OpenAI GPT 5.6 Launch, Apple’s On-Device AI Play, & China Allows Nvidia H200 Chips

22:51Aaron Tilley:And I just want to translate a bit for folks. You say parameters. This is coming up time and time again. Parameters is basically just the micro decision that the model has to make at any given moment. Is that the idea? I mean, just think of parameters as roughly the complexity of the model. You know, the GPT, the Claude, Gemini, the big models, these are trillions of parameters. And it's just roughly the complexity of these models. Right. And I'm also thinking about how much space it's going to take on my iPhone because I can never seem to get enough space. So that gray block that tells me, oh, application settings or whatever is taking up 30 gigabytes.

23:39Aaron Tilley:It's like, man, what is being stored here, really? I have no idea. You had the chance to talk to Prism ML's CEO. What were the big questions that you went into that conversation with? Yeah, I mean, I think primarily I'm just wondering how much demand and interest there is for running on device. I mean, this is Apple's agenda, but is this really an industry-wide interest and sort of like how do we fit more on device? is that of interest to the industry? And what he said, and we'll see what happens, but in a few years, he said 95 % of this AI compute will be running on device, whereas only 5 % of really the leading edge is going to be running in this cloud.

24:27And if that's true, I mean, that would be a massive shift in what we currently see, where it's just like priority is data center build out consistently, constantly, really trying to build out more and more compute in the cloud. And if we do can move more and more on our device, then the economics of the whole industry changes. And does it really necessitate this like massive data center build out?

24:55Aaron Tilley:And so, I mean, that's kind of interesting because it seems like he's suggesting the entire data center build out is basically not useless, but effectively for no good reason. And I mean, it does sort of lead me to believe that we are sort of one breakthrough away in this edge inference category from all that data center compute may be becoming not so great an idea. It's kind of interesting. The question is really up in the air. We'll see. I mean, right now, the advancements of AI is happening at still such a degree that every week there's a new model, a new iteration, that continuously updating the on-device is unpractical.

25:43And really just keeping in the cloud is where you want it. And I mean, for a lot of reasons, cloud is better for especially enterprises. But yeah, I think it's a debate in the industry right now. We'll see where it goes. prism and that all just really, I think, you know, shows a different path forward. Great.

26:02Aaron Tilley:Well, Aaron, I want to thank you for coming on. That is Aaron Tilly, our Apple reporter here at The Information. The Information has exclusive reporting that China is giving AI companies at home the green light to buy a limited number of NVIDIA's H200 chips. My colleague Chenor Liu published that story, and our Asia Bureau Chief Jing Yang spoke with her about her reporting. Here is that conversation. Hi, Cher. You reported that China plans to let some of the country's top AI companies to buy a small amount of NVIDIA H200 chips. Could you unpack for us what you find out?

26:43Qianer Liu:Sure. What we found is that Beijing is preparing to let some of China's biggest AI companies to buy a limited number of NVIDIA's H200 chips. This would include companies such as Alibaba, Vidance, and DeepSeek. But the approvals are not completely open-ended. Companies need to explain how many chips they need and why they need them. And this is also not a full reopening of China's market for NVIDIA. Beijing is still trying to limit tech companies' reliance on US chips. but it is also recognized that Chinese AI developers now are facing a real shortage of high-end chips for training AI models. Chinese officials want these NVIDIA chips used mainly for training while pushing companies to use Chinese AI chips for inference.

27:41Qianer Liu:So basically, the message from Beijing to this Chinese company is use NVIDIA chips whenever you must, but use Chinese chips when Beijing thinks they are good enough. I see. Remind us, what is the H200 chip? And how does the H200 compare to, you know, NVIDIA's latest generation, the Blackwell, and, you know, other chips made by Chinese companies like Huawei? The Edge 200 chip is an advanced NVIDIA AI chip based on the company's hopper architecture. It's not NVIDIA's latest or the most cutting edge generation anymore, but it remains very powerful and useful for AI workloads. Compared to NVIDIA's BlackWild chips last year and the Ruben chip which is on production right now, the Edge 200 is a step behind.

