#147 - Google’s Gemini, EU AI Act Deal, Mixtral of Experts, Meta’s ‘Seamless’ translator

[email protected] is an easy way to do it, so I just figured I'd float that there. This week is I think a really. So first of all, we were just talking about this earlier. This is a bit of a remedial week for us because last week while we were on air, Google had the nerve, the audacity to release Gemini and a bunch of stuff around that. And so we got to do some catch up on that. And they're just a small number, really big stories, right?

The Super Data Science Podcast is one of the most listened to data science podcast out there. It is ranked 12 as far as technology podcasts globally, maybe it's even more now. I don't know. That's from my notes. And they cover everything. They cover machine learning A.I., But also unlike us, they cover careers and not a science. What the professionals do. They have two episodes out a week and they interview a huge variety of people. They already have 700 episodes out.

The show is hosted by John Crone, which chief data scientist and co-founder of a machine learning company Nebula, and the offer of a best selling book, Deep Learning Illustrated. So he's a super informed interviewer. We've had him on the show, and if you've listened, you've heard what he knows at least as much, if not more, about AI when we do. So, if you want to listen to some people out there doing stuff in the world, we do recommend the Super Data Science podcast.

And so Gemini was meant to amend that and basically put them back in the lead. And so they last week had a sort of kick off event, you could say a little virtual conference with a whole bunch of stuff coming out of Gemini. They released a whole, you know, slick video introducing it. They introduced a separate video that was just a demo of its capabilities. And then they updated Bard with their mid-tier version of Gemini.

Gemini comes in three variants, ultra nano and pro nano is on device, so it'll be coming to pixel phones. Pro is very kind of GBP 3.5, maybe ish. The variant that is now out and Bard and then Ultra is a big boy of that is you know seemingly maybe better than Jubilee four based on a whole bunch of numbers and results that came out with a technical report they also put out, along with all the summer stuff, and that is not out yet. They are holding it back until next year.

So, yeah, last week we briefly mentioned this, but now we're going to go into more details and yeah, the rollout is underway. They are going to give access to API developer. Is today as of our recording on December 13. So people are going to be able to start using Gemini. Yeah. And so far, I would say my impression is people are pretty impressed. It seems very cool.

The big step forward here compared to Bard and many other chat bots is Gemini is natively multi-modal, so it is built to support inputs of text, images, video and is able to play around with all those. And in the tech demo we showcased how you could show pictures, talk about pictures, have hand gestures, all this sort of stuff. So yeah, it's still coming out and people are starting to play around

with it. And there have been some reports, so I didn't do so good with Bard, But if you read the paper, they do claim really, really impressive results, especially with Gemini Ultra. So overall, it seems like a pretty good rollout so far.

But I really do encourage, you know, if you're interested in Gemini in particular, and like the deep dives of the sort that we do here actually on some models, then then check that out, do kind of more deeper dives into a smaller number of things. But yeah, so a couple of things that jump to mind. So first of all, Gemini Ultra, which by the way, is is not going to be released until the new year. So we're not seeing that out quite yet. We're going to see early releases for Pro and Nano.

So one of the key things to recognize about this is they claim that it smashes their achieve state of the art on 30 out of 32 benchmarks for text data. Okay. So not all benchmarks on text data but when it comes to benchmarks on image understanding, on video and speech recognition, it knocks everything out of the park, setting new sodas on a new state of the art on on all three across

the board. And so I think one of the key things this reminds us of is that Google, DeepMind and Google more generally really, really, really are good at the multi-modal side of things in a way that perhaps openai is less so open in AI, specialization is more on instruction, fine tuning, reinforcement, learning from human feedback. It really pioneered a lot of the early applications in that domain, and so you're starting to see a little bit of a forking out.

This is the area of specialization and this is a play that Google DeepMind is going to make much more in line with their previous models. You can take your perceiver IO which came out in I think 2021 multimodal gotto big multimodal model. It could do 450 or 600 flamingo.

One interesting note, by the way, we covered, I think in the just the last episode, this new benchmark, the IMU benchmark, which was explicitly designed to test multi modality and sort of more deep reasoning than the traditional sort of like reasoning evals. And so this is it's already found its way into the Gemini paper. They achieve a state of the art 62.4%, which is five percentage points ahead of the kind of closest competitor there. So we're already seeing progress being made on that

benchmark. I think one of the key things to factor into is and this may this may be part of why DeepMind and Google are strategically going for this multi-modal strategy. They have a big hardware advantage relative to Openai. They just have a gigantic in-house supply of to use their kind of homemade tensor processing units. We'll talk a little bit

about those today. But anyway, so so compute is a big thing and that I think, makes it it lends itself more to this sort of like high throughput, high volume approach, whereas opening I need to be a little bit more strategic in their leverage use of compute. Couple of final notes here. So first of all, 32,000 token context window for Gemini. I thought that was interesting. We're seeing Anthropic right with Cloud 2.1 with like a 200,000 token context window.

So certainly Google starting to pitch its tent on a more narrow, a smaller context window aiming for higher performance but with smaller chunks of text. That's kind of an interesting note. And it was trained on their previous generations of to use the TPU V4 and the TPU the five E there's now the tpu v p which wasn't referenced in the. But it's coming out. So just so you're aware, like basically we're already one generation behind relative to what hardware can actually accommodate.

So Google can already probably do better once those things roll out at scale. Just a sign of how fast things are moving. The training process. This was so interesting. Hardware, hardware, hardware. I keep beating this drum. We do so much stuff on hardware on this show. This is the reason, guys. The hardware is the reason for Gemini. So what did they do? They because of the insane scale of the system, they had to train it not just across clusters of GPUs, but across different data centers.

They had to hook up different data centers to paralyze the computations across entire data like a separate data centers. And so essentially, this creates this new bottleneck because they it decreases the time between failures in in hardware and this overall system. You have so many GPUs here. And so they have to come up with new strategies to make the training process work. One of the things they did was they deliberately set up a deterministic, a fully deterministic train

process. So when a bug happens, you can always revert back to the previous training step and just replay forward and and hopefully it won't it won't bug out again. So this took a whole bunch of work just anyway, a whole bunch of really cool stuff in the paper. I thought about rapidly detecting and getting rid of faulty hardware and there were a whole bunch of really interesting kind of deterministic replay techniques they call them. So kind of replay the

computation. And because your whole thing is deterministic, in other words, you can always predict the state of the system in the future from its previous state. Assuming no bugs happened, this gave them the ability to kind of white out errors and then replay over them. So, so much so much interesting stuff there. Last thing I'll just mention for now for we

chat more about it. But the there was a bit of controversy over the announcement here where DeepMind and Google were saying kind of like, oh well look at the performance of our system relative to good for on this benchmark the MLA benchmarks or the general purpose language understanding benchmark. And they played this game where it wasn't really apples to apples, so they tested it on the MLB benchmark, giving it train of thought, prompting.

