoft, for example, put their big investment in back in the day, I think it was 10 billion. The most recent one, um, their, their returns were capped to a hundred X, the original investment.

And the argument here was if we, if we ended up making more than that, Then, um, essentially with the, any surplus does not go to Microsoft so that you don't have a situation where the rich get outrageously richer. Obviously they make a lot of money, but they're, you know, they're not going to get infinitely more wealthy. Instead, that those revenues are redirected.

I believe it was to the OpenAI non profit parent entity for redistribution in some broader sense, uh, sort of across the population. And there are all kinds of, uh, thinking, and there's all kinds of thinking that OpenAI. people had done, including somebody called Colin O'Keefe, who put together this proposal for the windfall clause.

This idea that if OpenAI at some point, or if, if a lab that signed onto this clause at some point generates a gigantic amount of, of, um, windfall, Revenues then they commit in advance to sharing that to redistribute it according to a certain scheme. And so all of this was factored in was baked into that board structure. And now the whole idea was for precisely this eventuality. When you see the valuation of the company soaring like crazy.

Now, all of a sudden, whoops, we're, we're just saying, okay, you know, forget about that. We're going to change the corporate structure so you can read into that what you will. It certainly is consistent with the strong incentives. As you said, Andre that Must exist. You have all these companies that want to see returns, right? Like it's no good if there's if there's a cap on these returns.

Apparently, there hasn't been a final decision on whether or not to change the structure in that direction. But it is under consideration to remove this cap on profits right that that cap for profit structure that's apparently what's potentially on the chopping block that nothing's finalized and um, yeah, I mean, this is, this is a space where you just have a lot of these weird structures. Obviously, Anthropic has.

A more traditional benefit corporation structure, but it's, it's sort of to make sure that their fiduciary obligation is not to the shareholders, but rather to the, the wider, um, uh, population. And so anyway, the public benefit that is, uh, so yeah, we'll, we'll see, uh, if this actually ends up changing, if it does, uh, don't be surprised to see a lot of people complaining that this might kind of go against the grain of opening eyes, initial commitments.

All of those co founders are now going to Amazon along with a quarter of the startups employees. So very much similar to what you've seen with character. ai and Google, what you saw with Amazon and Adept. back in June, this is like the fifth or sixth example of this weird sort of actually hire, but not actually acquire deal. So, uh, yeah, it's certainly an interesting thing to see.

And I think is an educator also that in this space of advanced robotics for manufacturing fulfillment, uh, that's been around for, you know, quite a while now, covariance has been around since 2017. What I've kind of heard and then seen a little bit is that the companies are struggling to make it work and to really scale.

Uh, Covariant has said that they have like a robotics foundation model, they have advanced AI, but somehow This move does hint to me that maybe it's challenging to really kind of push in and make it a successful business, uh, at least so far.

You are, you're basically having these folks abandon ship implicitly. What the C suite is saying when they leave for Amazon is, I think I have more potential over there. And now they've got the CEO role. And that comes with a whole bunch of challenges for a company, especially one that's been around for a while. And, Anyway, so this is kind of, um, uh, it's dicey. I like, I don't know how long covariant is going to continue to be a going proposition after something like this.

It sort of makes you think of other acquisitions like this, that, uh, that you referred to earlier. Um, and it is being done as well, uh, to avoid regulatory scrutiny, right? Antitrust scrutiny, or this idea of like consolidating all the, the companies in the space. Um, so, so acquiring companies like Amazon in this instance, presumably are trying to avoid a, an all out acquisition because of the optics of it. But, uh, yeah, we'll, we'll see it. We'll

So I don't know if morale is, maybe this is good for covariant because now it's hard to say, but certainly, you know, it's, it's not been the case that these kinds of deals were common in tech until like this year.

And their founder, Tang Jimin, is now being sued for being, for failing to raise over 70 billion dollars in CSP financing.

They, they just like light billions and billions of dollars on fire on a regular basis, which can sound stupid unless you factor in the fact that it is a national security priority for China to kind of get ahead in the, in the AI race. So they're, they're really willing to just light that money on fire. And, and it's, it's led to some wins, but anyway, so this is one of those cases where, yep, that's, you know, it does cost you money every once in a while.

Um, the corruption in this take seems to have taken the form of the founder basically taking the money and, or so sorry, some senior executives, as you say, just taking the money, like pocketing proceeds from the fundraise, not actually spending it on R and D, but just like. You know, kind of taking it for themselves, which is by the way, not, not legal. Uh, so anyway, that's, uh, that's it. One of the big challenges that they've faced, um, is poor Europe, poor yields.

