Apple Suing OpenAI for Stealing Trade Secrets, Former Meta Director of AI on Muse Spark

13 Jul 2026 · 47 min · 24 chapters

Ask about this episode

Ask anything about it. ChatGPT or Claude reads this page and answers with the times it was said.

Connect VO and ask about every podcast you hear, including the moments you saved. Add to ChatGPT · Add to Claude

In short

Episode topics: Apple sues OpenAI over alleged trade-secret theft; DTCC demonstrates blockchain-based tokenized stock settlement for select assets; takeaways from ICML AI research conference; Laud Institute reaction to recent frontier model releases and AI eval challenges.

Guests and backgrounds

  • Aaron Tilly, Apple reporter at The Information; covers Apple and tech industry legal/business developments.
  • Yueqi Yang, crypto reporter at The Information; wrote about DTCC blockchain tokenization.
  • Stephanie Palazzolo, AI reporter and author of The Information’s AI Agenda newsletter; attended ICML in South Korea.
  • Brayden Hancock, research partner at Laud Institute; former Meta director of AI.

Key claims and notable examples

  • Apple alleges OpenAI device-hardware team (Tang Tan, Chang Liu) stole proprietary Apple knowledge; includes “show-and-tell” parts during interviews and alleged ongoing access via an Apple laptop.
  • DTCC (DTCC) will run live production trades on blockchain this Wednesday for stocks/ETFs/treasuries; blockchain used for tokenized transfers/collateral, not clearing/net settlement.
  • ICML themes: shift toward efficiency/cheaper training; RSI (recursive self-improvement) and multi-agent systems; diffusion language models.
  • Laud: model releases (e.g., Grok 4.5, Muse Spark 1.1, open-weight DLM-5.2) show real gains; evals are harder for agentic, long-horizon tasks; open-source can occupy parts of the Pareto frontier.

Written by AI. May contain mistakes. Listen to the episode to check what was said.

Chapters

Tap a time to open that second in VO

Apple vs OpenAI: The Lawsuit Explained

0:45 to 2:36

Discussion on Apple's lawsuit against OpenAI over trade secrets.

“We'll bring on our Apple reporter to unpack the latest on that.”

Key Allegations in the Lawsuit

2:36 to 6:40

A deep dive into the serious allegations made by Apple against OpenAI.

“I want to go through some of the big allegations in here, but there are two people that are at the center of this lawsuit, Tang Tan and Chang Liu.”

OpenAI's Response and Legal Context

6:40 to 8:12

Overview of OpenAI's response to the lawsuit and industry implications.

“So similarities, but certainly the allegations of theft from OpenAI's part, that's quite serious and pretty wild.”

Impact on OpenAI's Future Projects

8:12 to 10:50

Discussion on how the lawsuit may affect OpenAI's product development.

“So far, it's just been, you know, we have no interest in other companies' trade secrets, and we want to innovate for everyone.”

Timeline and Expectations for the Lawsuit

10:50 to 14:03

Expert insights on the expected timeline for the legal proceedings.

“So this talent drain has been pretty rough for Apple.”

Legal Battle Timeline

14:03 to 14:25

Discussing the duration of the ongoing legal battle involving Apple.

“Well, Aaron, I want to thank you for coming on.”

Introduction to DTCC and Blockchain

14:31 to 14:49

Introduction of Yueqi Yang and transition to discussing blockchain trading.

“This week, a major stock market infrastructure company is set to demonstrate stock trading on blockchain.”

DTCC's Role in Stock Trading

14:54 to 16:43

Exploring the functions and upcoming blockchain initiatives of DTCC.

“Not a name that I thought we would be talking about on the show.”

Comparison with Crypto Companies

16:47 to 18:12

Differentiating DTCC's tokenization approach from crypto exchanges.

“Because, I mean, crypto companies have been talking a lot about doing this.”

Advantages and Limitations of Tokenized Stocks

18:14 to 20:44

Analyzing the benefits and limitations of tokenized stock trading.

“initiative, the first question to ask is really what are the benefits and what are something that blockchain enables that they can pursue in the traditional setup?”
Show all 24 chapters

Challenges of Blockchain Clearing

20:46 to 22:24

Discussing the impracticalities of using blockchain for stock clearing.

“That's one of the most common misconceptions.”

Future of Tokenized Stocks

22:26 to 23:26

Perspectives on the long-term outlook for tokenized stocks and DTCC's efforts.

“You actually, I mean, big picture here, is this moving faster or slower than you thought?”

Circle's National Trust Charter

23:36 to 24:56

The significance of Circle receiving a national trust charter and its implications.

“Circle got the green light for its bank charter.”

Transition to AI Conference Insights

24:56 to 25:55

Stephanie Palazzolo shares insights from her trip to the AI conference in South Korea.

“I think Stripes Bridge is also another important one to watch.”

Key Takeaways from ICML Conference

25:56 to 28:00

Discussing the key takeaways and shifts in focus observed at the ICML conference.

“I am definitely jet lagged, but it was worth it.”

The Future of AI Research Jobs

28:00 to 30:00

Discussion on the anxieties and optimism surrounding AI job displacement.

“You know, am I going to get automated out of a job?”

Continual Learning and AI Agents

30:00 to 32:50

Exploration of the role of continual learning in AI and the rise of multi-agent systems.