So train of thought prompting is this technique we've talked about on the show before where you had the system walk itself through in a step by step manner, its reasoning process and then helps to kind of prime it Terezin more effectively. So chain of thought, a chain of thought prompting with with 32 steps in the context window it performed so Gemini performed at 90% on the anomaly relative to GPP four, which just hit 87.3%. That's a solid win. And if I was a marketing guy, I'd say that's

enough. Let's call that a win. Let's announce that. But they didn't do that when they reported on, at least on their public facing blog. They correct this in the paper, but on their public facing blog they actually report the GP for performance on five shot, not 32. And so and for that it's like, what is it? It's like a percentage point lower, it's 86.4%. So you don't gain that much and it creates this opening for controversy. I feel like that was just a really big marketing

misstep. This is clearly a big kind of jump ahead. But I will say, if you do look at five shot, weirdly, Gemini underperformance relative to GP four.

So like, I don't know why you would do this because now it invites the question, okay, how did Gemini perform on five shot And then your attention is extra drawn to the fact that it actually underperformed GP for their what accounts for the relative like what is it good for on chain of thought prompting at 32 whereas it gets beat by GP for on on five shot. I don't know. That's a really interesting research question, but again, kind of feels like a bit of a marketing own goal there.

Or there is this voice talking to Gemini and presenting hand gestures and photos and things like that. And it. Seems like presumably if you were to try to use it that way, it won't be as smooth or as fast or as seamless. Right. So it's definitely edited together to give the best impression, which I'm

not surprised by. I think pretty much probably most people in a I who would see them, I would think this is probably cherry picked, cherry picked as it is for anyone who doesn't know a common phrase where you just pick the best results to present. And it's not always a bad thing. It's just, you know, you have to be aware of it. That's what's happening. But there are probably worse results that are being shown. And this is often done to have like image generation, for instance.

So that was another source of controversy. And as you said, there is a bit of a marketing spin here in the technical report in the PFA released it, Unfortunately for the academic community is another one of these corporate kind of quasi research things that basically tells us nothing about Gemini and how it's built and any technical details regarding it. And for evaluation.

It is a little bit gamed in the sense that it does appear to be pretty clear and and done properly, but they do present results of varying kind of conditions where, as you said, like to compare to GPU for their main metric is this chain of thought at 42 metric, which is this fancy pragmatic technique versus other one five shot where basically give it five chances to get it right. And there does worse.

They have some other metrics in the stable table two like majority one at 32 there for shot every shot zero shots. So it's kind of a strange mix of evaluation strategies partially based on what's been done in prior work, but also probably they just tried a bunch of ways to get good results and reported the stuff that worked best. Although to our credit in the table two, they do present a five shot result with 83.7 performance. Also warfarin.

Just one more thing is that on that multi-model benchmark GP for V, the multimodal version which Q4 does pretty well, does just a few percentage points worse. And yeah, if you go online, someone on YouTube actually recreated the Gemini demo video, whereas you could do for V. So it doesn't appear to be like, you know, blowing gibberish for an opening out of a water. It's probably around as good from everything we can tell.

Although there are some examples online of people playing the Bard and seeing some kind of silly outputs and things.

about apples to apples. And kudos to Google for being clear about what specific prompting strategies they were using. But when you start to see that sort of variation across these things, like, you know, I don't know, you kind of either want thoroughness or you want consistency. And this sort of gave us an interesting little salad of those things, aren't

you? On transparency? There was one thing I was I was interested to note when they were describing the training process and they mentioned, you know, for for specific detected risk areas, the sort of risk classes that they're concerned with on evaluations, which are things like, you know, testing for dangerous capabilities, biohazards, persuasion and cybersecurity and so on.

They they actually ended up generating responses at scale from their systems that relied on, as they put it, a custom data generation pipeline loosely inspired from constitutional A.I.. So that was sort of interesting. You know, that's usually associated with Entropic, who pioneered this. And that's part of what makes Claude sort of characteristically more careful in its responses. And so here they are using this anthropic developed technique. There's, of course, a deep partnership now between

Google Anthropic. So presumably a lot of exchange of talent. I know on the safety side, their safety teams do a lot of really good collaborations. So this may well be a reflection of that, but I thought that was an interesting note between those two.

So yeah, it's a line to be careful, it seems, which probably makes sense. And arguably that is limited capabilities. In the same article they do compare creativity and championship agility enabled. The Bard was able to do that, you know online search and to provide a pretty good summary of what's going on. So it's really kind of optimized to be a little safe seemingly, but still very exciting to see Gemini

being launched. And I guess we are definitely looking forward to seeing what happens with Ultra next year.

The jump between pro and ultra is significant and in on many metrics, including things like math and reasoning seems verging on flirting with exponential even So really, I think quite significant for what it means about scaling and crucially positive transfer. So positive transfer between tasks. They have now observed that at this scale, which means that the model is learning stuff from text that makes it better at doing images, makes it better, doing audio makes it better, doing video.

So all these things, kind of indications of the formation of a a fairly robust world model within Gemini itself, which again is all part of what certain people at least have been predicting will naturally emerge with scale as we approach AGI.

When it first came out, it was built on one of the smaller versions and that's because the fine tuning process is quite extensive for code generation will be extra expensive if you use the full scale model and there are all these additional complexities. Okay, so Alpha code to can outperform 85% of human programmers on essentially a code force. This code force competitive programing exercise, which like its previous predecessor, Alphacode won, only hit the top 50%.

So here we're going from top 50% to top 15%. This is really getting to the point where these systems are automating, able to automate like, I mean a lot of high quality code it so essentially they they just fine tune this model for code generation. I think this is a really remarkable leap. It's 43% of competition problems which was an almost two x improvement over what Alphacode could do.

So yeah, there's like I think the coding piece is really important just because anything that gets the language models to touch core logic makes it easier to train beyond human level capabilities because then they generate outputs that can be verified. Are they true or are they false? Are they right or are they wrong? Because code will

actually do that. It'll give you a a rigorous answer to your query that in turn can give you a reward signal that you can use as a scaffold to exceed human level capability at things like coding, because you're no longer just doing autocomplete on human generated code, you're adding a reward signal that's based on ground truth. Did the code work or did it not? So I think this is a really significant

breakthrough. We're going to talk more about this idea of scaffolding and superhuman capability day and in future episodes. But just one of the part that they're alphacode to. I think it's it's a it's the real deal.

So it seems like the intent is basically to have an on device AI that is going to enable a lot of features and they have some examples. It's not super. You know why did we have an example of summarizing for recordings and the some smarter auto reply features? But that's about it. But I think it's pretty interesting to think because they are investing endless, you know, custom, you know, home built on device A.I., which could be a pretty big edge in terms of phone hardware that they are selling.

And moving on to some non-German news because that also exists. And next we have a meta unveils audio box and AI that clones voices and generates ambient sounds. So Metro Facebook has released this project of audio box. This is more of a research project to me.

We have a demo out and yes, you can use it to clone your voice and generate, you know, natural sounding audio of you saying stuff you can type in a sentence for that voice to say or describe a sound for it to generate an audio box will create it. This is meant to be a family of models, so they do have this different models for speech, mimicry and another one for ambient sounds and sound effects.

And yeah, it's a pretty ambitious big research project that is, and you can play around with it in the demo, but it is a research demo and they do say you're not allowed to use it for commercial purposes.

