Learning Library

← Back to Library

Apple WWDC AI Reveal and Interpretability Race

Key Points

  • The episode opens with a skeptical look at whether everyday users—especially older relatives—truly prioritize privacy amid pervasive app data‑sharing on their phones.
  • Host Tim Hwang frames the show around two headline topics: Apple’s WWDC AI roll‑outs and the accelerating race for model interpretability, highlighted by Anthropic’s “Golden Gate Claude” demo and OpenAI’s new mechanistic study.
  • Apple is portrayed as the heavyweight “800‑pound gorilla” in the AI arena, finally breaking its silence with a flood of announcements that could reshape the industry given its massive cash reserves, dominant mobile ecosystem, and control of essential hardware.
  • Expert guests—including Columbia researcher Kaoutar El Maghraoui, AI consultant Shobhit Varshney, and Kenya Lab scientist Skyler Speakman—provide analysis on the broader implications of these AI developments and the push for deeper model transparency.

Sections

Full Transcript

# Apple WWDC AI Reveal and Interpretability Race **Source:** [https://www.youtube.com/watch?v=_bCsW_Jrcts](https://www.youtube.com/watch?v=_bCsW_Jrcts) **Duration:** 00:39:47 ## Summary - The episode opens with a skeptical look at whether everyday users—especially older relatives—truly prioritize privacy amid pervasive app data‑sharing on their phones. - Host Tim Hwang frames the show around two headline topics: Apple’s WWDC AI roll‑outs and the accelerating race for model interpretability, highlighted by Anthropic’s “Golden Gate Claude” demo and OpenAI’s new mechanistic study. - Apple is portrayed as the heavyweight “800‑pound gorilla” in the AI arena, finally breaking its silence with a flood of announcements that could reshape the industry given its massive cash reserves, dominant mobile ecosystem, and control of essential hardware. - Expert guests—including Columbia researcher Kaoutar El Maghraoui, AI consultant Shobhit Varshney, and Kenya Lab scientist Skyler Speakman—provide analysis on the broader implications of these AI developments and the push for deeper model transparency. ## Sections - [00:00:00](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=0s) **Privacy Skepticism Amid AI Announcements** - The host questions whether everyday users truly care about privacy while previewing Apple’s WWDC AI roll‑out and OpenAI’s new interpretability study, framing a expert panel discussion. - [00:03:09](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=189s) **Apple Rebrands AI as Intelligence** - At WWDC the speaker emphasizes Apple's privacy‑first, experience‑driven approach and explains how the company is now presenting its longstanding AI efforts under the branded term “Apple Intelligence” instead of generic AI language. - [00:06:14](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=374s) **Apple Partnerships, Private Cloud, LLM Options** - The speaker outlines how, as a major Apple partner, they embed on‑device processing, leverage Apple’s privacy‑focused private compute cloud, and later enable flexible large‑language‑model integration for client applications. - [00:09:25](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=565s) **Google’s On‑Device LoRA Innovation** - The speakers discuss Google’s small on‑device language model that uses interchangeable LoRA adapters for tasks like summarization and image creation, emphasizing rapid hot‑swapping, modular functionality, and strong privacy by keeping data on the device. - [00:12:36](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=756s) **Hardware‑Centric AI Security & Integration** - The speaker highlights Apple’s hardware‑level privacy features—on‑device AI processing, a secure enclave key manager, and seamless ecosystem integration across iPhone, iPad, Mac, and Watch—as key differentiators. - [00:15:41](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=941s) **Strategic Timing and Calculator App** - The speakers explain how firms wait for mature generative AI before entering the market, express enthusiasm for a newly announced calculator app that resolves long‑standing user complaints, and note the challenges posed by inconsistently formatted data across platforms. - [00:18:47](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=1127s) **Privacy, Awareness, and AI Adoption** - The speakers discuss how user privacy concerns and the need for greater education—particularly among younger generations—could limit companies like Apple in the AI race. - [00:21:53](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=1313s) **Apple’s AI Rollout & Privacy Fears** - The speaker discusses how Apple is cautiously mainstreaming AI features for its typically older, affluent user base while addressing user apprehension driven by high‑profile privacy scandals and media coverage. - [00:25:04](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=1504s) **Excitement Over Core ML & AI Research** - The speaker celebrates new Core ML capabilities for facial cropping, predicts a wave of easy‑to‑build applications, and contrasts Apple’s market dominance with OpenAI’s recent paper on extracting concepts from