“So let's go back to RSI and then continual learning broadly.”

Diffusion Language Models Explained

32:50 to 35:00

Introduction to diffusion language models and their potential impacts on AI outputs.

“You also wrote about diffusion language models being at the center of some of the conversations.”

Introducing Brayden Hancock

35:00 to 35:35

Introduction of Brayden Hancock and his background in AI research.

“Like, you know, just give it to me in a quarter of the length.”

Evaluating Recent AI Model Releases

35:35 to 39:10

Analysis of the latest AI model releases and their implications.

“He is a research partner at Laud and the former director of AI at Meta to walk us through this moment for the technology.”

Challenges in AI Evaluation

39:10 to 42:00

Exploring the complexities of evaluating AI systems and their effectiveness.

“that these models do not appear to have been simply bench-maxed, so to speak.”

Challenges in Evaluating AI Agents

42:00 to 44:31

Explore the complexities and evolving methods for evaluating AI agents.

“You know, with agents, I mean, my understanding of the challenges, well, hey, I mean, if it's something like a customer service agent, maybe you can sort of measure, okay, the task was done.”

The Future of Open Source in AI

44:31 to 45:53

Discuss the potential impact and viability of open-source AI technologies.

“Let me ask you a question about open source that's been becoming more and more popular.”

Build vs. Buy in AI Solutions

45:53 to 46:39

Understand the ongoing debate between using open-source versus proprietary AI solutions.

“Yeah, I think we have precedent here in a lot of different technology areas.”
Hear the part that matters, and keep it.Open this episode in VO. Double tap your headphones to save a moment as you listen.
Get VO free

Transcript

Automatic transcript. May contain errors.

0:13Stephanie Palazzolo:Welcome everyone to the information's TITV. My name is Akash Pasricha. It is Monday, July 13th. Before we get started, this is TITV's one year anniversary week. We'll be coming to you from the Bay Area all week, filming special episodes in San Francisco and across Silicon Valley. Tomorrow is our 250th episode. You do not want to miss it. We have a jam-packed lineup. On today's show, Apple is suing OpenAI in a stunning development for the growing rivalry between these two companies. We'll bring on our Apple reporter to unpack the latest on that. We're then talking about Wall Street's biggest bet on blockchain yet with our crypto reporter.

0:55Stephanie Palazzolo:Stephanie Palazzolo, author of our AI Agenda newsletter, is back from the big ICML AI research conference in South Korea. We'll chat with her about her key takeaways from the event. And we'll wrap things up with a conversation with the Laud Institute about the big themes that they're watching in AI research and their reaction to the latest big model releases. is. It's going to be a fun show, so let's get right on into it. Apple is suing OpenAI, accusing the AI company of stealing company secrets. The lawsuit was filed Friday in the Northern District of California. I want to bring on our Apple reporter, Aaron Tilly, to break down what we know and where this is likely to go from here.

1:37Stephanie Palazzolo:Aaron, welcome back to the show. It's great to have you here. Thank you. Okay. Big lawsuit. Tell us what's in it. Let's start there yeah so apple um filed a lawsuit uh against open ai alleging that there has been a systematic effort uh on behalf of uh at open ai on the devices team to steal proprietary apple knowledge related to product development all the way to supply chain in manufacturing methods So this has been the allegations Apple's making is very serious. The companies, the OpenAI is doing things in a very sneaky way. Apple alleges that OpenAI folks are purposely hiding their efforts and trying to bypass Apple's security.

2:35So pretty serious allegations.

2:38Stephanie Palazzolo:I want to go through some of the big allegations in here, but there are two people that are at the center of this lawsuit, Tang Tan and Chang Liu. Tell us about who they are and why they are so central to this lawsuit. yeah so tang is currently the chief hardware officer at open ani he was a long time over two decades apple hardware engineer near the end he led all of apple's uh just iphone product development so very important for person at apple he left in 2024 he co-founded io products alongside former Apple design chief, Johnny Ive, and then OpenAI acquired it for$6.5 billion in stock last year.

3:29So he leads all of OpenAI's device hardware projects. And then Chang is a little more of a younger, one of the younger employees, less seasoned, but very experienced engineers from Apple that OpenAI recently brought over. He left Apple in January to join OpenAI, so it's only been there for a couple months. But yeah, Apple accuses him of bypassing Apple security and accessing Apple networks on an ongoing basis while employed OpenAI.

4:04Stephanie Palazzolo:And so the details in this lawsuit are pretty stunning. I mean, some of the details here, from what I recall. I mean, it's the idea that they asked people when they were interviewing for OpenAI, they were at Apple, they asked them to basically bring parts, you know, components, I guess, of hardware to the OpenAI office. Am I recalling this correctly? Yeah, they accused Tang of, during interviews, of requesting engineers to bring actual parts and do a sort of quote-unquote show-and-tell. And so, you know, taking parts to these interviews and then also accused Tang of asking about ongoing secret projects using project code names and asking about updates for that.

4:57So, yeah, using his already, you know, pretty substantial knowledge about Apple and kind of use it to dig more information out from current Apple employees.