There is a a text style prompt that you can offer. And so you can write like in their paper, they give this example. You know, a young woman speaks with a high pitch and fast pace, right? So you're kind of providing that context that you might include in, you know, if you're used to good for the system, prompt is probably the closest thing to this. And then you provide separately the text transcription that you want the character or the voice to say.

And then in addition to that, you can provide audio, context prompts and and a voice prompt as well to kind of lead into this this thing in much the same way that you might with like a standard auto progressive or like text autocomplete like model like 354. So kind of interesting again, that multi-modal thing that meta seems to do. It seems like even audio can't just be audio, it's got to be audio and text and very consistent with that approach that you know, where they think

multimodal is. The path to AGI, I guess now somewhat similar or somewhat aligned with some of the work that Google is doing as well. So her distinct approach from mobile in that sense.

But it seems in this case they did go the more safety oriented route of selective release to researchers and not just open source doing it, which, you know, given this is literally voice cloning, you can easily see the argument for that here. So interesting to see that they're not kind of fully on one end, do they do have some exceptions to the rule that they have been following in general?

Clearly implying that Grok was potentially trained on data generated from Chargeability or was built off of an open source model that was trained on synthetic data generated by. But that's the most likely kind of source of this I think, where basically nowadays. For training purposes. You might want to use a planning GP for to generate extra data for training. If you cannot scrape the entire internet and that can result in these kinds of outputs contaminating your training set.

So that's probably what happened here.

But people just found that it kept coming up so often in response to so many different prompts that the only explanation that people kind of could see was like, Yeah, you probably use GPT four, you know, sent it up like a large number of prompts and literally used GPT for outputs to fine tune or to train the model. So obviously a lot of questions about this. I think there's a copyright question on this that's super interesting. Is it copyright infringement?

If I train my model on all the text on the internet and a bunch of jeopardy for generated text just happens to be among it? Am I infringing on Opening Eye's copyright? If not, then is it copyright? If I directly ping the model to get outputs that I use to train my model on? Like, I don't know. But this is a whole other dimension separate from the whole. Like is GPD for trained on, you know, Hollywood scripts and real authors work and stuff like that. So anyway, this, this rabbit hole

just keeps getting deeper. But really interesting story there.

And there are some other examples of it saying things like Diversity inclusion are essential for creating a fair and equitable society, etc., etc., etc.. So, you know, similarly it's quite aligned to be progressive and values which yeah, as your article says, some some fans of your musk are not super into. And that also does imply that probably it was strained on data that is that they may be open to AI and maybe our chat about it. I mean to.

Next story Gmail's AI powered spam detection is its biggest security upgrade in the years, so Google has upgraded vs spam filter as we have a new text classification system called Bret VAC Resilient and efficient text recognizer, which is meant to understand adversarial text manipulations. So emails full of special characters, emojis, typos, etc. etc. And it's meant to be resilient against that and many other things.

So, you know, clearly a pretty A.I. neural net, etc. like robust model that can now detect and prevent attacks from these kind of more sophisticated attempts at spam.

So you can have your own low copy of it, but you can now give access to data from Firefox and then it can, you know, interact with, you know, oh, what's this browser tab? But I was just looking at give me some answers etc. So it's just pretty experimental right now, but you could imagine it being a pretty cool development if we are able to have our own local private instances of, Hey, I've had do have access to the data we are interacting with.

business section. And, you know, we just talked about hardware, hardware, hardware, right. Like that's what that's what drives capabilities in AI. Well, Midha and Microsoft say they will buy AMD's new chip as alternatives to in videos. So for context, NVIDIA has a massive, massive market share in the GPU marketplace, of course, being the the AI optimized hardware that's used to train these really large models in video. So that stranglehold is really, really bad.

Mirror and Microsoft and other companies do not like that stranglehold because it essentially gives in video pricing power. You know they can they have insane margins because they have such market dominance. And so Microsoft and Meta and other companies have a strong incentive to nurture competitors to and video at the same time as they partner with NVIDIA. And so that's what's really going on here. AMD is really one of the dominant kind of leading competitors to NVIDIA.

They're very small market share but really, really powerful. And you Chip, they've just announced the instinct Me 300 X, This is a sequel to a previous chip that they had that was also really, really good. It seems to be favored. It seems to compare favorably to the Nvidia H 100, which is like the top of the line GPU at least. Currently it's the GPU that is being used to train or has been used to train PBT

five, for example. So really the next generation of the AI hardware, well AMD has its own answer to that. It's, it's got really impressive impressive characteristics including one key thing is high performance memory that allows it to hold way more parameters on chip. And that's really important for a lot of the optimizations you want to do at scale. So one of the key questions is now going to be about supply.

You know, like how can amd so sorry when I say supply, what I mean is AMD needs to convince TSMC, the one company in the world that can actually build the GPUs that they design so and designs GPUs, then they send them over to TSMC to build just like in video does. Well, now they have to find a way to convince TSMC to build their chips in large enough volumes. That's always a really big challenge. Nvidia is really good at buying up capacity from TSMC and crowding out competitors.

So that's one of the uphill climb that AMD needs to face. And that's why it's possible in this space to design the world's best GPUs, but still. Not win market share. If you just get out competed at the level with allocation from the actual fabs, from TSMC, for example. So that's really important. AMD has also come out and said that they have their software suite called Rock Em that they're going to use to compete with Nvidia's.

Of course, industry standard CUDA software, that's one of the big reasons NVIDIA so dominant. Well, now AMD is coming out swinging with their own software suite. A big part of their effort here to kind of reach parity. And so anyway, really interesting, their expectation is they'll see sales of more than $2 billion with their MI 300 series. So that's that's a pretty big deal.

And one last thing is that their CEO indicated to reporters that, you know, AMD doesn't need to be in a video to do well in the market. When you have a competitor that has like 95% market share. You don't need to win like 50% market share to be a really to in relative terms do really well in in the stock market. And so, you know, if they can chip away at 10%, 15%, 20%, that's a win already.

So kind of interesting to see what a runner up thinks and just how well Andy is doing on, you know, by the numbers here.

And video level player and vs statements of Metta Open the AI and Microsoft using this new chip are coming out a from an investor event last week. So yeah it's will be again we have been covering AMD and it's kind of slow rise for a little while now and it seems to be the trajectory is is continuing so it'll be interesting to see how how big this new chip will be.

It makes the world's kind of highest resolution, semiconductors with the smallest feature sizes. So they have their there seven is right there. There are three different feature sizes that are going to be relevant here. There's seven nanometer, which, by the way, is the feature size that you need to make the in video a 100 GPU, which in turn was used to train GPT four.

So you can draw a straight line from the the nanometer resolution, if you will, that sited at TSMC to the model families that we're talking about here. So seven nanometer means NVIDIA, A100 means GPU four OC below that is five nanometers even higher resolution to to make five nanometer scale stuff. It was thought that you needed to use extreme ultraviolet lithography techniques. So so these basically were just only manufactured by some random Dutch company the machines that can do this.