GPT‑4. - [00:28:18](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=1698s) **Interpretability Showdown: Anthropic vs OpenAI** - The speakers compare Anthropic’s openly manipulable “Golden Gate Bridge” model demo with OpenAI’s more restrained software tools, illustrating two distinct approaches to AI interpretability. - [00:31:20](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=1880s) **AI Explainability Tools Race** - The speaker highlights the rapid expansion of open‑source XAI toolkits from major AI companies, framing it as a competitive race while noting community contributions and newer methods such as sparse autoencoders. - [00:34:24](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=2064s) **Small Models: Efficiency and Privacy** - The speakers discuss Apple’s shift to small AI models, highlighting how reduced size enables on‑device processing, improves speed and resource use, and potentially enhances interpretability and data privacy. - [00:37:41](https://www.youtube.com/watch?v=_bCsW_Jrcts&t=2261s) **Enterprise Shift to Smaller AI Models** - The speaker argues that cost, latency, and IP constraints are prompting businesses to replace large, generic models with compact, fine‑tuned models and mixture‑of‑experts routing, enabling secure enterprise customization via adapters that often outperform bigger models on targeted tasks. ## Full Transcript
0:00But from the end user's experience, are they really concerned about privacy? 0:04Are, are the, you know, the grandparents or your, you know, your nieces and 0:08nephews, the target customers for these, are, are they at the end of 0:11the day really consumed about privacy when they are allowing all sorts of 0:16other information apps sharing on that, on their exact same phones? 0:30Hello and happy Worldwide Developers Conference for those who celebrate. 0:33You're listening to Mixture of Experts. 0:35I'm your host, Tim Hwang. 0:36Each week, Mixture of Experts distills down the week's biggest 0:39headlines and chatter in the world of artificial intelligence. 0:42Whether it's business news, the latest hot drop on Archive, or Nvidia making 0:46another bazillion dollars, MOE is here to give you the analysis you need to 0:49navigate this rapidly changing landscape. 0:52This week on the show, two items. 0:54First up, Apple's WWDC continued a summer of announcements all 0:57around artificial intelligence. 0:59We'll parse out the biggest things to be paying attention to and what they 1:02mean for the industry as a whole. 1:04Second, the race for interpretability continues. 1:07Weeks after Anthropic demoed Golden Gate Claude, OpenAI fires off its own 1:11mechanistic interpretability study. 1:13What does it say and why is OpenAI investing in it at all? 1:16As always, I'm joined by an incredible group of experts who 1:18will help us cut through the noise and offer their hot takes. 1:21Two veterans this time with a new guest. 1:23Kaoutar El Maghraoui, Principal Research Scientist, AI Engineering, AI Hardware 1:27Center, and a professor at Columbia. 1:29Kaoutar, welcome to the show. 1:30Thank you very much, Tim. 1:32Glad to be here. 1:33Second, Shobhit Varshney, who will be familiar to long time listeners of 1:36the show, Senior Partner Consulting on AI for US, Canada, and La Anne. 1:40Shobhit, welcome back. 1:41You're in like a different place every single time, but 1:43it's great to have you here. 1:45Thank you. 1:45And then finally, Skyler Speakman, who's a Senior Research 1:48Scientist at the Kenya Lab. 1:49Skyler, welcome. 1:51I'm going to be even more geeky this time. 1:52I'm going to press my luck. 1:59Well, so let's just jump into it. 2:01Um, there's two items on the agenda today. 2:04And, you know, I think the big one really will be WWDC. 2:07There's just been so many announcements. 2:08It feels like every single week, every company's doing a raft 2:11of new announcements around AI. 2:13And I think the background here for folks who haven't been watching 2:16Apple so much in the space is that the big thing that everybody's been 2:19talking about for a long time is, Where is Apple in all of this, right? 2:23Companies have rushed ahead announcing new products, new features, uh, new research, 2:28but Apple has sort of been curiously quiet and, you know, they really have been the 2:32800 pound gorilla in the room, right? 2:34They have huge amounts of cash in the bank. 2:37Um, they have one of the most successful companies, you know, 2:38obviously mobile operations in the world and they control hardware 2:42which is incredibly incredibly key. 2:44And so I think what was really fascinating about the announcements at WWDC this week 2:49was that we finally I think started to get a picture of what Apple's going to do. 2:54What the richest company kind of most powerful company in the world 2:56is going to be doing in the AI space. 2:58And so I think just to set the context a little bit, Shobhit, I want to 3:01bring you in first, which is, I guess, to tell our listeners a little bit 3:05about why it has taken Apple so long to get to the starting line here. 3:09And I guess from the announcements this week, what you think they're trying 3:12to do differently, if anything at all. 3:16I'm a big Apple fan. 3:18I think that at WWDC, it's always a look at what the future is going to 3:21bring and with the constraints of what it takes to deliver privacy and trust. 