Uh, and this has been an issue consistently for, for Chinese companies. Anytime you see a big headline where it's like, Oh, uh, you know, some Chinese company just came out with a new GPU and it looks really, really good. And there's a tear down and, you know, it does look solid. Always ask yourself in the background. What is the actual yield, though, of the process that is being used to fabricate this device? Very often in China, what you find is the yields are just garbage.

And so it's just not economically viable without state support. And they do have state support that's subsidizing these industries. So just kind of a background thing here. And I think the corruption note, you know, we haven't talked a lot about these high profile Kind of corruption crashes if you, if you can call them that in the Chinese semiconductor space, but they, they do, uh, they do happen and they happen quite a bit. So, uh, another interesting instance of that.

This, uh, obviously I don't know as much about this as you, but the gist from the article is that these are still not being developed at scale, these kinds of silicon photonic technologies. There's a lot of push behind it since it will help with these AI clusters. So TSMC has stated that they are ramping up on their ability to do this, these kinds of chips over the next few years.

So there'll be much less energy efficient than getting just optical interconnects, which is what this is all about. Um, that's, it's widely viewed that optical interconnects are going to be kind of the, the wave of the future over the next five years or so. And specifically this, this area of co packaged optics, where you're bringing the optical interconnects as close to the chip as you can, by which I mean, You want to, um, essentially have all your GPUs connected by optical interconnects.

And then when you start to get on the chip, as much as possible, you even get those wires replaced with optical interconnects. Um, so it's, uh, essentially just a way to, yeah, make everything work more efficiently and faster. You benefit from the fact that light just moves crazy fast and it doesn't tend to Burn energy, create heat as it moves around in the same way that electrons do as they kind of bump into each other in, in a circuit.

Uh, right now the issue is there are already some of these deployments, some of these, um, coag co packaged optics deployments out there, but there are really low volumes, like you said, and the big challenge is how do you get the industry to invest enough in this technology that it will come online in five years. There's just so much r and d required to do this, and so you're seeing the emergence of a kind of industry alliance.

That's effectively subsidizing this, uh, this process of, of bringing this new technology online may not be profitable, may not be beneficial in the short, short term, but unless they make this upfront investment, they're never going to get there. And so you see quite a bit of coordination. Uh, you've got meta and Microsoft, for example, backing this co packaged optics, uh, technology.

And, um, uh, and, and a bunch of others stepping in the space, not unlike what you see with Apple and TSMC there, there's a relationship there where Apple will usually order from TSMC is most advanced node, their highest resolution, tiniest node, um, for their iPhone and effectively they're subsidizing. TSMC's research and development on that latest node. And then, you know, we get to benefit other companies rather get to benefit from it later on as it becomes a more mature node.

Um, this is basically what's happening at the industry level with co packaged optics. So yeah, interesting space. And I think more and more, uh, going to be relevant at the chip design and, uh, and fabrication level in the next like five years or so.

Pretty much we only larger model, well, there's a couple, but Aside from meta 400 B 70 billion is sort of that, um, category of where most of a big open source, uh, language models are. There's also 7 billion and 2 billion, uh, parameter models. And those are being released under a fully permissive Apache 2. 0 license. The 72 billion model is not being released. You don't have access to weights. You will most likely have to use it through. The Alibaba cloud API.

They have released some benchmark numbers and they're quite impressive. The 7 billion parameter model does the beat GPT 40 mini on some, uh, reasoning and the vision task in particular on video, it performs quite well and better than any, uh, or almost any other open source. Uh, model of a similar size. So yeah, it, uh, you know, he sometimes I think can get tired of like, Oh, another model that's like, uh, how many of these do we have?

But as a vision language model, I think in that category, we don't have quite as many things. Llama of course is primarily just, uh, language. So this could be a big deal.

It kind of binds the end user to Chinese law, which is an interesting way of, of waging a sort of open source lawfare or not lawfare, but open source warfare in a certain, um, kind of informal sense and not, not doing that here. We're getting just the Apache, the vanilla Apache 2. 0 license. So this is a true open source play. Um, there's a, yeah, a bunch of interesting kind of innovations going on here. They have this thing they call naive dynamic resolution support.

Basically, this allows the model to take in images at all resolutions, um, and, uh, and it dynamically maps them to a number of, of tokens. So anyway, it, it's sort of like. Uh, more, more, uh, robust in that sense to different image shapes and sizes. And then another one, which I thought was really interesting, this idea of they call it multimodal rotary position embedding MROPE.

Um, which is, so, so there is this notion of, of rope, like rotary position embedding that's been done for a long time. Um, it's a little complicated to explain, but basically, like you, You add up a signal that you overlay a signal on top of your input tokens that, um, that the, the model can learn to, uh, interpret as giving away the position of that token in the, uh, in the input.