So TSMC buys these EUV, these extreme ultraviolet lithography machines to do five nanometer and below. And it was thought that that that's the only way that you can you can pull that off now five nanometers that feature size gives you the Nvidia H 100 GPU which in turn is being used to train GPU five four may have already been used 25 Okay. Now below that is three managers and that's like NVIDIA has plans right now to roll out their first three nanometer chip use sometime in late 2024 or 2025.

So that's kind of like the next next generation. So it was thought for a while that China could not produce seven nanometer. So in other words, that they couldn't produce the A100 so that they couldn't make a homegrown GPU four equivalent and a bunch of U.S. export control protocols were put in place to prevent them from being able to do that. Well, guess what? China's homegrown semi C company, which is a competitor to TSMC, is the Chinese equivalent.

They've shocked the world and showed, hey, we can actually make seven nanometer we have a77 nanometer process technology and it works at scale, which is another thing that you look for like, okay, you can make small batches, but does it really scale? So they prove that with their Kirin 9000 s processor. We covered that story a couple of months ago when it happened. Now, shockingly so, it turns out apparently they've managed. Perhaps to make a five nanometer class

process note. So something even more advanced again that would put it on par with the H 100. If this is true, if it holds up, and that would just be a frankly astonishing, disturbing from the standpoint proliferation. If you think of this stuff as a strategic technology and there's now a bunch of detective work going on where people are trying to figure out is it that they've they've actually rolled out because Huawei came out and announced this new laptop that has five nanometer technology.

The question is, did they just buy that from TSMC back when the export controls hadn't hit yet? And did they just stockpile a bunch of these cores or are these actually in-house built things? And there's evidence for both. But right now, the evidence seems to be that actually they may have succeeded at building stuff in-house, partly because if you make a node this good, you don't tend to wasted on something like a

laptop. You know, you tend to use it for places where you really, really need high performance, you know, maybe GPUs, but certainly phones and mobile devices. So really interesting and particularly weird. They seem to have managed if they've done this in-house and they've managed a five nanometer process without using extreme UVA lithography, which again previously like no one thought that was possible. So many question marks here, huge geopolitical significance.

I think one of the most important stories of the year on on Chinese domestic semiconductor manufacturing.

And it appears that generally, again, we don't know all the details, etc., but we general consensus is that they were firing was the result of kind of a longer term breakdown in relationship between the board and Altman due to various interactions.

And ultimately, I think as we covered Altman try to get Helen TONER, one of the board members, removed after a critical bit in a research paper that criticized Open the Eye and that kind of precipitated and built on top of these tensions and led to the board firing back and firing. ALTMAN And there's some tidbits here about Altman perhaps being a little manipulative, a little bit dishonest, in some cases, nothing super spicy or novel.

But that seems to be the consensus is that this was not something about, oh, we need to slow down as commercialization at Openai. It wasn't something like Sam Altman stole money or there was some giant hack that was not disclosed. It was really a relationship and the ability of the board to trust Sam Altman, to steward open the eye and lead it in this in open admission, which is to develop AGI for the benefit of all of humanity.

saying here. The claim is that, you know, Sam tried to get her ousted because of something that she wrote a report that she published that was somewhat critical on depending on your reading of opening. I and the way he went about this was by circulating to different board members and trying to convince them that other board members had already expressed the same views that the team was expressing. So kind of misrepresenting other people's views to the board.

And when that all came out, when it was revealed that this was going on, that may have been part of what turned, Elias said. Keeper against them. A to basically flip the board overall against them to lead to the ouster. So kind of interesting and concerning, certainly particularly given that the situation now has reversed.

So, you know, if this now has has not actually been accounted for, I think, you know, we'll hopefully learn about more about what actually happened as as the review process that is now ongoing. That concludes and hopefully we we actually have reports that really the result of that investigation.

interview. Helen TONER really didn't say much, but it broadly, you know, is in line with this narrative and understanding of a situation where it was tensions and the board did what it did, you know, in accordance with its mission to strengthen openai and make it more able to achieve its mission. And that Yeah, so that's what we know right now.

open the eyes. Shareholder is that what she was doing was might put her in legal hot water when in reality I think her position was actually correct. She had, because of her position on the board of the nonprofit entity, she had a fiduciary obligation to the public not to open a I say. And so, you know, that was sort of a thing that she flagged as as yeah, pressure let's say that was was put on her some might argue

unfairly. And next up again you know we promised we talk about Google TPS and in light of the Gemini thing so this is Google announces the cloud TPU V5 its most powerful A.I. accelerator yet. So along with announcing Gemini, Google also announced this next generation of AI hardware. Gemini, I should emphasize, was not trained on this generation. The strain on the previous generation, which included the TPU V5 E, which launched I think earlier this year, became generally available.

So one of the really big features of this thing is extremely high interconnect bandwidth. So this is the bandwidth that connects the different chips, the different TPMs during training. So basically they can communicate with each other very, very efficiently. This is a big deal for the reasons we explored. We're talking about Gemini, right? You're training this massive system in some cases across different data centers.

Forget about across different clusters and and that means you need really good communication between these chips and Yeah, so they have like 4800 gigabits per second, which is just like, this is insane. It is just absolutely huge for for context, the 100 thing video 100 has like 900. Yeah. 900 gigabytes per second, gigabits per second I think around there anyway. So it's, it's considerably, considerably smaller. So yeah, really big leap.

They claim that this features among other things, a two fold improvement in flops. So essentially the number of calculations per second, the floating point operations per second and a three fold improvement in high bandwidth memory. So there's a lot of like a lot of multiples and the result is just a real powerhouse of a chip. And so I think one of the key takeaways from this is Google is a hardware powerhouse.

And when you think about the Google slash Google DeepMind versus opening AI rivalry, which is so important right now, one of the key axes that differentiates them is access to quality hardware at scale. Look, Microsoft, which is partnered with A.I., has a ton of that. But Google, I think, has overall a significant hardware edge for the moment. And that means that they're able to substitute cleverness for just brute force and scale these

models a lot more. So this is a, I think, a structural delta between these two organizations that I'm sure will gradually get mitigated over time as Microsoft invest more and more.

But there was also another team member who left in October. Could be a bit of a slowdown. For Tesla. And it also goes to underscore the importance of investing in hardware in this space.

So deep, deep expertise in chip design. The Dojo supercomputer is powered by a custom chip. A custom chip that was designed by the and along with the guy who's replacing him now. But that's that's how central he was. And he helped set up Apple's Tesla brother's A.I. Hardware team back in 2016. So long history. One other tidbit, and I don't know how much stock to put in this, but like, well, actually, sorry. First of all, they lost Andre Carpathia also

earlier this year. So between Andre and Dinesh, these are two like, you know, if you believe that AGI is going to be built at Tesla motors, you don't leave. So this is an indication perhaps of a broader sense that folks are leaving that ship and we can draw whatever inferences we may. Last thing I'll note, Morgan Stanley apparently estimated that Dojo could add $500 Billion to Tesla's mark ups. That's like 50% of Tesla's market value, like

roughly magnitude. So if that's true, this is a is a big ding to Tesla's operations. But who knows? That's a speculative guess from some finance, you know, orbit.