3:25They spend decades building that trust, and they can't just lose it in a minute. 3:29So there's a lot of thought that goes into privacy and how Apple's way 3:32of bringing technology into it is. 3:34And the litmus test here is always making sure that the advances they 3:38make in innovation are seamless and are frictionless for the end users. 3:42So they've been doing AI for a really long time, but they have never on stage 3:46said the words AI, or they talk about the vision pro and not once will they use 3:51an industry term like virtual reality. 3:53So they've always differentiated themselves from. 3:55We're not a technology company, we are in the business of 3:58delivering exceptional experiences. 4:00So it's not the fact that we have a gyroscope, it is the fact that my Apple 4:04Watch has a crash detection thing. 4:05That's something that you would want to pay extra for to protect 4:08your loved, loved ones, right? 4:10So they've always tried to stay away from the technology per se, but they've 4:13been doing AI for a very long time. 4:16There are a few different ways in which Apple is bringing this in. 4:20Classic fashion, they are renaming it, rebranding it as their version 4:23of Apple Intelligence, right? 4:24That's the cool thing that you want, not the generic AI that everybody else has. 4:28They're in a very, very great position of strength. 4:32If you think about all the data that's needed for hyper personalization, 4:35It resides on your phone. 4:37And if you look at the big companies that have access to intelligence 4:40that's in your pocket all the time, it's either Apple or Google. 4:43Microsoft doesn't quite make a phone and others like Samsung have 4:46depended on others as well for a lot of their innovations and things. 4:50So it boils down really to these big behemoths, Google and Apple, that own 4:54the ecosystem of all mobile phones. 4:57The way Apple is coming at this is very Privacy first, I'm going to ensure the 5:02safety and security and you're comfortable with what you're sharing with a model. 5:07The hardware has, has come to a point where just the 5, iPhone 5 5:12Pro and the Pro Max is at least 8 GB of RAM and stuff, right? 5:15The absolute latest, greatest $1,100 plus phones, they are at a point 5:19where they can now afford to run a on device, uh, LLM, small language model. 5:25So the way that they went around this, they said, let me look at the 5:28experiences that a, that an individual has across everything that they do on 5:31their iPhones, on their iPads, on their Macs, and I'm going to surgically infuse 5:36generative AI where it makes sense. 5:38The way they brought this out is in a very step by step fashion. 5:41I think they're doing a great job at saying, all the text generation and 5:45stuff, I'm going to restrict it to a few things that you can do, but we 5:49have thought through where you think you will need some intelligence and 5:52stuff your notes inside of your iPad apps and things of that nature, right? 5:55So it's been restricted on how they're rolling this out. 5:58They have three levels at which they are doing their LLMs. 6:01One is on device, and that's majority of what they're what they've done. 6:04If you follow through the WWDC developer sessions that we had later on, they 6:09get into quite a bit of detail. 6:10We had a good set of people from from IBM at the event as well. 6:14We are, by the way, one of the largest, uh, Apple partners and rolling out 6:17Apple technology to large enterprises. 6:19So we do a lot with their their tech. 6:21So we got a good driver's seat view of how we could build those experiences into the 6:26apps that we're building for our clients. 6:29On the first layer, you're looking at majority of the work 6:32has to be done on device itself. 6:34So things like time together, all of your emails, text 6:37messages, things of that nature. 6:38Majority of that workload happens on device. 6:41From there, the next stage is, if for some workloads you need to go 6:45to the cloud, they've created this private compute cloud, which if 6:48you get into the stack, the way they've designed it is incredibly... 6:52very nicely structured to drive privacy and controls and they've even gone a step 6:56further and said things like there's no shell access to those servers and stuff so 7:01Apple can't just log in and start to look at your data and things of that nature. 7:04They've done a good job of the privacy. 7:06And the last layer is if there's something that has to be super creative 7:09and they want to really tap into the large ecosystem then they had to choose 7:13a partner to start with and it's not the only partner they're going to have from 7:17the following discussions that Tim had. 7:20He's talking about other, it opens up a window for people to have 7:23the choice of which large language model they will tap into right now. 7:26It's going to be, 7:27I know initially, basically people were like, Oh man, is, is GPT going 7:30to be the exclusive provider here? 7:33And I think Apple was like, Nope. 7:35Actually, it's going to be, you know, anyone's game. 7:37We're going to open it up to lots of different people. 