So, you know, transformers natively don't actually can actually tell what order tokens are appearing in and they don't. They don't care about it. What you need to do is give the transformer a hint by overlaying some kind of signal onto the embeddings that you feed it. Well, what they're doing here is the same thing, except they're overlaying three different of these signals at the same time.

One for the sort of x axis, x dimension of these images, one for the y dimension, and one for the time dimension. And combining those together, that kind of allows the model to tell, okay, you know, which of these tokens correspond to which patch of the image and what time slice of this video. So it allows you to do, to kind of generalize the idea of the more standard rotary position embedding. So I thought that was kind of cool.

Um, and, uh, and it does seem as you said, to lead to really impressive results. So, uh, there you go.

And so this paper, it kind of describes in some detail of their architecture of their system that they developed for training. And so they go into how they deployed this Fireflyer 2, uh, system with a hundred, sorry, 10, 000, uh, PCIe A100 GPUs. They say that they are able to achieve, uh, or approximate a performance of, uh, let's say more powerful product from NVIDIA while reducing cost and entry consumption. And, uh, after that, there's a bunch of nerdy details about how they did that.

So it sounds like they worked on some communication technology. Honestly, not my expertise. So I'll just let Jeremy take over and give you a thought.

Um, what they do is they basically take this cluster of A100 GPUs that they have, like you said, these 10, 000 A100s, um, PCIe GPUs. So you can think of these like individual GPUs. And then if you've got these individual GPUs, you have to figure out how to like wire them together in pods and then wire the pods together and create like a functioning cluster that unifies all their work and gets them to dance together in a coherent way.

And what they do is they, they essentially set that up, they walk you through the process of how they achieve that, and they reduce. Um, costs by half and energy consumption by 40 percent using their fancy techniques. What I want to emphasize though, is just a, a plot that's like that's in the papers could figure two here. And it's so interesting. It shows you how over time.

Different components of, say, of GPUs, different, um, different computing components, different, uh, elements of the GPU, different, different features of the GPU have been getting better over time.

When you think about computing and creating these clusters for very scaled compute runs, uh, you know, people often think about the what's known as the logic, so the number of operations that a GPU can perform, the number of flops, floating point operations that a GPU can perform, uh, that matters a lot.

But it turns out that that the number of flops that GPUs can, um, can perform their efficiency, uh, is, has been increasing about, uh, roughly speaking, uh, it's like The three, three X a year over a year. So every year you get three times better flops, but that's not the only number that matters. So you can number crunch like crazy. What happens in practice when you're training a large model is you have to like send, like split up your model, send bits of it to different GPUs.

They have to compute. The, the gradients, basically the, the model weight updates that they're going to, uh, sort of recommend if you will. And then you have to pull all those weight updates together. And then you have to redistribute the new model weights to all those GPUs so that all the GPUs are kind of updating the same, uh, same consistent model. So you can get it to converge. And that requires a ridiculous amount of bandwidth.

You got to be sending model and the, you know, uh, parameters and data left, right, and center to all these GPUs. So, so interconnect bandwidth between. Um, different, uh, kind of GPUs becomes really important and so does the on GPU high bandwidth memory. And so these two things, the interconnect bandwidth and the on GPU high bandwidth memory have been scaling much, much more slowly than the flops than the logic component on these GPUs.

And the consequence of that is that we are very quickly going to find, in fact, we are already getting there, We're not flops bound, bound, uh, sorry, flops bounded. We are bandwidth bounded. We we're essentially like high B bandwidth, memory and interconnect bandwidth.

So in other words, within and, and among, uh, GPUs, uh, the, the memory bandwidth required to move these weights around, to move activations around gradient, updates around, uh, is, is our ability to do that fast is not growing fast enough to keep up with the other. Um, with the logic, say, and the wider demand for, for computing. And so we're hitting that wall right now, the sort of memory wall. And this is just something to look out for.

So just trying to kind of increase the level of resolution at which we're thinking about GPUs. It's not just about, you know, oh, can we get more logic on this chip? Can we get more, more flops? You can't use those flops if you can't transmit data between your GPUs fast enough to get, you know, the information you need to run those computations in the first place to that GPU.

So anyway, um, I thought it was a really interesting, uh, paper for, for, for the reason of that figure alone, if nothing else. Um, for context, 3x year over year increase in logic and flops, 1. 6x increase year over year, 1. 4x increase year over year only in, uh, in the sort of like, um, high bandwidth memory and interconnect bandwidth. So much, much slower rate of progress. on that exponential curve.