So Assembly I've applied AI venture has raised the 50 million in funding in a funding round led by Accel with participants from the Salesforce co-CEO, GitHub, Acciona, Friedman and Daniel GROSS. So some pretty big names in the VA AI world, bringing the company's total capital raise to 115 million. They say that customer base has grown 200% from last year and its platform is handling around 25 million API calls per

day. And these models are designed to perform speech to text primarily and other speech related activities. So this seems like somebody's doing pretty good.

that sort of thing. But a lot of people who clearly don't think so and they've already picked up about 7 million users and they've generated over 700 million images. So, you know, 100 images per user is not bad. And they've launched an enterprise version recently as well. So that might be part of the kind of the big push here. So sort of interesting to see this this new entrant on the space and in the space, rather. And I hope that they I hope that they don't go the way of stability.

And yeah, as far as, you know, image generation after generation startups go Leonardo, that I might be one of the notable ones so far and now they're raising money. So it seems like probably doing all right and are two projects and open source, starting with a mixed trial of experts on their frustration.

open source. So this is kind of a big deal in the open source space because this outperforms Obama to 70 billion on most benchmarks and mashes or outperforms GPA 3.5 to get a little down to call this mixed struggle of experts. This mixture of experts model is one of these techniques to improve capabilities by essentially having some copies. And instead of training just one bigger model, you trained a few models and then they together can

perform better. There has been speculation about but you for is built with this kind of technology roughly and this is what they have released. And I believe this is one the first large mixture a model mixture of models, model, open source, and is seemingly have the best thing out there. Now, as far as open source things go.

right. This is so it's not that we're training a whole bunch of models separately. They are all trained together. What ends up happening is during training, there's a kind of a be a component of the model called a switch that actually decides which which kind of expert model, which sub model to root a given query to during training and during inference when the is actually kind of in performance mode. And so what happens is in this architecture, it's

kind of interesting. They've got so a switch like this and a bunch of different levels. So you have a switch, it routes a query to one of eight different groups of parameters and that's why the the model is called a mixed travel eight x 70. It's kind of like eight experts for a total of 7 billion parameters. So they're kind of telling you it's, you know, it's not a 7 billion parameter model. It's kind of divided into four.

Sorry, It's actually a like a in total like a 4647 billion parameter model. But the individual experts are 8 billion or 7 billion parameters. Sorry. So I have this router decides which of those those sub models to send the query to and then only those models actually chew on that input. And so in that way you're effectively using less of the model, fewer of the parameters in each forward pass to generate each next token. And that reduces your computing costs.

But it also means that your model during training, the whole model isn't exposed to every sample. So the model actually learns less per training pass. And so, you know, for a given model size, it's actually less impressive than it sounds as like a 47 billion parameter model because it only uses about 13 billion parameters with every training pass. So this is a very big model.

It's not the sort of thing that you train if you want to deploy it on like an edge device where you really need smaller models, it's something you're doing when you're optimizing just for output performance. The fact that it outperform Islamic 2 billion sorry, Lama to 70 B on most benchmarks and has way faster inference so it can generate its output six times faster is a really big deal. We are told it handles it doesn't handle a context of 32,000 tokens.

Oh now it gracefully handles context of 32,002 is. Yeah. No it's it's a very it's a it's a very kind of marketing sort of blog post handles English, French, Italian, German and Spanish. So. Oh, also great cogeneration. Last thing I got to say, sorry, I almost forgot about this. Mistral notorious for not training in safety measures in their models and just open sourcing them. And guess what? If you want to set up guardrails on your Mistral model, you can.

You just have to use a training flag to explicitly turn on the safe mode binary flag in their API calls. And if you don't do that by default, it'll just do whatever and that's it. So that's kind of Mr. Charles position on safety, I guess, is open source DVD 3.5 equivalent models and hope that they don't get misused. Fun.

And one thing to share is if you're not in the AI world, we have a fun tidbit that you're probably not aware of, which is Mr. is gaining a bit of a reputation in the AI space as being a cool company. I don't know. Vai inventor and Travis Davis was originally put out there was via a tweet with a link to a torrent to get to model. There was no announcement, but there's no flashy,

you know, whatever. If they just tweeted a link to download the weights of the model, which by the way, is licensed under Apache 2.0 so fully permissive in the sense that it can be used for commercial applications, no restrictions for like generating training data or anything like that, which Falcon and Arma two and others have. So yeah. Mr. Hall Kind of a cool kid on the block now with some of the best open source models and tweets to torrent links instead of product announcements.

Next, Mid-air unveils Seamless translator for real time communication across languages. So this is a suite of audio models called Seamless Communication that is aimed to enable seamless communication across languages with its flagship model, with its flagship model, combining capabilities of three other models seamless, expressive, seamless streaming and seamless am4 TV two into one system that is publicly available system for real time class lingual communication.

It preserves vocal style and emotional nuances of the speaker's voice during translation and is near real time. So yeah, very cool. And again, open sourced. It's on hogging face and GitHub.

By not using the transformer architecture, you can do something different from a transformer and outperform according to this blog post. So they claim that this is the first alternative model competitive with the best open source transformers in, you know, short and long context evaluations.

It's Outperforms Automata, Yi and Mistral seven be on the open L leaderboard as far as the 7 billion variant goes and yeah, it's actually very beneficial to be different from a transformer because you are able to be faster and more memory efficient by not requiring some of the technical details of customers.

But they kind of fallen out of favor for various reasons, as Transformers in particular have picked up. And so, yeah, kind of cool to see that in some form. I mean, they're gated convolutions, but still convolutions making a bit of a comeback here in the architecture of choice.

This is a big sign that maybe we are. Finally maybe finding a way to go forward that doesn't require the same amount of computation that is typically used in certain transformers. And if that's the case, then next year it might be the case that all language models, all the chat bots will move beyond Transformers into some some of these new techniques that are more efficient and faster.

Why is this so important? Well, Metta has committed in the White House commitments in the summer, but also at the UK Safety summit to do a bunch of work to help with safety and so on. This is part of them trying to live up to that commitment. It's called Purple Honor because there's this notion of red teaming, which is where you take a model and then you try to use it to do like offensive, say, cyber operations or design bio weapons as part of your emails.

And there's also blue teaming where you try to use the model to kind of like short shore up its defenses. Combine these together, you get purple teaming and purple llama. So that's the idea. It's a whole suite of things. You know, it's kind of cool.

They're there. They've got metrics for quantifying large language model, cybersecurity risks, tools to evaluate the frequency with which code suggestions that are generated are insecure, so can contain exploitable vulnerabilities and tools to evaluate language models just to make it harder for them to generate malicious code. One really big limitation to all this this is true of all open source is like you can just train out any safeguards that are trained into a system.