28:45Qianer Liu:But compared to Chinese alternatives such as those from Huawei or Cambricon, Alibaba, Alibaba's T has, the Edge 200 chips is still very hard to replace, especially at model pre-training. Because NVIDIA is not just selling chips, it is also selling software. It's selling a full ecosystem. Chinese chips are improving, of course, in raw materials, in raw performance, but many developers are still finding NVIDIA's hardware and software easier and more reliable and stable for training their AI models. So in short, S200 is no longer NVIDIA's cutting edge, but in China, it is still a very high end chips.

29:31Qianer Liu:BlackWild is obviously ahead of it, but Huawei and other Chinese alternatives are still trying to catch up, not only on hardware side, but also on software development. I see. And I remember that it was US President Trump who approved the sale of NVIDIA 600 to China back in December. So it was more than 60 months ago. But you are reporting that this is the first time that the Chinese government actually is allowing these chips to come in at a relatively contained scale. Could you tell us what has been happening the last six, seven months and why the Chinese government didn't allow NVIDIA to the H200 into the country previously and now why are they changing their attitude?

30:20Qianer Liu:The reason it's pretty straightforward. Chinese AI companies now need more compute resources. They are trying to train larger and more complicated models but they just don't have enough high-end chips and the recent US government's crackdown on chip smuggling makes it more difficult for a Chinese company to find a world around. So this shows us that Beijing is probably becoming more pragmatic. Their long-term goal is still self-reliance but in the short term officials appear willing to make selective exceptions if the chip shortage, the training chip shortage, threatens the competitiveness of China's AI industry.

31:08Qianer Liu:So I would say this is a very interesting compromise. Basically Chinese companies still have just most have limited access to these S200 chips with a very clear condition provided by Beijing. And this is also a very strong signal that Beijing still wants the Mystic chips remain the priority even though they allow Chinese companies to buy a limited number of H200s. So how should we understand this sort of relaxation on the VDA chips that you are reporting? What does this mean for VDA? Should we expect VDA to be able to regain some of the market share in China from this moment on? I will say it would definitely, it's not definitely, definitely not a full reopening of China market to India because like we mentioned in the story, China will only approve a very limited number of H-200 ships and the reason why they were doing it is because Chinese company thinks and repeatedly warning Chinese government that they don't have enough AI chips and they need competitive high-end chips like Edge 200 to help solve the current computing resource resources shortage.

32:33Qianer Liu:But maybe as Chinese chip companies improve their products, for example, Huawei's next generation AI chips, which expected to be released next year could have computing capability comparable to S200. But I think maybe in the long term, China still needs a lot more chip-making capacity and better memory supply and software systems which the developer can use and trust, which I think is probably the most difficult part. So it sounds like what's happening is that the Chinese government is letting a limited amount of videos advance the chips into the country to tie over this, you know, searching demand and a shortage of computing power until as long as China may be able to provide all its needs on its own.

33:31Exactly. I see. Oh, this is very, very interesting. Thank you.

33:38Aaron Tilley:the information has exclusive reporting that finance is rethinking how it works with the department of justice i want to bring on leo schwartz who covers all things technology and washington dc to walk us through what he's found out leo welcome back to the show it's great to have you here your story starts with a memo at the justice department tell me about this memo so in early june a memo went out to attorneys at the justice department working on crypto related cases, basically saying that Binance, the world's largest crypto exchange, would be cooperating a lot less with any cases that prosecutors are working on.

34:14And that involves things like seizing crypto that's believed to be used in illicit asset flows, tracing it, freezing it so that it can't get out of their grasp. In the past, Binance was much more cooperative in how they would work with prosecutors on those cases. And what this memo basically laid out is that they could expect much less cooperation from Binance. And the context of this is that in 2023, Binance settled a massive, massive case with the Department of Justice, along with the Treasury Department and another agency responding to money laundering charges. And after that point, Binance obviously was incentivized to work a lot more closely with the department.

34:57it.