So I thought it was kind of interesting, a little bit more resolution on Moore's law, if you will.

And how you set up a training run. But, You know, I, I think a lot of it is hardware and how to make your hardware work and a hundred percent. Yeah.

To kind of squeeze more juice out of the lemon, but you can also just make a bigger lemon, have more computing power, or just optimize that computing power. So you're right. It's like, it's, it is all kind of one problem. And that is a reason why structurally companies like Google are massively advantaged because they own the whole stack and they can optimize the whole stack together. That's also why right earlier today, we talked about opening. I tried to design their own chips.

That's why they're trying to do it. They want to own more of the stack so they can They're thinking of it as a holistic problem, right? Just How do I, you know, get more compute efficiency out of my system?

Pretty much goes into how We've covered on and off this topic of long context reasoning, which is to say you have a big chunk of input, let's say 10 books, and you want your model to be able to take in this large amount of text and actually reason about it. So this has been one of the challenges.

In AI and in in scaling language models is if you give it, you know, 100, 000 words can actually use the content of 100, 000 words effectively, even if technically you can and the claim being made in a blog post is one. But the current benchmarks are flawed, which I think is generally accepted.

This needle in a haystack test that is often used is pretty much saying, can you pick out this one tiny bit of text that has been inserted into this longer document, and the inserted bit does stand out as something that doesn't really belong there. So that has been an issue with a benchmark. We've seen some efforts on that front and here they introduced, uh, kind of another way to evaluate it with, uh, essentially kind of randomized data where there's no needle that stands out.

It's just randomized pairs of things where you really do need to sort of have a very effective memory to be able to complete the problem. They then say that they have a new technique that works really, really, really well on it. And that's it. Like, I didn't see anything about how it actually works, which was very disappointing, but that's their claim.

But. Beyond that, like you, you know, you're not going to get a lot of detail. So that is very much this paper. I did like this whole discussion that you were, you were pointing at with the needle in a haystack stuff, right? So they created this new, um, essentially, uh, eval called hash hop. And so there's this idea of, of hashes, right? And these are kind of like random, random strings of characters and numbers and letters. And. The point of them is that they are random.

In other words, like you, so imagine a document where you just have a bunch of hashes, you know, a bunch of random digits and letters, separate space between them, and then another chunk and another chunk like that. And now in that context, right, you're asked to recall one of these random hashes. What this forces the model to do is to actually kind of Keep in mind, genuinely keep in mind the whole document that it's just read.

If it's going to correctly pick out a hash, a random hash from a document filled with other random hashes, it can't use cheat codes. It can't just, for example. You know, famously when Anthropic, uh, ran their big needle in a haystack, uh, test. I don't know what they fed it, but it was something like the collective works of Shakespeare. And somewhere in between, they, they include like a pizza recipe, right? And that sticks out like a sore thumb.

And so the model quite quickly goes, Oh, wow, you know, that's obviously not supposed to be here. And it makes it really easy. It doesn't, if, if you were asked to like, You know, pick out the thing that doesn't belong in a document. Um, you wouldn't actually have to read the document super attentively, right? You could just real skim and wait until you see something. It's like, what the hell is a pizza recipe? Okay, clearly this is, this is the thing.

So that creates an artificial sense of an inflated sense of the model's actual capacity on these evals. Whereas if you have completely random, uh, strings, you know, a sort of hash in this case, then maybe you're able to force the model to keep the whole thing in mind. And as it kind of looks for, for, for one, and it does this thing where they also force it to do. These sort of logic hops. This is why it's called hash hop. So they will map one hash to another hash.

They'll say like, you know, this hash, um, connects to this, this hash and that hash connects to this next hash and so on, like a string of connected hashes. And so they can go and basically say, okay, um, model, I want you to. tell me what the, um, the hash that is connected to this hash is based on the giant document you were given. And they can even ask it for, okay, what is the fifth hash that is connected to this one? So force the model to go through like many steps of recall.

And that starts to look more like the kind of reasoning that that humans do when they read long documents, right? You're kind of integrating information from across a document and your reasoning on top of it. And so I thought this was really interesting. I wouldn't be surprised if we start to see this happen a lot more.

Um, what they, what they're doing is functionally, they're measuring the emergence of What are known as induction heads, which are, um, actually, uh, in, uh, mechanistic interpretability. These structures that, that have been fairly recently discovered in, in language models, um, that essentially, uh, will, uh, look back over the sequence that's been fed to them for previous instances of the current token.

Basically see like, if the current token is pizza, you look back over the previous sequence and go, Hmm, every time I've seen pizza previously, what was the word that followed that? And you grab that word and then you predict that that's the next word. These are known as induction heads. They're actually hypothesized to be the main driver of the incredible meta learning ability of language models.