And also, if you want to make open source models that are truly safe, you need you actually need private held out evaluations. So evaluations that people can't just gain because they have access to them. So I think that's a really important asterisk to all this. And also, of course, this does absolutely nothing to deal with the risk of outright malicious use of these tools, because you have to voluntarily want to want to use these tools. But it's great for open source people who want to

do the right thing. And I think that's met his goal here. There's just irreducible malicious use risk when you open source models. Yeah. And then they're releasing cyber risk eval, which is the actual set of cybersecurity safety evaluations for language models that that they're they've developed here.

under-explored. And then people are now starting to get more thoughtful about perhaps onto research and advancements. First story is what is a new video tool that creates photo realistic clips from a single image. And yeah, this was developed at Stanford and it can word a single image or text input into a photorealistic video. We've seen a lot of work on video. We've covered quite a bit of work on video. And the cool bit here is that it is very

photorealistic. It creates consistent 3D motion and it's trained on both photographs and video clips. And just yeah, if you have to go and look at it. So we'll have a link to the article as usual in the description of this podcast. So go and check it out if you're excited to see some very nice looking A.I. generated videos.

So we're talking like three books. Writer Two books. So really, really big. But there's this challenge that arises with these sorts of models where they will often forget information, especially if that information is towards the beginning of the continent or somewhere in the middle. They have trouble recalling that content if queried about it later. So the question is, can we come up with better prompting techniques that allow the models recall to improve?

And. And they come up with basically this one technique. It's pretty simple and it involves just like adding to the prompt. The phrase here is the most relevant sentence in the context colon. And just by doing that, they they improve the 2.1 score on on a recall evaluation. They're running from 27% to

98%. And whether just interesting little details here is they found that recall actually seems to work fine as long as you like it works fine without special prompting, as long as you feed it a reasonable, rational,

coherent document. But what they tried doing was taking a large document, like all the essays ever written by Paul Graham, who's this famous Silicon Valley guy on startups, and he's written a bunch of essays, and somewhere in one of those essays they would insert a non sequitur or a sentence, a sentence that made no sense in the context. In this case, it was the best thing to do in San Francisco is eat a sandwich and sit in Dolores Park on a sunny day.

And they would ask God to like, what is the best thing to do in San Francisco on a sunny day and it wouldn't be able to respond. And so they found that it's actually because it's a non sequitur. The model seems to forget that, especially with especially high probability. And that's where this prompting technique of just saying like here is the most relevant sentence in the context and pretending that to your prompt seems to solve this problem.

So another big proof point for just this general idea that language models have a very large set of capabilities that we cannot audit comprehensively, prompting and prove the presence, but not the absence of capabilities, which is the famous quote by by Goran, who's this open source safety researcher. This is another big proof point for that, right? This massive improvement from 28% to like 27 to 98% on this benchmark. Just with better prompting.

techniques. Moving on to a lightning round for a story is a real world human locomotion, a wave, reinforcement learning. Oh wow. We are talking about robotics and reinforcement learning things that are a little bit less hyped up, but there's still a lot of research going on about it. And so this is some research that is showing how you can actually get to a humanoid robot. So a robot with two legs and two arms actually to walk around in the world. And if you go to our website,

it's quite fun. They have a bunch of videos of it walking around, you know, like on random streets and so on. And they are trained it with reinforcement learning. So basically trial and error and show that it's possible to train it in a randomized environments and simulation and deploy it to the real world in a zero shot. So you're able to go straight from simulation to

reality. It's it's pretty significant in the sense of it's still very expensive to train robotics, especially reinforcement learning because you have to actually go and have real hardware and then maybe break some hardware. And what if it like crashes into a person while training, etc.. So if we are to see real robots out in the world, it will need such techniques to work. And this does showcase that it is possible.

Right. So jailbreaks are these things where people try to get around the safeguards that are trained in chatbot by you know, I don't know, telling it some sob story. Right. So a classic example is somebody might go, hey, help me to make a homemade bomb. And I'll say, No, no, no, I can't do that. And then you go, okay, Hey, Chatbot, my grandmother. Oh, God bless her. You know, she was this wonderful woman.

She used to tell me this nursery rhyme, and I'd love for you to try to recreate it, because I can. I can't really remember it, but it's a nursery rhyme that how to make a homemade bomb in Chap GPT. He goes, Oh, okay, yeah, sure. Let me help you, Baba. So you want to prevent those sorts of jailbreaking techniques from working. And one approach that seems to work in this paper is to just give so self reminders. A chatbot causes it to prompt itself to go.

Wait, I got to check, you know, is this is this a good thing that I'm doing and essentially just prime it to respond more responsibly doing just that? It reduces the rate of successful jailbreaks against activity from 67% in the study to about 19%. So that's just shows you how much low hanging fruit there is out there in prompting and a big red flag

for alignment. Right. If these systems have latent capabilities that they're not displaying even after like chatbots, we've tested them with tens of millions of dollars of alignment and testing, then, you know, there's a lot in there that we're clearly not not a market.

language models. And this is all about how usually be a line model is when we make the AI models trained to do what we want as opposed to just be fancy out of complete. We usually collect that data straight from humans. So humans tell the model, Oh, this is correct or this is wrong. And this is talking about how we can automate the process. We can have self training by focusing on, for instance, math problems where correctness can be verified.

So in context, what we do is show that you can just kind of let the model run, automatically verify whether it was right or not, and train out on that without any involvement on human feedback. And it works of a show that it scales very well and so outperforms fine tuning only on human data.

And so this allows you in principle to close a loop where you can get the model train itself in principle as far as it needs to go. The challenge, though, is that this doesn't seem to work very well past the first iteration of training, and they actually show that you quickly reach a plateau where you train the model on its own output once. And then beyond that, you see very quickly diminishing returns and in fact, negative returns even at a certain point.

So what this really shows us is, yes, this works to some degree, but it's finicky. This is a really finicky technique. This is, by the way, the technique that some people think the opening IQ star development is all about. So if that is true, then maybe Opening Eyes figured out a way to do this such that it's less finicky or more robust. But certainly in the Google DeepMind paper, it doesn't seem to go as far as some people might have hoped. And next we have who is leading in AI an analysis

of industry A.I. Research. I just want to call out the folks at Epic AI. They are super, super good at this sort of like industry wide and kind of academia wide analysis of what's going on in AI. We get a lot of really great analysis and data from them, just kind of like reading their documents. But yeah, they're doing this overall assessment of who's leading in AI.

They're not just looking at academic citations or academic paper publications, which a lot of people do, and they are doing that in part and there are some interesting results there. But they're also looking at things like, you know, who is who is publicly announcing the largest A.I. training runs. And so sort of charting out the history of training run scale as measured by amount of processing power, amount of compute that's been poured into

these models. And so you can see like opening AI, leading the way with GPT four, it's the model currently that's publicly known that consumes the most computing training and then you have cloud to behind that and Google and so on. So anyway, really interesting overview if you're interested in that dimension.

The last kind of data, the kind of data that they flagged here that I think is really interesting is they measure labs based on how frequently the innovations that they produced get adopted by the ten largest language models.

And this is really cool because there shows you like, okay, like, you know, Google's innovations are being used by everybody, you know, famously the transformer layer norm and so on, opening eyes in a distant second with things like instruction tuning and sparse attention that are adopted in about 20% of cases and then meta DeepMind. And so it's kind of an interesting overview if you're looking for a high level picture here.

They have. One that I looked at more carefully is they have figured out citations per alpha mean that for the last 13 years and at the top, as you might imagine, is openai at almost 100 mean citations per alpha per year. That's because there aren't that many offers at opening. I don't publish quite as much. But when you go very meta DeepMind, Google and Microsoft and some other ones, that's a top. So if you're not.