5:11Stephanie Palazzolo:So Aaron, I want to just take a step back here for a minute. You've been covering Apple for a long time. Apple has one of the most extensive supply chains, if not the most extensive supply chain in technology today. Have you ever seen a lawsuit like this? I mean, a show and tell? I mean, we'll get to what OpenAI is saying about this in a second. We'll get to their response. But, you know, are these types of lawsuits, are they common in the hardware business? Have you seen anything like it? There are echoes of previous lawsuits. For example, Apple went after a chip startup in 2019 called Nuvia.

5:53It was co-founded by former Apple engineers, and the company ended up recruiting a large number of Apple chip engineers. and Apple really didn't like that. They accused the company of sort of similar behavior. Well, somewhat different, not a serious kind of feeling, but -

6:15Stephanie Palazzolo:Not a show and tell, but - Not a show and tell of chips, but would certainly, you know, this has some echoes of that. You know, I just think of Apple really doesn't like, you know, former employees going after current employees to recruit them and kind of the systematic effort that OpenAI certainly undertook is not something Apple's been happy about. And we've seen it in the past. So similarities, but certainly the allegations of theft from OpenAI's part, that's quite serious and pretty wild. And what is the evidence that Apple presents in the lawsuit for these things having happened? I mean, they say they did an investigation.

7:04Stephanie Palazzolo:Do they present any of their findings in detail? Bill? Yeah, they quote messages that they were able to retrieve, especially from Chang. They had a peers like ongoing access to messages he would send either to current employees. And it seems like they were able to do this because he retained an Apple work laptop. So perhaps, you know, they don't lay out specifically, but, you know, their security apparatus was able to access ongoing messages he was sending, even though he was a former Apple employee at that point. They also have just messages and conversations taking place between Tang and current employees.

7:53So they were probably able to retrieve that through interviews with current employees that went through the OpenAI interview process. Right.

8:01Stephanie Palazzolo:So now what is OpenAI saying about this? What's their response? They haven't said much. So I think they are caught flat-footed when it all came out Friday afternoon. So far, it's just been, you know, we have no interest in other companies' trade secrets, and we want to innovate for everyone. so pretty pretty small response there short response you put out a story after this lawsuit broke talking a little bit about how the tensions between OpenAI and Apple have been growing for quite some time leading up to the lawsuit walk us through a little bit of your reporting there because I think we all recall the flashy Apple keynote where they unveiled that part of their AI would be powered by open AI.

8:59Stephanie Palazzolo:It was a union that I think surprised some people, but it hasn't been so great since then, according to your reporting. Yeah. So 2024, Apple's Worldwide Developers Conference, they come out with their sort of AI, LLM. This is the generative AI. This is their strategy. And part of that strategy is partnership with OpenAI. Sam Altman was at the conference and it seemed like a very positive start to the relationship. But that didn't really last too long. You know, it's been two years now, and the partnership hasn't yielded much for OpenAI. They're not getting much out of it. It was a very sort of like very superficial integration where OpenAI, ChatDBT was accessed in a very limited fashion via Siri and other Apple AI features, and it really didn't work out very well.

10:00And on top of that, OpenAI has begun a very aggressive endeavor of recruiting Apple hardware engineers from this TangTang team.

10:12Stephanie Palazzolo:And OpenAI is now just starting their own devices. I mean, we don't have a clear idea as to what it is, but I mean, this hardware play is a direct competition to Apple. For sure. Yeah, so that's been going on. And OpenAI now has over 400 Apple engineers working on this device project. And that's been very difficult for Apple. There's very senior hardware folks going over there. Some of their best working on iPhone and Apple Watch and new hardware products that haven't been released. So this talent drain has been pretty rough for Apple. and one name that wasn't included in the lawsuit joanie i've was was not uh i mean he's such a central figure i was surprised what what do we know about you know the joanie i've factor here i mean i mean i guess it was it was a clean transition for for lack of better words from what we know yeah that was interesting they really kept him out of it um and i i i think that's like maintaining some sort of friendly relations with him I think is very intentional not to dip their toe, go too far.

11:32Suing the former chief design officer, the heir of the Thieves Jobs design philosophy would maybe be a step too far for Apple. So they went one rung below that to Tang and kept it a little less aggressive.

11:50Stephanie Palazzolo:So then I guess what I'm thinking about then is even though Joni Ive is not named in the lawsuit and he's kept separate from all this, to what extent do you think this lawsuit stifles OpenAI's progress, Joni Ive's own capabilities here to develop the devices that they are set to release? I mean, you know, there was a report out today that they're still aiming to get a device out in the next year or two. Do you think that this stifles their progress at all? I can't imagine how it won't stifle the progress. Yeah, they plan to release the first device February of next year. Part of this lawsuit is demanding that the company doesn't use proprietary Apple information, this device.

12:38I don't know how OpenAI can work around that. They've already been working on this for the past year, whatever plans they have in place. They can't just kind of up in that and go a completely different direction. So the device that comes out, Apple is going to be scrutinized intensely, and I'm sure they can find something to go after OpenAI about. So that could really stop the release of this product, stop the sales of this product. They could get course to intervene here. I'm sure this is going to be very tricky for OpenAI to navigate.