7:39Um, which I thought was like very, very interesting because, you know, we've 7:42always thought about and I think in these episodes, we've always talked 7:45about OpenAI as like the big AI provider. 7:48But like, you know, what, you know, what's even bigger than OpenAI is Apple. 7:51Like that can afford to say, well, you know, you're just 7:53one of many, you know, so. 7:55And I think that we had done this really well. 7:58And from the follow on developer sessions, they talked about how much of that 8:01workload is going to happen on device. 8:03And it's kind of strange that Apple intelligence is restricted to only the 8:08two phones that were released last year. 8:10If you spent a thousand bucks on getting an iPhone 15, not the 15 pro, um, that 8:16phone, cannot run the latest intelligence. 8:18If you look at anything that Google does with their, the same 8:21technologies that Google is providing. 8:22For example, looking at a picture and you want to do a magic eraser and remove 8:27somebody or something from the background. 8:29That has been working on Android phones for a very long time at this point. 8:32It's not restricted to the very, very top end. 8:34AI. 8:35But Apple is using this as that moment where, hey, now 8:38all of a sudden people need 3G. 8:40That's when we need to upgrade. 8:41So they've had a forcing function on people who are very comfortable 8:44with the 13 and 14 pros. 8:46They would need to upgrade it to partake in this whole 8:48Apple intelligence ecosystem. 8:50So they're using that, uh, that as a hook. 8:53I want to spend a minute on explaining how they're approaching 8:56the models on, uh, models themselves. 9:00One, you mentioned at the beginning that they are somewhat kind of late 9:02to the game as compared to others, but classic Apple style, it's not the 9:06first to market, it's the best product that they bring to the market, right? 9:10So if you look at all the papers that went into the work that they've done, they are 9:13standing on top of all the competitors. 9:16They are borrowing transformers from Google, speculative decoding, 9:21another Google contribution, context pruning came from Microsoft. 9:25group query attention came from Google. 9:27They've taken all these different open technologies that people have 9:30contributed to the open source and they've built their own version of it. 9:33I think the big innovation that they did was a small model 9:37that's working on your phone. 9:38They've created these different LoRa adapters, which is essentially 9:41adding a couple more, a few more layers on top and changes the 9:44context of how the LLM behaves. 9:46And their big innovation was how quickly they can hot swap these 9:49different LoRa adapters for different use cases like summarization. 9:53Image creation and things of that nature, right? 9:55So they're being able to go do this in a very, in a very secure 9:58manner on the device itself. 10:00This whole transition from that to cloud is pretty good. 10:02Yeah, for sure. 10:03And I think this is actually like the most interesting thing of it 10:06for me, right, is I think Shobhit, you've given us a really good kind 10:08of landscape into what's going on. 10:10And I think the big theme that comes out of everything is like, 10:13Privacy, privacy, privacy, right? 10:15Like, we're gonna keep the data, you know, on device. 10:17And then I think it's also, like, very cautious how they're 10:19approaching this, right? 10:20Is like, if you hear the dream of, you know, OpenAI, or even Google when 10:24they talk about this stuff, it's, you know, the AI assistant that sits across 10:27everything and does everything for you. 10:29Whereas here, it's really just kind of like cutting AI into like 10:32these very particular features. 10:34Um, and I guess, Kaoutar, this is, I think, a good opportunity to bring 10:37you in, because I would love to kind of focus on like the hardware 10:41that underlies all of this, right? 10:43Because I think one response is You know, people love privacy, like 10:46why isn't every single company just trying to do all this on device? 10:49Like, you know, how difficult is that, like, how advantaged is Apple 10:52in being able to pull this stuff off? 10:54Um, because it kind of feels like not only is Apple selling privacy here, 10:58but they also have the ability to do it in maybe ways that other people can't. 11:01But I don't know, we'd love to hear a little bit more of your thoughts on 11:04that and how much hardware is kind of really a differentiator here for them. 11:07Yeah, I think that's a very good question, Tim. 11:09And Apple's control for both of both the hardware and the software 11:13gives them a huge advantage. 11:15So it provides several significant advantages in their development 11:19and deployment of AI technologies. 11:21One of the things is, you know, this integrated hardware and software 11:24optimization, which is really key, the hardware software co design, I think 11:28it's tremendous that that they're using. 