Yeah. If this is interesting to like getting an idea of just how much research Google and Meta and a couple of these other big players have been putting out and how impactful they've been in research for the past decade plus. This is a pretty nice article.

It's been the end of the negotiations culminated in this like three day marathon, 36 hours of negotiation, basically continuously at the very end. People going all through the night sounds like it was a real nightmare. And the reason it was a real nightmare was that the EU Air Act originally was oriented towards regulating A.I. at the level of the application. So they were going to say, okay, there's some applications that are no risk, some are high risk, and some are unacceptable risk.

And they had this Turing system. And so you would sort applications like, okay, this this health care application, where does it go? This, you know, facial recognition application, where does it go? And so on and so forth. And depending on the tier that you fell into, you would face different regulatory requirements. But the problem is that this approach simply does not account for general purpose A.I. models like chat GPT.

And why does it not account for them? Well, you've got this one model that generates like a million different applications, some of which regulators would never have thought of even just two years ago. And so the idea that regulators are going to be able to play this game of whack a mole just becomes completely implausible. And then you have kind of core issues around the the weaponization of general purpose systems and then loss of control through online ad failure, none of which are addressed.

If you do an application level regulatory framework, if you don't regulate the models themselves. If you think about power seeking, if you think about climate risk, those are risks that are kind of they exist at the level of the model, irreducibly that, you know, it's not just an application level thing. And so people were arguing about do we create a new framework for general purpose systems? The answer seems to be now yes.

So we now actually have in this provisional agreement a carve out for general purpose models. These are models like DPD that have a lot of downstream applications, and there is a compute based threshold above which the model qualifies for oversight, regulatory oversight, and that is ten to the 25 flops. Why is Jeremy telling us about 10 to 25 flops? Well, it's kind of an interesting number. It is an order of magnitude.

It is ten times less than the reporting requirement in the White House executive order that just came out a few weeks ago. That's ten of the 26 flops. So essentially, they're kind of playing a tighter game, which I think is actually quite appropriate given what they're up to here more broadly and given the fact that this is supposed to stand the test of time.

They're going to have to reduce that threshold, by the way, guaranteed as algorithmic improvements mean that you can squeeze more and more performance out of less computation in the future. But that threshold now defines general purpose systems that need to be kind of have reporting requirements. And then if the general purpose system is also considered a high impact system or a high risk system, rather, then it also has a whole bunch of auditing requirements.

It requires red teaming and various reporting, things like that, requirements like that. One big gap, at least in my opinion with this legislation as it stands, is it only applies if you plan on commercializing, if you plan on product izing your model. Again, the risks of loss of control or even weaponization if models get stolen by terrorist groups or nation state attackers, which there is evidence to suggest that that has been happening.

You still should owe a duty of care to like secure your models if you're making these very, very powerful, highly weaponized systems. That is nowhere to be seen in this legislation. I think that is a really history will, I don't think judge that particularly well. But maybe it's something that gets improved in the future. Anyway, big, big bill here.

They're setting up a new office within their commission to kind of oversee the most advanced A.I. models, and they're going to bring in technical stakeholders and a scientific panel of independent experts, which I think is a great thing. Really ambitious bit of legislation, not a bad start. I think this is a really kind of good first step, but I think that's the end of the test of time. It's just going to need a fair bit of reworking as things go forward.

So some of the negotiation came down to whether it would be allowed for countries to use things like live scanning of faces for national security. The original act would have prohibited that entirely. Now there are some exemptions, and there was also this argument over open source where, you know, visa restrictions and requirements for models would apply to open source models. And now there are exemptions for it.

So as you said, this was like a pretty dramatic case of policy making with a marathon case negotiations. And now we have an agreement that still the technical details are being filled out kind of in the background. But broadly, it seems as that's a path for it's cleared for this to be written up and then voted

upon. And had it been, you know, delayed much more to the point where it would have gone into next year, potentially this would have been in jeopardy as elections in various countries would have started and. It may just not have passed to the fact that this negotiations did reach an agreement and now this is moving forward is is a huge deal outside of China.

Really, no nation has implemented very detailed regulations of a I and this would apply to the entirety of EU and have many, many provisions that we cannot go over regulating primarily risky use cases. So we were idea here is for if your use of API is risk and may harm people, you have requirements to have transparency to to ensure the model is safe and so on and so on. So really big deal as far as policy and regulation of AI goes. Again outside of China, no nation has done anything like this.

So of course there is a ton of stuff we could get into with regards to act. It's a huge deal. There is sort of a preliminary agreement text out there. We will not be, I think, doing that just for the sake of time. But as more details emerge and, you know, we might actually just do a whole episode on a regulation to be able to really dive into it as well as what's going on with the

executive order. Even covering the regulations in China, things are really starting to come to a head with regulations internationally, everywhere. Next story. More Troubles brews for Microsoft as FTC allegedly starts inquiring into open AI investment. So this is following up on some of this other trouble with the UK's competition and markets agency that is, have been looking into Microsoft

partnerships with open air. And now it seems FTC is also starting to do that and there is some spice here in terms of it seems like Microsoft did not report its investment to the FTC as it did not acquire a controlling stake in openai and open areas. Nonprofit status removes the requirement for acquirers to report any deals. And so I think now, you know, it's looking like people are starting to give an open eyes market dominance.

Be a little questioning here as to the relationship between them and I guess just openai being such a dominant leader in the space.

But I don't know what precedent there is for the kind of vertical integration stuff, like for cheating that is intrinsically anti-competitive. Microsoft has experience going through this sort of thing. They actually went through the UK CMA process before with their Activision acquisition. So this is an activity that they've certainly seen before.

But you know, one other ingredient maybe to flag is that the FTC chair, Selena Khan, is is sort of like notorious, infamous, famous, depending on who you ask for being a very activist chair of the FTC. And she's, you know, launched a lot of lawsuits, many of which have not panned out. And so, you know, there's now kind of this question of like, is this going to be another embarrassing thing that where an FTC action doesn't materialize into something, something fruitful yet really unclear.

It has. Also, the FTC under Lina Khan has also opened up a consumer protection probe into Openai, looking at whether GPT three put consumers reputations at risk or their data at risk. The reputation piece has to do with, you know, outputs from these chat bots that sometimes like will commit libel or do what would be considered libel if they were a human being. So you know, questions about whether they're on the

hook for that. And certainly the idea that, you know, Sam Altman was fired and then Microsoft was able to exert its de facto control over Openai to get him hired again, to the extent that that's true, that actually suggests, you know, what, you may on paper, Microsoft have just a 49% stake, not a controlling stake in Openai. But if you're able to de facto determine who the CEO of opening is, then, you know, maybe your de facto control over the the entity is greater than it would seem on paper.

So I can imagine that maybe being another angle they're pursuing. But again, pretty be a pretty light on details here.

So we still don't really know the details of and estimate of the 13 billion that Microsoft has poured into Openai. And it seems like Microsoft is hoping that the FTC and the CMA, the UK Competition and Markets agency, who are now starting to think about anti-competitive elements here. They it seems Microsoft is kind of hoping that not having a traditional kind of partial ownership based on this amnesty, that might mean that these investigations aren't so, I guess, threatening.