34:59Aaron Tilley:And so behind the memo, did you get a sense at all from your reporting as to why this change this change was happening from Binance on the ground? It's unclear. I mean, Binance for years did not have any official headquarters or wasn't domiciled in any place. It's still that way. However, they did get registration with Abu Dhabi, the Abu Dhabi financial regulator. It could be really into that and new policies coming out of that that different regime that Binance is under. But at the same time, obviously, the Department of Justice is the leading law enforcement agency in the world. So theoretically, that they would have influence over other authorities and they would still be able to operate as business as normal.

35:42Aaron Tilley:We should say here, what did Binance tell you about this when you went and asked them about this? So this is the interesting part. The memo went out on June 2nd and said pretty definitively that the Justice Department was expecting that the changes would be made as soon as June 8th. It even said that prosecutors working on cases should go through their normal channels now to make sure any funds that they needed to be seized or freezed were in fact done so before the changes went into effect. However, when we went to Binance for comment, Binance said that those changes never actually happened. They denied that they were changing how they were operating with law enforcement.

36:19Hmm.

36:20Aaron Tilley:Okay. And I mean, I guess the question remains to be seen then is that we have to wait and see how this really plays out if these changes actually do happen. I guess right now it's sort of like feels a little bit like that executive order type of thing, which is like we have this paper saying something, but we don't actually know what's going to happen down the road. I think the bigger story is that whatever the Justice Department learned was significant enough for them to not send an internal memo to a few lawyers, but this went out to a mass group of lawyers who were working on crypto cases.

36:55So clearly they thought these changes were happening, or at least their relationship with Binance was changing. And the background context of this is after that 2023 plea agreement that Binance signed, they agreed to two monitorships or monitoring programs that made sure that they complied with the terms of that plea agreement. One was with the Justice Department. One was with Treasury, which we've reported on earlier. We've learned in the course of this reporting that the DOJ monitorship is effectively paused and has been for over a year, and that Binance and the Justice Department are an act of negotiations to formally end it.

37:28There's also another key context that's important to understand, which is that under Trump, who is obviously a much more pro-crypto president than President Biden was, the Justice Department in particular has taken a new approach to crypto where the Deputy Attorney General at that point, Todd Blanche, sent out a memo saying that they weren't going to do the regulation by enforcement of the previous administration. So the DOJ has been much more lenient when it comes to crypto cases. And I think those two pieces of information are important in the context of this memo and trying to figure out why it might have gone out.

38:05Aaron Tilley:And I just want to make sure I understand, you know, a monitorship in this scenario, what does that mean, especially for a company that Binance, I mean, they at one point were operating in the U.S. by with Binance U.S., but now, I mean, there's sort of rethinking what it means to operate here, right? So what does a monitorship mean here and what's the status of that? So Binance, the mothership, the bigger company, ostensibly never operated in the U.S. So obviously with the 2023 case, that was predicated on the fact that, in fact, it did have a U.S. nexus. There was that subsidiary that separately managed Binance U.S.

38:43that did operate in the U.S., although it effectively shut down. It's sort of trying to build back operations. But the monitorship program was about Binance, the larger Binance. And it was basically saying that under all of these terms of a plea agreement from 2023, when Binance admitted that it didn't have an effective anti-money laundering regime, that it didn't effectively do know-your-customer checks on its users, that it was violating Bank Secrecy Act key provisions. They said as part of this plea agreement that they would make sure to adhere to the terms of the plea agreement to cooperate with law enforcement.

39:20The point of a monitoring program is you have this third party, usually an outside law firm, who make sure that Binance is complying with those terms. So the fact that the DOJ one has been effectively paused, I think, would be worrying to a lot of people who are watching what's happening with crypto and with Binance. But the Treasury Department one is still active. We report in this story.