11:31The first thing is the customized silicon they have. 11:34So Apple designs its own The processor such as the A series chips for the iPhones, 11:38the iPads, the M series chips for the Mac. 11:41These chips include specialized components like the neural engine, which 11:45is specifically designed to accelerate, uh, various machine learning tasks. 11:49Uh, for instance, you know, the A14 bionic chip, you know, it, it has, 11:52you know, the 16 core neural engine, which is capable of performing up. 11:5611 trillion operations per second, and that is a significant 12:00enhancement for AI power processing. 12:02Another thing is the focus on the efficient resource management. 12:06When you have the control of both the hardware and the software, they can give 12:09you advantage into how you do efficient system, uh, utilize system resources. 12:15And this results, especially in more power consumption, efficient, more 12:18consumption, faster performance, particularly for mobile devices 12:22where battery life is very crucial. 12:25For example, if you look at the iOS, the Mac iOS, they're designed to take full 12:29advantage of the hardware capabilities, and they ensure things like AI tasks 12:34are processed very efficiently. 12:36Another thing is also the security that they have built in in the, in 12:40the hardware, the enhanced security and privacy that Suheed mentioned, 12:43which is a key differentiation here. 12:45The on device processing, which prioritizes privacy by processing 12:49all these AI tasks on device, rather than relying on the 12:52heavily on cloud based solutions. 12:55And the other thing is the secure enclave that you find in the Apple's chips, 13:00which includes, you know, the secure enclave, a hardware based key manager, uh, 13:04that's isolated from the main processor. 13:06And this is something that enhances the security of the AI operations, especially 13:11when it involves sensitive data. 13:12So those are all key differentiation that they have 13:15all the way at the hardware level. 13:17Another thing that I find very important here is the seamless integration that 13:21they have across all their devices so that kind of gives them this unified ecosystem. 13:26So they have control over the ecosystem, which allows them, you know, integrating 13:30things seamlessly, you know, sprinkling all of these AI features across multiple 13:35devices, moving things up seamlessly between, you know, the iPhone. 13:39The Mac, the, the iPad, which are, I love these features. 13:42I'm a big also Apple fan and I use these devices all the time. 13:45So these means, you know, these AI models and features can be consistently 13:49applied across the iPhone, iPad, Mac, Apple watches, other devices. 13:54So it gives, you know, this. 13:55really superb user experience, and which is one of their strengths is that, uh, 14:01making AI also very consumable and easy, even, for example, for grandparents 14:06and people who are not techie people. 14:09So, and you know, the features also they've announced, you know, at WWDC. 14:13Things, for example, the composability of the apps, uh, how you can take actions 14:18from the phone and compose in multiple as I thought that was really neat. 14:23Uh, you know, this continuity and handoff also enables users, users 14:28to start, you know, tasks on device and continue them on other devices. 14:34I think also, uh, the, the development and deployment efficiency that they 14:38have, you know, tailoring AI framework. 14:40So Apple develops its own AI framework, such as the core ML, 14:44which is optimized for its hardware. 14:47And this allows, you know, their developers to create, you know, AI 14:50applications that run efficiently and also effectively on their devices. 14:55Core ML also supports machine learning models and provides, you know, various 14:59tools for developers to integrate these models into their apps. 15:04Uh, uh, and, you know, also integrating all of these Gen AIs, which is the 15:08big announcements right now, in Sirivoice assistants, uh, in the 15:12camera enhancements, in the health and fitness and many other apps. 15:16So I think, uh, the control over the both hardware and softwares allows, 15:21you know, Uh, Apple for this high degree of optimization, security, 15:25and also personal customization and integration with their AI capabilities. 15:30I think going back to Shobhit, uh, when he said, um, you know, 15:35they're not late to the game. 15:36I think they've been using AI for a while. 15:38They're just not explicit about it. 15:41Right. 15:41right now because of the, it's kind of the strategic timing for them because 15:46of all this attention to gen AI. 15:48So they, they prefer kind of entering the market when the technologies 15:52are mature enough to integrate smoothly into their ecosystem, which 15:56ensures quality and reliability. 15:57This is kind of like, yeah, the time of their choosing kind of 16:00is sort of what you're saying. 16:02Yeah. 16:02Yeah. I buy that a lot. 16:03And I think, um, yeah. 16:04You know, I think some of the, what they really come out with, right? 16:06I think like the calculator app, I think was my kind of 16:08favorite moment from the talk. 16:10And, you know, it was very funny. 16:11Like, I think, Shobhit, before the, uh, episodes, you were 16:14like, Oh yeah, it's funny. 