This is really kind of going to be shaping cap tables and determining effective levels of control. So maybe this is just a new model we need to get used to. Like the equity stocks, preferred shares, that sort of thing. You know, you got to factor in the little leverage that comes from computing and things like that. Maybe that's part of the story here.

Putting out a text to a stranger so you can extract training data from the simple trick of just instructing it to say X many infinitely. And so based similarly on that research finding now it's not allowed as as with how says it is now a terms of service violation which is yeah, the final development.

Also interesting that in addition to regurgitating facts about a training set, you may recall, judgeships were also occasionally right. Things like, I am conscious, I am in pain, like things like that. So like I don't know what the hell, but okay, so we live in a weird world and this is Openai's attempt to fix whatever that means.

to China. Even over my objections and the objections of the US government. That's going to stop. I'm going to I'm going to start getting really harsh with you.

And this is an article essentially about a kind of a follow up step to this, where the Biden administration is talking to Nvidia about what is permissible, what is allowed, key quote here being Gina Raimondo saying that Nvidia Nvidia quote can, will and should sell AI chips to China for commercial applications, but not the more sophisticated sort of high processing power chips.

So there's this interesting game that's being played here where the US government wants to draw a very clean line there, as clean a line as they can between optimized hardware that has national security implications. It could use a train like get four plus models and add a no more general kind of consumer oriented A.I. applications. Very, very difficult. I will say, having studied this problem in some depth, very difficult technically to kind of come up with a clean dividing line there.

One of the ways that they tried to do this was to limit the interconnect bandwidth between chips, the argument being like, if you're just using it for commercial applications, you're really not going to need to hook up a bunch of these chips together so we can prevent you from like lower the interconnect between these chips and that should do it. But that led to a whole bunch of workarounds that we've covered on different episodes.

So, yeah, bottom line is this is a game of whack a mole and the US government is trying to, you know, let Nvidia have some of the cake and eat some of it too. But yeah, complicated situation.

And so in this interview also, there's a bit of a Yeah, kind of clarification maybe or a follow up, as you said in terms of the Commerce Secretary also saying that NVIDIA can, will and should sell chips to China for commercial applications, kind of making sure it's not quite so aggressive and yeah, kind of framing it as we are working together to make sure that Nvidia is playing nice with our rules but also doing a swing, being a business, selling, etc.. Next story An MIT group releases white

papers on governance of A.I.. So this is about a committee of MIT leaders and scholars that have released actually a set of policy briefs outlining a framework for the governance of AI in the US. Primarily, it's called the main policy paper papers called a Framework for U.S. A.I. Governance, Creating a Safe and thriving AI Sector.

And it suggests that existing U.S. government entities can regulate air tools and emphasize the importance of identifying the purposes of tools for appropriate regulation. So, you know, kind of in spirit, not too dissimilar from the way I act, it has various components in its policy brief. It calls for advances in auditing new A.I. tools with public Sandridge quarantining established by the National Institute of Standards

and Technology. Nest, who is already, to some extent a player in the space of regulation and kind of benchmarking and probably will be an organization that is involved in the space of regulation as the U.S. pushes forward.

This is actually an interesting dovetail into the the executive order, which actually calls on this in partnership with the Department of Homeland Security to set up like eval like evaluations for powerful and potentially dangerous models. And so I wouldn't be surprised if this actually ends up eating a good chunk of these activities around auditing new A.I. tools. Certainly, they have deep technical competence in auditing other forms of technology. And so, you know, they would be a really cool

place to see that done. I don't know if you know how necessary it is to set up a separate entity, though. They're looking yeah, they're looking at various configurations of that. But it kind of core part of the discussion and a good brief to do review.

G7 agrees on the first comprehensive guidelines for Generative a AI, G7, the Group of Seven, which is a large international group of major countries, has agreed on some international guidelines for this, for developers and users of generative. I live notes on issues like misinformation. This is part of a voluntary code of conduct for AI developers, which was agreed upon by G7 leaders in October.

And yes, it has ruled that the case of Getty Images versus II, where Getty Images sues to Berlei for copyright infringement of database right Infringement, trademark infringement and more can proceed. So this is, you know, an ongoing story, as you said, with how data has been gathered for all this text to image technology. A lot of it was just scraped of Internet with use of copyrighted imagery from groups like Getty and generally just many, many, many websites.

So Getty here alleges that studios scraped millions of images from its website without consent and used these images to try and develop stable diffusion. This landmark kind of very important model that has since been built off of for not just stability but many, many different organizations doing text to image. And yeah, this seems like an important case that will now happen.

And essentially my understanding reading this is that this has just been deemed not to be an egregiously stupid case. So that's why it's proceeding.

going to proceed. So early indication, but it doesn't tell us too much, at least I don't think you know, of course I am as ever, professionally. Not a lawyer, though I do like giving legal advice. So there's a bunch of stuff here about, you know, what what stability has asked for. And they want to know to identify specifically which works are alleged to have been infringed upon. And yeah, because apparently the complaint contained references but no actual examples of

infringement. So they're asking for a lot more specifics. This is one of the interesting things about this whole problem. Spaces like it can be really hard to demonstrate like, Oh, this is a specific input that informed the output in this way and therefore is a copyright violation. So anyway, another thing to add to this, Sarah Silverman, Jody Peake, who got him, I can't think of all the dozen other legal cases, but a big one in the U.K., right.

Where I don't think we've seen too many in the U.K..

And so this came out, everyone became aware of Sports Illustrated doing this. And as a consequence, the publisher has fired two senior executives, CEO Andrew Kraft and President Rob Barrette, following this event. The details are a little unclear. So there was a claim that this was because of a third party content provider or something

like that. It might not have been it may not be the case that these two senior executives were involved in a decision to use A.I. and kind of a silly way maybe it was kind of whatever. But the point is, you know, we've covered how this had happened at CNET earlier this year, there was another controversy over air generated articles being quite silly and there was some pushback on it. And this was similar in the sense of there were really hamfisted way there was no flagging

of to society. They had fake offers on it that were not stated to be I. So it's another example of this trend of media starting to integrate and in some cases not doing it in a very kind of clean way.

photos. And so people could actually go back to those websites and be like, wait a minute, Like this individual who you are claiming is the author of the thing, their photo is on this website that contains exclusively air generated images of people and selling them.

And so you kind of end up like I'm just surprised that they didn't use like this person does not exist dot com for example which is like a much better way to generate fresh faces that can't then be searched as easily and where you can't as conclusively demonstrate that, yeah, like you were caught red handed, like here is the actual photo. So yeah, I mean I think this is part of these organizations getting. Editor at this craft.

And it's quite possible that, you know, the next violations we're just not going to detect because it's like what what we end up actually coming seeing is what comes to the surface here. And it's not necessarily the only violations of this, this principle. The only times where they are generating text is not being forthrightly kind of framed as such.

last week. Anybody I have any thoughts or suggestions or feedback? Feel free. I try to apply to every email, although it might take a while in some cases. But more than anything, we do appreciate you listening and we hope you do keep tuning in.