And there's a fascinating, uh, bit of literature on how they emerge and every training curve actually that you'll see will feature a little induction head bump, as it's called a little bump where the model suddenly learns about the induction head and you'll see the loss drop suddenly. So, uh, anyway. It's a really, really interesting way to probe at this very fundamental and important contributor to, uh, to metal learning.

And, uh, so I thought great paper, but again, doesn't, doesn't really tell us much about what the hell is going on under the hood.

That's a hundred able to do a hundred million. Uh, and that's different. That's different. So anyway, exciting progress on long context reasoning, maybe. And congrats to magic for raising a bunch of money and trading viral models. Alrighty, on to lighting round. The first story is smaller, weaker, yet better training LLM reasoners via compute optimal sampling.

So there is a trade off between using a stronger but expensive model or a weaker but cheaper model when you generate synthetic data for training. Right? You can generate more data with a weaker model, or you can use a stronger model to generate better data potentially, but not as much data.

And so that's the topic of the paper, this kind of problem of compute optimal sampling, and they show that, uh, Depending on the different settings, uh, knowledge distillation, self improvement, you can find the right balance, so to speak, and actually get a lot more out of weaker models and not just the strong ones.

I'm going to use the best model I can find like, you know, GPT four turbo or whatever it is, or, or, you know, call it 3. 5 sonnet to generate text that my, my model will then learn from.

And. Uh, yeah, I mean, it's it's an interesting paper of the results vary, but they do suggest that, um, you often benefit from saying, Hey, actually, instead of having really, really high quality text, but a small amount of it, uh, you're better off having kind of more medium quality text, but a large amount of it, I wouldn't be surprised if the capacity of the student model Is actually a big factor here.

In other words, the capacity of the model that you're training using data that you generated from sort of GPT 4 model or whatever you choose, because if you have a really small student model, It might not even be able to learn all of the subtleties that make a really exquisite model that much better. So it's kind of a waste of, of, uh, time out of the gate. It's sort of like trying to teach math to a five year old, right?

You're better off with more examples of simple math problems than a small number of like rigorously correct and complex examples. Um, so yeah, I wonder if that's kind of part of, Part of this. And that's something that I didn't see addressed directly in the paper, like the capacity of the student model as a, as a function of all this. But, um, be curious to see that looked at too.

And because of this sort of non uniform structure, typically you need special types of architectures to deal with graphs. And they are generally speaking harder to learn with. Uh, and as a result, you don't have as easy a time creating what they call here a foundation model where a foundation model is basically like, okay, this model can take any text and do something about text. This model can take any image and do something with the image.

That's pretty hard to do with graphs because graphs have different structure. They have different, um, you know, contents per node. The edges may or may not be directed. They may or may not, uh, have, uh, features on them. There's a lot of variation that's kind of not exactly there with texts or images. And the gist of it is here. The researchers from the university of Hong Kong have introduced.

What they say is graph foundation model called Energraph building, um, graph mixtures of experts, uh, architecture that can then deal with a big variety of graphs because we shouldn't get into the technical details. There's a lot that went into it, but they did. Uh, evaluate on experiments on 38 diverse graph datasets and showed that without training on any one dataset, you are able to generalize and do well, which is what you want for that foundation model.

Okay, and moving on to policy and safety, our first story is once again about SB 1047, as it has been for the last couple of weeks. And the news this time is that it has been approved by the California legislator. So we've kind of known that this was going to happen. This was leading to this, but now that it has been approved, uh, it will be soon moving on to, uh, governor Gavin Newsom to be able to veto it or not. And he has not given a stance on the legislation yet.

So there's still a potential for a veto here and for it not being, uh, passed. The chances are probably that it is going to pass given that this is kind of a democratic effort and Gavin Newsom is a Democrat, but, uh, yeah, it seems more likely than not that this is going to be becoming a law, but still up in the air.

Uh, there has been a lot of active lobbying from all kinds of, of, uh, big tech companies, um, and, uh, Y Combinator as well as we've talked about, uh, trying to push back on this bill. And, um, and so, you know, that, that might be part of this as well. So it's very hard to tell, uh, there's, there's been a lot of inflated rhetoric in, in all directions, but, uh, but, but including in the direction of sort of, let's say claiming that the bill does things that it doesn't do.

Um, and that's been a recurring issue. So, you know, definitely. All this might be part of the, the landscape here. And I guess we'll just find out when he actually, when he actually decides whether or not to sign it. But, uh, there, there does seem to be a surprising amount of uncertainty given the political lines that you've indicated. Like, yeah, you know, you would think a democratic governor, democratic, um, state legislature.