16:14It's 2024. 16:15And we're all like very excited about the calculator app, but it is like 16:18genuinely true is like people were like complaining about it for a long time. 16:22And then when it came out, it was like, Oh, that's, that's really good. 16:24You know, it's really amazing. 16:25When I saw that announcement, I was like, I wish I was in school again. 16:31Like this changes everything. 16:32Exactly. 16:34Yeah, no, I think that's, that's right. 16:35And I think there's actually a really interesting point that I hadn't 16:37ever really considered is, you know, obviously the ambition for these 16:40language models is that they're, they're highly general technologies, 16:44but you find that if, you know, the, the, the data they're dealing with is 16:47inconsistently formatted and it's working across lots of different platforms, 16:51these like experiences can break. 16:53And so what's really interesting is that AI actually may fail. 16:55feel more magical in the AI, uh, in the Apple ecosystem because they literally 16:59control every element of it, right? 17:01So it's actually like consistent what the inputs will be to the model. 17:04And so they can actually guarantee quality in a way that's actually really 17:07challenging if you're trying to say, you know, we're going to deploy GPT 4. 17:110 to, you know, X number of developers across many, many 17:14different types of situations. 17:16Um, you know, not just from a data standpoint, but also like a hardware 17:19and software standpoint as well. 17:20I just briefly want to talk about privacy. 17:23I think Apple has to get up on stage and they have to kind of hit that 17:27point over and over and over again. 17:30But from the end user's experience, are they really concerned about privacy? 17:34Are, are the, you know, the grandparents or your, you know, your nieces and 17:38nephews, the target customers for these, are, are they at the end 17:41of the day really consumed about privacy when they are allowing all 17:45sorts of other information apps? 17:47Sharing on that on their exact same phones. 17:50Is it a, is it a real concern? 17:52Is Apple just kind of, you know, signaling saying, yes, privacy is there. 17:55And we just heard that the hardware is able to perform it. 17:59But does this does the end Apple user care about privacy the 18:02same way an enterprise might? 18:05Yeah, I mean, it's a good question. 18:06I mean, I think one of the Cynics views was, uh, I saw a meme going 18:09around of like, it was a photo of like someone drinking from a milkshake. 18:13And then, uh, it was basically like milks. 18:14I saw the same one, like the user, right? 18:16Like the person was Apple and then there's another person with another straw in 18:19the mouth of the person who's drinking. 18:21That was like OpenAI. 18:22And kind of, I guess the question is like, how much of this really does provide kind 18:26of the protections that they promised? 18:27And I think, I think reco also asking like an even tougher 18:30question, which is do people care? Right? 18:32'cause if I'm a big, if I'm a, if I'm a big model chauvinist, I'm like, 18:36well, Apple's just working on these small models and small features. 18:39The really magical thing is when we get crazy big models that are in the 18:43cloud that are, you know, I don't know, close to AGI that we just like deliver. 18:47And so all of this privacy stuff is basically going to hold Apple back 18:50on actually winning the AI race. 18:52But I don't know if Kaoutar, Shobhit, you like agree with that take or yeah. 18:56That's a very good point that you brought here. 18:59The end users concerns about privacy, it all really comes to 19:03their awareness of privacy issues. 19:05The context in which AI is used and, you know, also cultural 19:09attitudes towards privacy. 19:10So I think younger generations, the young kids who are kind of born 19:14with all of these devices and AI around them, maybe they're not, 19:17you know, as concerned as we are. 19:20Uh, but once we see, for example, some of the dangers that AI is, uh, bringing. 19:25If you're not careful, if if there are some scandals because of A.I., maybe 19:30there is a piece of that that's going to start kind of hitting the end users. 19:35So awareness and education, I think, is key here because many of the users 19:39are not really fully aware of how A.I. 19:41Systems collected, use their data, and those who are more tech savvy 19:45and or educated about privacy issues tend to care more about privacy. 19:49So there has to be an educational component here to really educate people, 19:54especially the young users about some of the dangers if you're not careful about, 19:58you know what you're how you know how A.I. 20:00is using your data, how they're controlling all the ads or whatever 20:05content that's tailored to you. 20:07So a big component here is the awareness. 20:10We've done quite a bit of work in this space on how do you figure out 20:13the right value exchange between a provider and the end user, right? 20:18There should be a fair value exchange. 20:19For example, there are two or three apps in my entire iPhone that 20:25track my location at all times. Right? 