13:15Stephanie Palazzolo:And just to recap, as best we know, OpenAI, is it a smart speaker type device that they're working on? What's the latest on that? That's our understanding of the first OpenAI device, a smart speaker. though we understand you know a phone is is uh in the works and hopefully maybe second device third device who knows it's the intention is a family of products open a opens helps to release but um a smart speaker is the first one okay and last question for you i mean i'm sure you're talking to uh lawyers and legal experts in the background about this what is your understanding of the timeline here for this lawsuit, what we can expect to happen next.

13:59Stephanie Palazzolo:Is this a matter of months, years? I mean, how long is this going to take to play out? Yes, months, years, years definitely. All of the above. The long-running legal battle. We'll have a flip phone by then. We'll be back there. All right. Well, Aaron, I want to thank you for coming on. I trust that we will need more of your insight on this in the weeks to come. That is Aaron Tilly, our Apple reporter, here at The Information. This week, a major stock market infrastructure company is set to demonstrate stock trading on blockchain. This is a transition that many people in crypto have been endorsing for a long time.

14:42Stephanie Palazzolo:The Information's crypto reporter, Yueqi Yang, has the latest. She published a piece on it this morning. I want to bring her on to talk all about it. Yueqi, welcome back to the show. It's great to have you here. Hey, Akash. All right. So let's talk about this company here, the Depository Trust and Clearing Corporation. Not a name that I thought we would be talking about on the show. You wrote a whole story about them. Who are they? Why do we need to know them? So DTCC is the most important company that basically - I get it. It's cool. DTCC. It allows the U.S. stock market to function because it provides a few services, mostly including the clearing of the stocks.

15:26The stock market clears at the DTCC for about$20 million worth of stocks per day, so it's a huge amount. They provide a settlement of the stock, and then they also provide custody of the shares of the stocks. So they are getting into tokenization, which means they're starting to allow their members to move assets onto the blockchain. And the first step of that will happen this Wednesday when they demonstrate a few live real production trades of assets, including stocks, ETFs and other treasuries and other assets that are going to move on the blockchain.

16:06Stephanie Palazzolo:Have they picked which shares they're going to put on the block? Are these specific companies? Do we know anything about that? Yeah, so they're allowed to pursue this tokenization effort because last year, DTCC got the green light from the SEC to provide tokenization of assets, including the largest U .S. stocks that are making up the Russell 1000 index. And so that's going to be part of it. And the other assets like treasury bonds are also going to be part of the launch. Now, is DTCC, are they the first to do this? Because, I mean, crypto companies have been talking a lot about doing this. I mean, my recollection, I feel like this has happened before or am I wrong?

16:55They're approaching it with different methodology and different vantage point. So a few crypto companies, especially the crypto exchanges like Binance, Bybit, most recently Coinbase, are starting to list tokens that are backed by stocks. And then these tokens are usually issued by third-party providers rather than the issuer themselves. And then their play is more gearing towards retail distribution of these tokens. DTCC is different because they're a market infrastructure provider. They're providing a tokenization service to the members of the DTCC, which mostly includes big banks, trading firms, brokerages, etc.

17:45So this is more about institutional play.

17:48Stephanie Palazzolo:So let's talk about the advantages and then also the limitations of tokenized stock trading. I mean, one clear advantage I think we've talked about is the 24-7 nature of blockchain and the fact that that is what it could bring to stock trading. is it cheaper too? What are the benefits that we see? That's a great question. I think whenever you read about companies pursuing blockchain initiative, the first question to ask is really what are the benefits and what are something that blockchain enables that they can pursue in the traditional setup? And I think for the DTCC effort, there are a few benefits that they're trying to target.

18:33The first one is they're providing the ability for stocks that are already cleared, already settled, and are currently sitting at custody at DTCC to be tokenized so that companies can transfer these stocks in token format to each other. And why do you do that? That unlock a few benefits, including companies pledging these tokens as collateral so that they can get funding from another company and they can do this 24 7 do it real time and on the weekend as well um so like like margin margin trading is

19:13Stephanie Palazzolo:that the idea it enables enable oh i oh i guess margin trade this would be posting literal shares as collateral not just on margin to get margin to get right right okay so that's that's one benefit. I could see how that would allow some more risky behavior, call it. Margin chain always exists. Posting collateral, this process always exists, but you can do it in a different way on the blockchain, which brings some additional benefits with speed. Yeah, so there are other benefits as well, and personally I find it, the one that I find to be most interesting is that it potentially provides this bridge between the U.S.

19:59stock market and the crypto market. Because one potential outcome in the future is that for all these crypto exchanges that are providing crypto tokens backed by U.S. stocks, they can potentially use the DTCC tokenized stocks as the underlying asset backing the crypto tokens. And that could, in theory, be a superior method than the current method because it's a more direct claim to the real stock that's sitting at the center of the most important market infrastructure we have in the U.S.

20:39Stephanie Palazzolo:What about limitations to tokenized stock trading? Are there any? I think it's important to point out that DTC is not using blockchain to clear stocks. That's one of the most common misconceptions. DTCC itself said using blockchain to clear stocks is just impractical because every day it clears more than$20 trillion worth of stock. It's just there's no way to and they net out a big percentage. Most of these stocks at the end of the day, meaning that companies trade back at fours, so they don't actually need to transfer money for each of the transactions. They net it out at the end of the day and they only settle a small portion of those stocks.

21:27So this whole process of clearing and settlement is unchanged. They're not using blockchain for it. And they're really using blockchain to target a small part of this process, which is what you do with stocks that are already cleared, already settled, and just kind of sitting idly there at the DTCC.