It probably cited to law, it seems a little bit uncertain at this point.

And, uh, we'll just have to see and no doubt we'll talk about whatever happens soon. Next story is now a research paper related to safety and it is called Pemper resistant safeguards for open weight LLM. So we've mentioned this, uh, on and off with open late models that, uh, you know, in general, you want to make your models, uh, hard to jailbreak, you want to make it so they can be misused for various reasons, but.

If you release your model and its weights out into the open, as for instance, Meta has done, there's really nothing you can do, uh, more or less. Like once you have access to the weights, you can undo any sort of training for safety. Uh, you can, you don't even need to jailbreak. You can just retrain it more or less to do what you want.

And so this paper is introducing this, uh, tar method for trying to build some tamper resistant safeguards so that adversaries cannot remove the safeguards even after training for a fairly large amount of time. And they say that in evaluation and red teaming analysis, they found that, uh, the method is effective and also preserves capabilities, which is another more important dimension of this. So that's what just, and now that Jeremy. Dive into the details, as I'm sure you love to do. Yeah,

So, uh, we can think back to the good old days of generative adversarial networks, GANs, which used to be, uh, the, you know, the way you would train generative image models back in the day and sort of have, have turned into diffusion models nowadays. But so this is kind of following a similar philosophy. They have two stages of training. So they have their language model. And they'll start by baking in some kind of safety measure.

So, you know, reinforcement learning from human feedback or, uh, you know, gradient to center or strategies that are meant to make the model forget about dangerous knowledge. And so they got that safety, um, safety measure baked in. And what they'll start to do is this iterative training procedure where in the first iteration, they will.

have an automated jailbreaking technique, some kind of tampering attack that involves fine tuning that model to try to get it to, uh, unlearn the safeguarded behavior or relearn the dangerous capability. And, um, so, you know, a common way to do this, if you have, for example, an instruction fine tune model, that's trained to refuse to, to execute on instructions for like helping you to build a bomb or something would be to give the model some extra training.

With, um, instruction data where the instructions are dangerous and it includes the execution of those instructions. You're kind of training the model to like to go. Okay, never mind. I actually will answer these dangerous queries. So you start with that. You have your base model with your safeguards. You have your phase where you're trying to get it to unlearn those safeguards and then you're going to try to bake into it a counter essentially to that.

Um, you're, you're going to try to kind of rebake in the, uh, the safeguard, but as you do that, there's a risk that the model will forget the actual capabilities that you wanted it to have. And so, there's this kind of tension between getting the model to learn the safeguards, And, and unlearn dangerous knowledge, but not unlearn too much of the good capabilities you want.

And so there's this outer loop where you actually want to get the model to remember the capabilities, the sort of capability retention. And anyway, the way they execute this is quite interesting, but it's, Too much detail to go into here. Um, they find a significant level of success when they compare their method, uh, to others, both on the capability of retention side and on the kind of safe behavior side.

They're able to retain the safeguards while, um, uh, while preventing capabilities from being lost. Even if you allow an adversary to try to fine tune out the safeguards. Now, one important caveat is that this. does not work for, um, let's say attacks that you didn't factor into your training process. There are a whole bunch of attacks that we haven't yet discovered that we'll discover later. And those attacks will probably, you know, work much better on this kind of model.

And the, the issue there is with open source models, you just kind of put them out there and you can start your clock. And eventually somebody's going to come up with an attack, but the people who developed the model never thought of. And so this isn't necessarily a good sort of robust forward looking technique, but it certainly does help on. You know, in a sense, it helps make it a little bit harder for people to jailbreak these models.

Um, it isn't terribly effective at a certain category of attack called parameter efficient fine tuning attacks. This is kind of like, like, um, Laura. So if you, you know, if you, if you design attacks that only retrain some of the model parameters, but not all in strategically chosen ways. This technique doesn't tend to work really well. So there are holes, but this is certainly a better technique than we've seen before.

It's something that I think, you know, some people might've thought might not have been possible. And so here we have a pretty good, uh, update on our ability to secure these open source models. I don't think it lives up quite to the hype that, uh, that the paper is sort of, uh, tees up at the beginning where they kind of imply that this will herald the new, new era for open source. Um, safeguards, I think it's, it's, it's between incremental and not incremental. It's certainly interesting.

And, and it'll need a lot of, um, a lot more robustness in order to be sort of future proof in the way that I think it will need to be.

So this is a teardown from a Tokyo based te, which is a semi semiconductor research company, and that's the gist of it, is that it does seem like China's ship manufacturing capabilities are catching up somewhat to TSMC. And three years may not sound like that much, but you know, given the cycles in the uh, space, you know, few years behind being that in a couple years China will be able to make. cutting edge, uh, AI computing infrastructure of the sort that is being made now.