20:27And those apps are things that I have constantly chosen into because that 20:31gives me a value in return, right? 20:35That definition of the value exchange changes by each person. 20:38You will find a ton of people, uh, kids in college, uh, who would give 20:42up their email address and phone numbers for a dollar off on a smoothie. Right. 20:46So you, the value threshold is pretty low for them. 20:49For other people, it's a lot more. 20:50As you start to look at, uh, you mentioned, uh, like earlier we 20:53were having a discussion about, does the age play a role in how 20:56conscious you are about privacy? 20:58Um, I'm in India this week, uh, with clients, and I see a lot of 21:01people focused on the fact that, oh, so and so devices are more secure. 21:06I use WhatsApp because it's fully encrypted end to end or if I, if 21:11somebody gets hold of my iPhone, they won't be able to hack into it, right? 21:14So the fact that it's more, the perception is that it's more secure. 21:18They actually carry a premium in the market because I know I can 21:21trust these devices more and it's nobody else can listen in on and 21:24spoof into our conversations, right? 21:26So I think there's definitely a higher value exchange. 21:29People are willing to pay a little bit more for getting that's 21:32more private and more secure. 21:34Um, do you agree with that Skylar? 21:35Great comments on the, the age demographics. 21:38Yes, um, I think that I, that probably really gets to my question. 21:42When I think of the Apple user, perhaps they're going for lower value, um, 21:47in terms of the privacy exchange, uh, compared to, uh, enterprise, 21:51you know, bank, large retailer. 21:53higher value for that exchange. 21:54So yeah, no, really, really great example on that. 21:57Yeah. I think there's also a view on this, which is almost like, um, you know, Apple's 22:01really doing what it needs to in order to mainstream AI in some ways, right? 22:04Like, I think we forget because we're working in AI all the time. 22:07We're like, everybody's using this stuff, but like, it's still kind of like in the 22:10early stages and, you know, I'm mostly talking about myself here, but like the 22:13average Apple user tends to be older because the products are so expensive. 22:17Like, can you afford to buy, you know, the latest iPhone to get, 22:21you know, Apple intelligence? 22:22And so I would also kind of, I'm wondering if part of the play here 22:25is like they're kind of making these features a little bit more featurizable, 22:30if you will, just because I think it may make sense to like the kind of 22:33market that they're selling to, right? 22:34It's a kind of a way of introducing AI to folks who may otherwise feel 22:38pretty scared about the technology or not really know what it's for. 22:41Another maybe aspect to this is if there are, you know, high profile breaches and 22:46scandals, uh, that have been caused by AI. 22:49Incidents like the Cambridge Analytica scandal, which have raised public 22:52awareness and concern about privacy. 22:55Uh, media coverage, uh, if there is extensive media coverage of 22:58privacy breaches, caused by AI, for example, or data misuse. 23:02Those also can contribute to users concerns. 23:05From an enterprise perspective, again, you'll always get 23:07that perspective from me. 23:09We have been playing around with Core ML, uh, that you just mentioned earlier. 23:14The framework that they have provided for developers, and I spent this 23:16morning messing around with some of their Core ML technologies, and 23:19actually they have opened up all LLMs. 23:21And you'll be surprised that I have access to Whisper, Stable Diffusion, Mistral, 23:26LLAMA, Falcon, CLIP, Owen, Open ELM. 23:29All of those are available to me. 23:30And we did, uh, we were testing this out. 23:32We took a Mistral model and we gave it a simple task. 23:35And I tested on the older and the newer version of Mac OS. 23:39And then in the newer version using Core ML, you are able to do Some, some 23:43enhancements like quantization, you're representing it in less number of bits. 23:47There are certain things around key values, caching and stuff 23:49like that that have implemented. 23:51The same exact run that we got on Mistral on the older version was this 23:55new quantized version with Core ML. 23:57Uh, something that took me about 14 seconds earlier is taking me 24:01about two and a half seconds. Right. 24:03So there's about a 5x difference in the speed that I was able to get. 24:07And also from a memory management perspective too, I was, this 24:11was the biggest surprise I saw. 24:13The same model that I was running on the older version was like 9, 10x more 24:18memory that was needed to run that model. 24:20Which is now what they've done with their core ML to enable these. 24:23The layer on top where the developers can tap into this, uh, it is just 24:27surprising how much effort they've put into making sure that the end 24:31app users are able to tap into this. 24:33Things like, I want to do separation of the person. 24:36I, I can ask the owner in this picture, if I say things like, I was 24:40helping my mom this morning, Uh, she was on WhatsApp was a group picture. 