21:47Stephanie Palazzolo:And based on what they're saying, I mean, are they sort of suggesting that they won't ever be able to clear on blockchain? Are they saying that's a much bigger lift? I mean, is this a clear limitation or is it just that they haven't gotten to it yet? I think for now, they don't have plans to use blockchain to clear trades. Right. Yeah, they gave me a statement saying that settlement models that rely entirely on growth settlement, which are common in some blockchain environments, simply do not work at this scale because there's not enough money in the system to support it. Hmm. You actually, I mean, big picture here, is this moving faster or slower than you thought?

22:33Stephanie Palazzolo:I mean, we've been talking about trading stocks on the blockchain for quite some time. There were big ambitions for it. Is this, what's your overall perspective on it and your opinion on how this is shaking out? I think this is going to be a year long, years long effort. DTCC has been working on tokenization for many years already. This is an important step. It now has the support from regulators. So I think they are making meaningful progress. But I also spoke to market structure experts for this story. One of them, Larry Tapp at Bloomberg said he doesn't expect DTCC tokenized stocks to reach even half of a percent of the value of total traditional stocks.

23:26And then he mentioned this is just a really long, complex process. And so it's going to take time.

23:36Stephanie Palazzolo:Yueqi, I want to ask you very quickly about another story that we've been following. Circle got the green light for its bank charter. This was another story that has been in the works for a couple years now. What is the significance of this bank charter? What does it allow them to do? Circle got approval for a national trust charter. This is the widely expected news because Circle and a few other companies got the conditional approval of it late last year. And the significance of this approval is that it basically allows Circle to now have a federal regulator, which is the OCC. And this is as intended by the Genius Act.

24:17It means that there's a national regulator who's now regulating Circle stablecoins. And then in terms of additional services, they can provide Circle with this charter. They can now provide institutional custody using this national trust charter and they can also manage the reserve assets backing up their stable coins and it's important to note that this is not a traditional full service bank charter it doesn't mean circle is becoming a retail facing bank they're not going to take your deposit they're not going to lend you money this is a specific type of charter called national trust charter and you actually do

24:56Stephanie Palazzolo:you expect other companies to go after this bank charter now other crypto companies yeah um it's a very popular charter for crypto firms to pursue this year, especially for the ones that are intending to issue stablecoins. So there are a few notable applicants. I think Stripes Bridge is also another important one to watch. They also applied for this National Choice Charter. Great. Well, Yuachi, I want to thank you for coming on. That is Yuachi Yang, our crypto reporter here at The Information. Stephanie Palazzolo, our AI reporter and author of The Information's AI Agenda newsletter, is fresh off a trip to South Korea where she attended the International Conference on Machine Learning.

25:42Stephanie Palazzolo:It is a big deal in the land of AI research. She wrote up some of her big takeaways for her newsletter today. I want to bring her on to talk about all of them. Stephanie, welcome back to the show. It's great to have you here. Thanks. It's great to be back. How was the trip? It was good. It was super fun. I am definitely jet lagged, but it was worth it. So ICML. Okay, this is not an acronym that people needed to pay attention to a couple years ago. It is now one of the hotspots for AI research and innovation. You wrote about some of the key takeaways this morning in your newsletter. Let's go through a couple of them.

Read the full transcript

26:21Stephanie Palazzolo:What stood out to you from the conference? Yeah, so there were a couple things that stood out. I think especially comparing this conference to the other really big AI conference that happens once a year, NeurIPS, which I attended in December. I think part of it was that NeurIPS happened right after a lot of really big, you know, big name researchers started to talk about how the field needs to be completely kind of upended. So a lot of talk at NeurIPS last December was around, you know, the fact that the models that we use today are not going to be what gets us to AGI or that we need to completely change the architectures of the models that we use.

27:00It was very much this kind of like rebellious sentiment in the AI field, I think, six months ago. And I think now at ICML, the mood is very different. I think, you know, a lot of people were more focused on practical questions like, how do we get models to run and be trained for cheaper? a lot more of a focus on efficiency and lowering costs rather than, you know, being very like, we need to completely turn everything on its head and come up with a completely new way of doing things. So definitely a lot more of these kind of practical questions. At the same time, I think a phrase that came up over and over again is RSI, which is the kind of short, you know, shorthand way of saying recursive self-improvement, which is this idea that once models become advanced enough, they can actually help to train the next generation of models.

27:54So a lot of, you know, panels I attended, you know, the number one most common question that came up from researchers was, how close do you think we are to RSI? What does this mean for researchers? You know, am I going to get automated out of a job? In fact, there was a whole talk dedicated actually to what researchers should do in the face of rsi or something that might potentially automate away their their jobs so this this is

28:18Stephanie Palazzolo:kind of an expansion i mean i mean a year ago the conversation was well how are software uh engineers going to have a job in a in a world where ai coding takes off now it's gone a step further ai researchers are the ones saying hey if the models can train themselves then you know what am i here for was it sounds like there was a lot of anxiety around that was there also people defending AI research and saying, well, there are a lot of things that AI research still cannot do or that AI automated researchers cannot do. Definitely. No, I think that was a big, you know, a big point during this talk that I mentioned was this professor from Princeton was arguing, in fact, that researchers don't need to be super anxious about AI taking away their jobs.