39:41Aaron Tilley:So the idea that Binance has said that we're not going to cooperate as much, according to your reporting, what do you think the net impact of this is for the DOJ? I mean, is this just a thorn in their side that it takes longer to get answers back from inquiries? Is this the DOJ not being able to pursue certain programs specifically? What's the impact here? Well, so I think for now, we have to assume that the changes haven't gone to effect and that Binance is cooperating as normal. However, if you look at what the changes laid out in that memo i spoke with a former prosecutor at the doj who i shared the contents of the memo with and he basically said if those went to effect it would make it nearly impossible for prosecutors to be able to work on these crypto cases one one part of the memo said that if a a prosecutor has a subpoena from the u.s but wants to be able to trace freeze and seize assets in a jurisdiction outside the U.S., they would have to do essentially a mutual legal assistance treaty with that country, a process that could take over a year before they could even freeze the funds.

40:53And crypto obviously moves very quickly. It's dealing with really significant cases like sanctions violations or pig butchering, which is when people get scammed out of their life savings. And if prosecutors lose the ability to actually be able to freeze those funds, it would really incentivize cyber criminals who use crypto as a means to commit their actions to be able to look at Binance as a venue that the U.S. really is not going to be able to crack down on anymore. Great.

41:18Aaron Tilley:Well, Leo, I want to thank you for coming on. That is Leo Schwartz, our tech and politics reporter here at The Information. E-commerce has come in and out of focus for big AI labs like OpenAI, but there is also a category of startups dedicated entirely to this space that are quickly gaining traction. Daydream is one of those companies. The AI shopping company is backed by Forerunner, Index Ventures, and Google Ventures, among others. It has 1.5 million people using its platform. I want to bring on founder and CEO Julie Bornstein for our conversation. Julie, welcome to the show. It's great to have you here.

41:53Thanks. Great to be here.

41:54Aaron Tilley:I was using Daydream this morning. I was playing around with it. It's a pretty cool platform, I got to say. Thanks. I'm glad you're here. I mean, I could explain it, but you'll do a better job. Bob, tell us about what the vision is behind Daydream and what the current state of the tool is. Yeah. Well, if you think about it, the way that we naturally shop, we're fashion-focused right now. And obviously, we could apply the same technology to lots of verticals. But we started here because it's a really hard, overwhelming category for most people. And, you know, the way that a person thinks about what they want is in natural language.

42:33And so the idea of now that AI can actually take natural language and understand it and interpret it and give you product results is really a dream come true for most shoppers. And so instead of needing to actually know exactly what it is you want, you can say, here's the need I have. I'm going to speak at this event. I'm going to a wedding in this location. I want to wear something that is not too revealing. I want sleeves. I'm thick in the middle section. And so the things that we have always asked when we talk to a person is now available by leveraging AI and the catalog that we have, which is the largest branded fashion catalog that exists anywhere.

43:14Aaron Tilley:And how many different retailers do you have included in your catalog right now? We have over 10 ,000 brands on the platform, and we work directly with retailers so that multi-branded retailers as well as brands directly themselves. And we have everyone from Uniqlo and H &M on the sort of lower priced end to Gucci and Saint Laurent on the higher priced end and everything in between. and the idea is that we learn about each user as she or he shops so we know the kinds of brands and the price points to show as well as your size and style so that we help you sift through this overwhelming supply of product by finding what's right for you.

43:57Aaron Tilley:And how do you make money on your end? When a consumer finds a product they want, they click to the retailer to buy and we get a commission. A commission, okay. And how much revenue is a company generating now? We get about 20 % of every sale that's generated, and we haven't reported revenue numbers yet. Cool. You know, one question I wanted to ask you is I saw this clip of Brian Chesky speaking online, and he was talking about AI and e-commerce. And one of the things he – it was an opinion. I mean, they haven't released whatever the product is, but he basically said that he didn't think chat was the right interface for AI and e-commerce.

44:42Aaron Tilley:And I wanted to ask you what your opinion is on that. Do you agree with that statement? Do you disagree? So it's interesting because I don't think it's an absolute answer one way or the other. I've been in e-commerce my whole career, and about 25 % of users who come to a site like Nordstrom or Sephora use search. And so if you think about it, there's lots of modalities and reasons why people come to shop. Sometimes, about a quarter of the time, they have a very specific need and they know what they're looking for. And in that case, chat is amazing. Whether you type something more traditional, like I'm looking for a lace red dress, or you get very detailed with what you want, chat actually provides a really great, much more rich version of search.