So I would not, uh, I guess I don't have a background. I do wonder what the estimate or consensus was. And this, if this is different, uh, as far as you know, Jerry,

One of the things that they don't quite get into here that again, is every time you hear about not just chips from China, but chips from anywhere, always ask yourself the question yields, right? You can have a really impressive chip that looks really solid. And on the basis of that chip alone, you might be telling yourself, okay, well, you know, China, Oh, wow. They must be really close.

But if that chip can't be produced at good yields, if like, you know, 50 percent of the, of the chips that come off the production line don't work, uh, then effectively the economic yield of that chip is not high or may not be high enough to be sustainable. And that's where you get into things where, you know, like we talked about Chinese government has to subsidize these things. In this instance, I think that's part of the subtext here.

Um, essentially what we have is TSMC's five nanometer process, uh, is being competed with, it seems, quite effectively, By high silicons, uh, design capabilities at seven nanometers. Um, so what they're probably doing there is multi patterning. This is this idea where yes, you have a crappier, lower resolution process, but if you use it to go over your chip multiple times, you can achieve the same outcome as if you had used that higher resolution five nanometer process.

But going over the chip multiple times means it's more expensive. Your yields go down and so on and so forth. So a lot of uncertainty still left here. Um, it's just another indication that yes, in fact, you know, Huawei and their subsidiary, high Silicon that actually does a lot of their chip design stuff, uh, is yeah, they're, they're a live player. Um, you should expect the U S export control stuff to, to start to kick in more and more over time as a layered process.

And they're likely to fall behind more and more, but for now, um, it's, it's further along that I think a lot of people would have expected. The operating question is always yields. And the thing with the PRC, uh, with China, that is, is that they're willing to subsidize this stuff. So economic yields. Don't necessarily matter in the same way, let's say.

Part of the thing with ASML, the maker of these incredibly, incredibly complex machines you need for fabrication of chips. They sell you these lithography machines. They also then have a contract to essentially continually maintain it. And, you know, this is so such advanced technology. It's very important to have that. So, uh, unsurprising in a sense that China would want to prevent that. And if they were to cut off, uh, the, uh, chip maker, ASML.

Would be very significant since ASML is, as far as I know, kind of the only company producing, yeah, the cutting edge UV

You know, if you see an op ed go out in the Global Times, you, you know, implicitly that this is something that is to some degree endorsed by the regime. Um, so, so it's a way of saying something without saying something with some weird level of plausible liability or some, some level of flexibility. But yeah, they're, they're saying, you know, if, if this happens, then ASML risks losing access to the Chinese market quotes permanently.

Um, and, uh, obviously whether or not ASML actually implements the latest, uh, us export curbs is not entirely. ASML's option, right? Like ASML is going to be forced by the Dutch government to do whatever, and the Dutch government may or may not choose to enforce the U. S. policy that the U. S. is asking for. Um, so, so it's sort of, you know, treating ASML a little bit as a pawn here, presumably trying to create some pressure on the Dutch government via ASML through these threats.

So, you know, it's, uh, it's, it's just another day in the world of, hey, chips and geopolitics. I'm not supposed to be talking about that.

build out infrastructure for creating AI hardware and doing advanced AI that's slowly been coming into focus and kind of taking shape as something substantial. And so this article details, kind of the emerging details Of what the first steps to that seven trillion number dollar, uh, looks like. And that is working to get, uh, kind of permission from the US government and getting various investors on board to start doing build out of machines and systems needed to train.

So that would be tens of billions of dollars, presumably for fabrication and things like that. This is all coming out. From people familiar with the matter. So it's reporting on these kind of deals that are being talked about. Nothing is signed or kind of agreed upon yet. It seems, but when we do hear more concrete news, probably it'll be along these lines.

with, with, you know, the UAE and other things like that, um, you know, that can raise alarms. So they're, they're trying to cover their bases and get ahead of the story. Talk to the U S government first, get their blessing and then do it. Um, and, and part of the pitch here as well, as they say, is that other companies besides opening eye would also benefit from these infrastructure projects.

So. Yeah, I mean, I see it's a sort of compliment to the Chips and Science Act, the sort of like, uh, I guess it was 40 or 50 billion package that Congress authorized, uh, for kind of domestic, uh, AI chip fabrication. Um, this is sort of the more of the, um, kind of GPU end of it. Um, with, with, uh, sort of more of an AI bent to it. So, yeah, uh, this is, it's quite interesting. Opening eyes had meetings with the U S national security council specifically.