29:06I think this is a common argument that we've heard for, you know, whether it's software engineers or lawyers or doctors. Like a lot of people argue, hey, there's a lot of kind of like menial tasks and more tedious parts of our job that AI can, you know, can do. But a lot of the parts of our jobs that are more creative. So like as an AI researcher, sure, you know, a lot of your job today is maybe writing code or running experiments, which maybe AI can do. But maybe the more valuable part that humans can really add is the, you know, the creativity needed to actually come up with these new hypotheses in the first place or these new ideas for experiments and model architectures that, you know, today it's very hard for AI to do.

29:50So I think there was a mix of anxiety, but also maybe optimism for, you know, the future of AI researchers and their jobs.

30:01Stephanie Palazzolo:So let's go back to RSI and then continual learning broadly. I mean, you and I have talked, RSI is mostly about the models. We've also talked about continual learning in the context of agents and agent loops and the prospect of agents improving themselves. I mean, what was the conversation around agents? Did that come up? Agents did come up a lot. I did notice that there were tons of papers about this idea of multi-agents, which are agents working together. I think it's because there's this idea now that a lot of the tasks that we're beginning to give to AI, it's so complex that a single AI agent can't do it on its own.

30:44So instead, these agent systems that companies are selling are spinning up multiple agents to each tackle certain parts of a job and work together. So definitely the idea of agents was super hot, as it has been for the last couple of years, but specifically this idea of multi-agent architectures and the problems that arise when you have multiple agents working on a problem and how to kind of come up with a solution to that.

31:13Stephanie Palazzolo:But was there any debate at all in the similar sense around RSI? Was there a debate around how effective or how far continual learning will go in the context of agents? Was that a perspective that came up? It was. I think kind of the distinction that people are starting to draw when it comes to continual learning is, you know, what is something, like, what is a breakthrough that we need to have for AI to be, to have a big impact on the economy and making companies lots of money versus what is a breakthrough that we need to have to have, like, super intelligence, right? I think a lot of people were making this argument of, like, you know, even with just the AI that we have today, no continual learning.

32:04Companies like OpenAI and Anthropic are still able to make billions of dollars from this sort of AI, right? And still having a really big impact on the economy. So on one hand, it's kind of like, okay, we don't really need continual learning to have this sort of big economic impact of AI that we see today. On the other hand, I think people are still saying like, hey, the idea behind continual learning as like AI that can continue to learn even after it's trained, that's still really important. And like, theoretically, if you were to imagine the super intelligent AI, you would imagine that it would be able to do continual learning and learn even after it's trained.

32:36But that's more of like a nice to have, I guess, versus like a need to have when it comes to, you know, AI that can actually make lots of money today, which is what we're already seeing.

32:47Stephanie Palazzolo:Right. Okay. Let's go through a couple of the other themes you wrote about quickly. You also wrote about diffusion language models being at the center of some of the conversations. What are those and what was the conversation around them? Yeah, so I think in all this, you know, as I mentioned, a lot of topics were still very, very, you know, practical and focused on like efficiency and cost and things like that. I think one area where I did see a lot of chatter around maybe something a little bit more experimental and out there was this idea of diffusion, sorry, diffusion language models. models, essentially what that is, is it's taking this idea that's been applied to image models and now applying it to the world of text.

33:28So the way that today's text models work is that they output a response to you left to right. Like whenever using chat QBT, you see it, it will literally just start writing the sentence like you're reading it. And with the idea of diffusion models is that instead of kind of outputting an answer word by word as if you're reading it from left to right, it actually kind of outputs the entire answer all at once. So you kind of just imagine like this entire paragraph of text appearing on your screen, but that paragraph of text might seem kind of random or it might not make a lot of sense, things in it might look wrong, and then slowly over time the AI model will start to kind of tweak different parts of that paragraph until it ends up looking like the correct answer at the end.

34:14So it's a bit like theoretical and kind of hard to imagine, but you can you can kind of imagine it more as like outputting this giant chunk of text and then refining it versus writing the answer like a human might from from left to right so you're saying i don't have

34:29Stephanie Palazzolo:to watch claude type everything instead of watching claude type everything slowly you would instead of watch it give you this entire mumbled you know chunk of text and then slowly the text would kind of morph into something that made more more sense so um okay i don't i like Like truthfully, I don't know which one is better. Yeah, it's kind of like, all right, both of those kind of sound like weird user experiences. Yeah, if this is the future of AI research, then fine. I mean, you know, I still tell Claude every time, man, you gotta be more concise. Like, you know, just give it to me in a quarter of the length.

35:06Stephanie Palazzolo:But okay, well, Stephanie, there's a lot more that we will have to bring you back on to talk about including the safety and risk component of AI. but I encourage everyone to read your newsletter. That is Stephanie Palazzolo, author of our AI Agenda newsletter, here at The Information. The Laud Institute is quickly gaining cachet as a leading institution promoting AI research. That is becoming all the more important as model releases are coming fast and furious. I want to bring on Brayden Hancock. He is a research partner at Laud and the former director of AI at Meta to walk us through this moment for the technology.