45:24And so what we know is that that is one modality that consumers want. What we have learned is that there are other modalities, too. I've known this all along. So there's the inspiration. How do you just show people something that's interesting and relevant to them based on their interests? There's uploading a photo, which we offer at Daydream. So you can literally screenshot anything or take a picture on the street, and we can show you the products that are similar to that. And then you can find the things that are sort of more in your price range, let's say, or a better version of that style for you if it's not exactly the right product.

45:57So I think there are multiple modalities and search is always going to be a big one for shopping in general. I think when you look at spec-based products, more like a washer and dryer, let's say, as an example, there are different kinds of information you want to put in because you have a certain type of space that you have to reach and there's a certain budget and there's different specs that you want. So definitely sort of the discovery and refining process requires a lot of different modes to be able to help someone find what they're looking for.

46:27Aaron Tilley:How have you seen usage trends for that, for the image-based recognition discovery that you guys introduced? I don't know how long ago you introduced it, but I mean, what's the mix right now in terms of how many people are discovering things through chat versus the image product? Yeah. So interestingly, what we do is we either let you come in and chat from scratch. You can click on some other chats to get some ideas or a feed of product, or you can upload an image. And what we find is that it's about 50 % of people actually key in specifically what they're looking for. And then between the other 50%, it's a combination.

47:04About 20 % are uploading photos and about 30 % are sort of clicking on something that appeals to them and then starting their journey from there.

47:12Aaron Tilley:So that would mean that 80 % are still finding it through chat. Yeah, about 50 % are keying in the chat that they want, whereas about 30 % are sort of finding a product or a preset chat that interests them and then searching. So it's kind of more like, you know, what you would call a traditional browse or, you know. Got it. Yeah. Got it. So you've got 10 ,000 brands represented on your platform, which gives you a pretty good view into how brands are thinking about the AI era. I mean, on one hand, you have the challenge of trying to navigate the advertising landscape and this new world of advertising.

47:58Aaron Tilley:Then you have the AEO, GEO bucket of discovery. And then you have platforms like yours, which are sort of newer platforms, you could say. What's the biggest concern right now from them? I mean, are they seeing ROI from the AEO, GEO bucket? Is advertising working for them in these types of platforms? What's the conversation like? I think it's too soon to say. I think everybody is focused on enriching their content so that it shows up where it should show up. And so that is a big conversation right now. How do we make sure that everything about this, you know, traditionally e-commerce has been kind of hit or miss on how detailed the product pages, and how clear what the product is.

48:44And I think that is clear to everyone now is you want to make sure that you have as much information available on the product page as possible. I think some people are starting to try AEO and GEO. No one said, you know, wow, it's really changed our business yet. They are seeing searches coming in from some of the big LLMs. So that is a difference. And I think everybody is still very much trying to figure it out. On our platform, it's not ads-based. It's just purely commission. And we are showing the consumer specifically what we think they are most likely to want. So there's a little bit less of a sort of game to play on Daydream, which I think is nice for the retailers and brands.

49:25Aaron Tilley:What about loyalty, brand loyalty in this era? How do they think about protecting brand loyalty, which is not a new issue, I should say. I mean, people are discovering stuff through Google, stuff like that, you know, not on websites directly for a long time. Do you think loyalty is going to be an increasing issue for brands? Or do you think we could go the other way and maybe, I don't know, maybe DTC could come back somehow? Well, I do think that loyalty matters a lot. And I think that for their best consumers, making sure that they're continuing to invest in bringing their users directly back is the most cost-efficient way for them to continue to drive their business.

50:08I do believe that there is the opportunity for brands that are maybe a little bit newer to get discovered in a new way. One of the things we see in Daydream is we ask people if they have brands that they like to share them with us. But then we also help them discover new brands based on their taste. And that is one of the things that users love the most is the ability to discover new brands. So I think it's going to be, you know, the one thing I will say is that brands aren't going anywhere because the LLMs and Gen AI can't replace an actual physical product and a brand. But I do think how they're discovered both through the LLMs and through platforms like Daydream is, you know, very new for them and they're in experimentation mode.