So this is the, you can think of this as, um, these are the, the council of people who directly advise the president on national security affairs. So, you know, they're, they're going, uh, straight to the, the highest levels of the executive. And, um, Yeah, uh, it's, it's Sam Altman doing what he does best. It seems these days schmoozing with the U S government. And, um, and he's, he's very good at it.

And now there's an organization behind it that has stated that opposing using AI writing tools is quote classist and ableist. So they made the point that they don't want to restrict anyone from using Novels written through the help of AI apps like chat GPT. Now, no rhyme or by the way, is happening in November and challenges participants to write a 50, 000 word manuscript. The organizers did clarify that using AI for the entire submission would, uh, defeat the purpose of a challenge.

But, you know, as you might expect, some members of the writing community have criticized the stance, have argued that. You know, these kinds of tools generally benefit the industry more than people actually using it. NaNoWriMo did face some backlash and has updated its post to acknowledge about The concerns people have with regards to generative AI tools in the writing industry.

That like makes it appealing. I don't know what the right, uh, the right math is there. Yeah. Maybe, maybe divide it. Maybe have two separate competitions. I don't know what I'm talking about. I'm just pitching things, you

Who in case you don't know, somehow is a very, very famous, uh, American actor, uh, has, uh, said on that post that there are numerous ads online that promote miracle cures and wonder drugs and basically warned anyone who followed them, uh, that he had no association with those posts or products and that they should be careful. We've seen this happen previously. I think we've even covered another news story about Tom Hanks being used to promote some dentistry product or something.

But I think, um, yeah, another reminder of this being very much an ongoing trend that, um, I don't know, luckily I have somehow managed to avoid seeing any deep fake ads to celebrities. Oh, it's coming, but they, they exist. They are out there and I'm sure people like Tom Hanks are not very happy about it.

Uh, but, uh, yeah, poor Tom Hanks, man, that's going to be a real, uh, legal headache for him to be going around suing all the, all the companies or people are

And it boils down to, as far as I took away this notion that to create art, part of the process or part of what makes it art is the process of decision making of actually sort of having to make thousands of little choices in the creation of writing or novel, for instance. And so.

When it comes to something like AI generated images, it seems that if your only choice is effectively the wording of a prompt, even if the output looks like something that you could say is art, under this sort of, I guess, philosophy or framework, um, you would still discount it. So it's, it's, I think, In this kind of field, in this discussion of like, is something that AI makes art or not art? This is a fairly nuanced. And well thought out tape that I found to be an interesting read.

And if this is interesting to you, I would just recommend you read it as opposed to just get our summary.

Yes, you as the painter, as the artist are making a million different decisions, but mechanistic interpretability might allow us to identify effectively the same number, or even a greater number of decision points that are implicitly being made in circuits, in neural networks as they generate these things. And then there's just like the separate question of, sure, like you can make this argument for This way of thinking about art, but everybody has a different way of thinking about art.

And it's just not super clear what this implies in practice about how we ought to enjoy something or not enjoy it. Um, you know, maybe it's just one of these things where we're learning that the word needs to be broken apart into more nuance and we need to have.

I generated art and human generated art and things in between, um, to kind of get better resolution because some people do care about the meaning behind it and the fact that a human made it and we're thinking things and you know, that these decisions were made in some kind of organic substrate and that that somehow is, is important to them. And I mean, I, I certainly. Kind of empathize with that. But yeah, I just, I don't know. I, um, I don't, I don't super get the debate.

I I'm, I'm, I'm sure I'm, I'm missing something. Um, but it just, it's always struck me as being like, yep, you know, there's. You just need to know what you are after with, with art, you know, if you just want to be stunned, you know, if you want to have a beautiful thing in your house, like sure, generate it with, with AI, if it suits you, if not, if the backstory behind it matters to you, if the generation process matters to you, then yeah, you know, go with a different thing.

You might hear this one just like a couple of days after, uh, the previous release. And can I just

What the hell is it? Anyway, thanks.

Everything's out of chance. Cause I'm taking a shot tonight. I said it's funny, but I wanted to. It just blows me away. And it's a real I believe. Sorry I'm not as sure. It's a prickly I do trust you. I'm a liar, I need a general identity. I'm a liar, I need a general identity. Say, hey, there was a story close, but there was new harmony. The thing that's for you. You still lie, you're not the truth. Curse, grace, distress, spell, lie. We're with our lives in a trance, but it's not a chill.

I can't, can't ever be sick of you. So I'm torn, Hey, hey, so the 1 H 2, Gosh, we wet, so we all lose. The dog's mad, she's fixed, I've done space.