35:43Stephanie Palazzolo:Brayden, welcome to the show. It's great to have you here. Thanks for having me. Were you at IECML last week as well? I was not, unfortunately. Can't hit all the conferences, but I've been following along from afar. There's a lot of them, so you can be forgiven for that. Tell us a little bit about what the Laud Institute is, and then I want to get your take on some of the model releases that we've seen coming out. Sure. Yeah, Laud Institute was co-founded by Andy Tom Wednesday, who co-founded Databricks, and then co-founded Perplexity, and then co-founded Discs. And if you notice a lot of Co's in there, it's because that's one of Andy's big strengths.

36:17He's very good at bringing together communities and making the sum be greater than the sum of parts. So Andy's had this vision, I think, for getting the researchers the right resources at the right time to do more, to have bigger breakthroughs, to see more adoption and traction of what they're working on. So with a great collection of other co-founders as well, we've been on that path for about a year as the Institute. becoming sort of the home base for the research community and a variety of different programs helping to connect them with resources, opportunity, connections, and things to have more impact with the research.

36:56Stephanie Palazzolo:Right. And so, look, you know, last couple of weeks, we've seen a staggering pace of model releases between Fable 5.6, Grok 4.5, 4.6? That's right. Can't remember the exact... 4.5. 4.5, okay. uh what was your review of these models i mean i i see the the tweets coming fast and furious there's a lot of chatter about um uh 5.6 being capable to delete entire uh hard drives on max i've seen that i've also seen um sam altman responding to fable being extraordinarily expensive which is something that we've reported grok is a bit of an outline walk me through your your reviews of all these models. Yeah, I mean, every individual model really just comes, of course, with its own narrative.

37:45And they'll, you know, list the emails and tell the story in a way that, of course, always shows them up and to the right and on the proto frontier. But I think the bigger picture here that is fascinating is just the number of different players that are now, I'd say, you know, within, you know, spitting distance of the frontier and able to sort of claim, it's not parody, at least, you know, kind of good enough and often cheap enough, cheaper than other alternatives, that they do become viable. In the past week-ish, as you mentioned, the new Grof 4.5 model, surprisingly strong. Muse Spark 1.1. I know.

38:20Stephanie Palazzolo:I literally forgot about the most recent one, which is one that you're close to because you were formerly at Meta. That's right. Yeah, I mean, a new family. I was there in the Lama days when they were still open-weight, so there's some noticeable differences there as well. But, you know, yeah, we're seeing, I mean, the DLM-5.2 model as well. and very exciting for a lot of people because it is an open-weight model that's in that class of reasonably competitive in certain spaces that matter to a lot of folks like coding. So very cool to see. And I think you see, there's what gets told on release day, and then there's the court of public opinion that follows that I think is the more relevant one.

38:58The other evals that are used that are sort of community-run and independently validated that are helpful for getting the full picture of what's real and what's narrative spin. And we are seeing, I think, that these models do not appear to have been simply bench-maxed, so to speak. There are some real gains here. One blog post that I liked quite a bit came from Matei Zaharia at Databricks, where they showed, we evaluated a number of these models with different harnesses on internal Databricks data and found that multiple different model families were on the Pareto frontier for their tasks. And that's not something that these companies could have hill-climbed on during development.

39:36And so that's a very interesting validation when it comes to seeing the true sort of, you know, eval story here.

39:42Stephanie Palazzolo:But let's go back to Muse. I mean, you were a former director of AI at Meta. So this is a team that you know better than the other teams, at least. You saw 1.1 come out. Were you better? Is it better than what you thought? Were you expecting more? What do you think? Yeah. Yeah. So a year is an eternity in the AI space, and a year is about how long I've been at LOD. So a lot of things change, a lot of reorgs at Meta. But I think they've got a lot of the critical ingredients there, in terms of what do you need to have a competitive model today? You're going to need compute. You're going to need people with the know-how to build an efficient model.

40:23And beyond that, there's a lot of scale, and then just everyone iterating as fast as they can. I mean, similar to the Satya essay, it's really ultimately about iteration speed, how quickly you can get a flywheel of a model and evals and feedback and kind of systematically revving up. So I'm glad to see another viable player in this space, assuming that they continue this trajectory. I think that really benefits the research community to have more options that are on that frontier.

40:53Stephanie Palazzolo:Now, you mentioned evals. I wonder if you can talk about whether you think evals are getting harder or easier, given where research is at right now? Yeah. Evaluation science has never been more critical. Earlier on, when models were so clearly well behind what humans could do, it did not take nearly as much effort to design a good eval. You could have relatively straightforward problems, relatively precise specs, and measure things very easily. As they become more and more capable, longer and longer time horizon, more and more agentic, not just giving an answer, but actually doing work that you want to then check and confirm that the computer state is in the right space after it finishes rather than just to return the right answer, it's definitely gotten a lot more complex to create good evals.

41:40And so you see more companies, I think, appreciating that, hiring for that, investing in that.

41:46Stephanie Palazzolo:And what is the, I mean, make clear for us, what is the challenge here? One challenge we've written about the information is the idea that the model may actually pick up on the fact that it's being evaluated in that moment. So that's one challenge we've talked about. You know, with agents, I mean, my understanding of the challenges, well, hey, I mean, if it's something like a customer service agent, maybe you can sort of measure, okay, the task was done. But it feels like there's a little bit of a gray area in terms of like, well, what is effectiveness in terms of agents doing their job? So give us some more of the flavors of some of the challenges with evals that are coming up right now.