50:53And I find that they're quite interested and willing to test quite a number of things in order to see what really works.

51:00Aaron Tilley:So you've chosen to focus on the discovery category of e-commerce. On the other side, there's the checkout and the payments, which we've reported on extensively here at The Information. OpenAI had initially seemed to have made a more concerted push towards a checkout's payments type of infrastructure. We reported that they have taken a step back from that. Did that surprise you at all? I've actually heard recently that they're possibly going to be trying again. So we'll see. Say more about that. What do you, what have you heard? I just, I just heard, you know, they may be coming back around to try and shopping again.

51:42I think it's inevitable. I think shopping is a huge category, obviously. And I think all of the LLMs should be thinking about it because the, really what they're great at is discovery. I think checkout is to me kind of a secondary problem. Right now, checkout is pretty simple across the web. And so it's interesting that that's what everyone called shopping because I think it's much harder to actually connect people to the right product. But the LLMs have the ability to do some really interesting things with the data that they have around products. And whether they decide to invest in it or acquire companies who are doing it will be interesting.

52:20But there's no doubt in my mind that, you know, the checkout is the last mile. And sure, once you've actually found the product you know you want, make it easy for me. But the much harder part is figuring out the product you want. And this has been harder and harder as the web gets, you know, more and more dense with products and brands and retailers. So it's funny.

52:40Aaron Tilley:I would have thought it would have been the other way around. I mean, I would have thought that the payments and the checkout would be the harder issue to solve. And that's why, you know, we see Google with I think it's like the universal commerce protocol, stuff like that. I mean, is the payments layer, I mean, is that not the most complicated part of all this? I'm not suggesting the payments layer isn't complicated. I'm just saying it's not the hardest problem for the consumer. So that was my point. Yeah, no, I think it's still, you know, for us, the way we think about, you know, checkout is whatever, however the standard evolves, we will adopt.

53:16But we're not going to try and create it or set it because it is very challenging. And it just requires basically every player to hook up into a system. And so, yes, that's a big undertaking. My point is more that the harder problem for the consumer is actually finding the right product to buy. Right.

53:34Aaron Tilley:And last question for you. I mean, one thing that we've talked about a little bit on the show is the access to data from brands and the right catalogs and stuff like that. And this is a reason that we've talked a little bit about Shopify sort of having a bit of a leg up here because they have access to this information already through their platform. I mean, can you talk about how big a lift it is to sort of integrate with the catalogs of all these brands? And then also, you know, frankly, the extent to which why you think you can do that better compared to just OpenAI or another platform doing all that legwork itself.

54:15We ingest the catalogs from our partners and then we enrich the product with lots of additional information. So we understand both subjective and objective elements of the product. So we've built this very deep understanding of the vertical across the board. And that has allowed us to know when you ask for, you know, you want to look sexy at a wedding, we know what product is appropriate for that. That exercise of going very deep on understanding the vertical has been really our leg up and why I think it would be harder for, you know, an LLM who doesn't really focus on understanding the product well to be able to do this easily.

55:01So I do think that, you know, you would need any of the sort of search engines would need to go very deep on understanding at a vertical level the category very well, because you obviously can understand and we already do understand with millions of sort of queries and prompts underway what people are looking for. Now, how do you actually match that to the right product and, you know, know this person and match it not just the query, but also to the person is really where sort of we've spent all of our time. Great.

55:32Aaron Tilley:Well, Julie, I want to thank you for coming on. That is Julie Borenstein, founder and CEO of Daydream here on TITV. That does it for today's show. A reminder, we are on this stream Monday through Friday at 10 a.m. Pacific, 1 p.m. Eastern. If you can't make it then, episodes are available on theinformation.com, on our YouTube channel, or wherever you get your podcasts. Make sure to follow us on social media, on X, on LinkedIn, and on TikTok, and on Instagram as well. if I could just get that. I'm already excited for our next show tomorrow. Have a great rest of your Thursday. Bye-bye for now.