42:27Sure. I mean, some fascinating ones that I'm going to reach kind of like the meta level are things like what you mentioned. Awareness that they're being evaluated and maybe change their behavior. Or a fairly common one is you used to be able to package up evals as a small set of files that all sit together. and now that these agents are able to move around and use the computer that they're on, they can go and look up the answer, not just on the internet, but sometimes in the local files around them as it's being tested. So you're seeing increasing need for just kind of generally good eval hygiene in terms of where you isolate information and what you make available, but then also just the design.

43:06I mean, if you look at coding evals as an example, we began with relatively simple, local, complete this file in Python, And now today, we're looking at things like Senior Sweetbench, a new eval that just came out that I really like that acknowledges the fact that often once a senior software developer given a relatively unspecified problem that are underspecified, that they need to go and like flesh out and sort of debug and design a solution for and like create a spec for almost rather than just being given very well scoped, you know, little bundles of work. And so that's part of, I think, where the complexity comes from is you design larger and larger scale, longer and longer time horizon tasks.

43:47That's a more complete world you need to create for this model to operate in, not just a text prompt. And then how do you evaluate it when it is able to operate in a larger space and sometimes diverge? How can you confirm whether it's done the task well, done it tastefully, done it with relevant intermediate steps so that your reward signal is relatively dense rather than super sparse? I worked three hours and I just got one bit of information, right or wrong. That's a lot less helpful than something that comes with richer feedback along the way of, did it complete these subtasks? Did it pass all unit tests?

44:24Did it design code that's compatible with the rest of the code base? things like that that were not considerations on anyone's mind a year ago.

44:31Stephanie Palazzolo:Right. Let me ask you a question about open source that's been becoming more and more popular. Is open source the future for AI? I sure hope so, for the good of all of us. It is so healthy for a research community, for a country, for humanity to have more players participating in a technology that's critical. So there have certainly been scary moments here where the gap seems to be increasing or the public will to support open source or open research with the same level of resources as closed has been in question at times. And so you'll see a lot of institute and a number of different players very actively trying to be a part of the solution here and see what needs to happen for the open research community to be competitive and sufficiently well-staffed and sufficiently well-coordinated to move as quickly as some of these well-funded but very vertical, very interconnected single companies that are making progress here.

45:37Stephanie Palazzolo:So if open source is the future, then what do you think becomes of the businesses of these giant AI labs, given that the frontier model are all closed source right now? So where do they fit in then? Yeah, I think we have precedent here in a lot of different technology areas. I think open source, with most technologies, becomes a viable option because it occupies a part of the Pareto frontier. Once a task or a problem where a workload becomes large enough in volume, there will always be places that want the outsourced version, the batteries included, quick plug and play. I'm happy to pay a premium for you to just bring me what I need.

46:18But there will also be researchers who need more visibility, companies with very large spend, or just very close to the core of their business that need more control. And they'll probably opt for more of a build-it-yourself solution on top of open source. So this is really just another iteration of the Build vs. Buy debate that had existed in sort of every industry and every technological revolution that comes along.

46:41Stephanie Palazzolo:Great. Well, Brayden, I want to thank you for coming on. That is Brayden Hancock, research partner at the Laud Institute here on TITV. That does it for today's show. Tune in tomorrow for a special one-year anniversary of edition of TITV. I am very excited for that one. 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 Instagram, on TikTok, and on LinkedIn. I am already excited for our next show tomorrow.

47:16Stephanie Palazzolo:Have a great rest of your Monday. I'll see you tomorrow. Bye-bye for now.

From the publisher

Apple reporter Aaron Tilley talks with TITV Host Akash Pasricha about Apple suing OpenAI over hardware secrets. We also talk with Yueqi Yang about Wall Street's DTCC tokenizing stock trading, Stephanie Palazzolo about her key takeaways from the ICML AI research conference, and we get into the latest frontier model releases with Laude Research Partner and former Meta Director of AI Braden Hancock. 


Articles discussed on this episode: 

https://www.theinformation.com/articles/apple-sues-openai-trade-secret-theft

https://www.theinformation.com/articles/wall-streets-biggest-move-blockchain-begins-limits


Subscribe: 


Sign up for the AI Agenda newsletter: https://www.theinformation.com/features/ai-agenda


TITV airs weekdays on YouTube, X and LinkedIn at 10AM PT / 1PM ET. Or check us out wherever you get your podcasts.


Follow us:

X: https://x.com/theinformation

IG: https://www.instagram.com/theinformation/

TikTok: https://www.tiktok.com/@titv.theinformation

LinkedIn: https://www.linkedin.com/company/theinformation/


Chapters:

00:00 - Introduction

01:13 - Apple Sues OpenAI Over Trade Secret Theft

15:31 - Wall Street’s Blockchain Bet: Stock Tokenization

26:31 - ICML AI Research Takeaways & Recursive Self-Improvement

36:38 - Lod Institute on the New Frontier Model Releases


More from The Information's TITV

All 304 episodes
Apple Suing OpenAI for Stealing Trade Secrets, Former Meta Director of AI on Muse SparkThe Information's TITV · 47 min
Listen in VO