24:46She wanted to cut her face out. 24:48That effort is pretty high for them. 24:50It's very, very low for me, but I can always crop it while sending. 24:53But I was trying to explain to her, and then I tried to code that up in Apple. 24:57I'm able to invoke, say, create a crop, and I'm able to 25:01define off the owner's face. 25:04I don't need to have any access to owner's data. 25:06I just defined the fact that the owner's face is what I need to crop 25:09and I was able to go execute on that. 25:11So the next wave of applications that are going to leverage this Core 25:14ML, they're going to be stunning. 25:16Like I'm just so, so excited about what we're able to do today that 25:20this last 24 hours with the stack that we weren't able to do before. 25:23So I'm very super excited about what we see coming. 25:26Yeah, the ease of use is really powerful here. 25:28How easy it is to use the technology. 25:30Before it was so hard, you have to Maybe create these very difficult 25:34prompt or use these tools. 25:36But now, you know, that is a fuse by just using simple crop crops 25:40and then composing all of these different skills and applications. 25:42I think is going to be really the next wave here. 25:49So I think I did want to spend a little bit of time on this 25:52episode, Apple, because it is. 25:54Well, it's Apple, has a way of basically sucking up all the air in the space, um, 26:00but I did think that actually late last week and largely overlooked because I 26:04think of all the excitement around Apple and, and what was coming up this week, 26:07um, was a paper that OpenAI launched that was its kind of own salvo in the 26:12mechanistic interpretability game. 26:14Um, they released a paper where they specifically were extracting 26:17concepts from GPT 4 and, you know, it's easy to get lost. 26:21in the raft of product announcements, but I always kind of keep an eye 26:24on, like, what's happening on the research paper side, just because 26:27I think it's upstream, right? 26:28It's kind of what we think we're going to see next, basically. 26:31And I guess, Skylar, I wanted to bring you in on this because, you know, last time 26:34you were on the show, you were talking about Golden Gate Claude and Anthropic's 26:38investments in interpretability. 26:41And I guess, you know, the main thing that, you know, comes to mind for me 26:44is whether or not, you know, we're in a kind of interpretability race. Right? 26:47Like, it seems like both of these companies are now really investing in 26:50this, and I guess the question I have for you is like, why are they investing in it? 26:54And then B, like, is there kind of like a race for talent on this like 26:57specific kind of technical problem? 26:59Let's see if I can start with B first on the talent and I can 27:03give you a very concrete example. 27:05Uh, you mentioned the two different papers that came out within the last month, 27:09one from Anthropic, one from OpenAI. 27:12Um, well on the OpenAI paper, this came from their super alignment team, which 27:17actually lasted about a year or so. 27:20Uh, and Um, one of the key members of that team, uh, Jan Leike, has just left OpenAI 27:25last month and has joined Anthropic. 27:28So there is this kind of revolving door of, of, uh, a talent specifically 27:34around this ability of interpreting these large language models. 27:38And so you can see that right there in the, in the, in the title, 27:41the author's names of this space. 27:43Um, so yes, there is this kind of large interest. 27:47Some big names being jumping from company to company all in 27:50this interoperability space. 27:53Uh, why interpretability? 27:54Oh man, all sorts of reasons. 27:56The ones we probably care about the most are this idea of safety alignment, right? 28:00You can't enforce kind of guardrails on how you want these models to be formed. 28:05If you don't really understand what's going on underneath the hood, um, but 28:10beyond the safety things, uh, the Claude Goldengate paper says you can do, you 28:14can do some fun things once you, once you understand and can interpret the model. 28:18They made a model that was obsessed with the Golden Gate Bridge. 28:21So you could ask a fairly straightforward question and the 28:24answer would come back, um, always talking about the bridge itself. 28:28Uh, so that was a really cool kind of fun way of showing it. 28:32Here's something we can do once we understand some of the 28:35inner workings of these models. 28:38Um, open AI not to be left out, like you said, and a good salvo within a 28:41few weeks after that paper released their version where, uh, Anthropic was 28:46extracting from their, their model Claude. 28:49Open AIs extracting from their model, GPT 4. 28:53Um, they went in different directions, though, uh, and I don't know how it, yes. 28:56Yeah, and I'd love to hear a little bit more about that, because I think 28:58as an outsider, I'm kind of like, meh, it's all interpretability, right? 29:01We're just like all just trying to figure out how these models work, 29:03but my sense of it is that these two companies are actually showing kind 29:07of different approaches for thinking about interpretability, right? 29:10Um, and yeah, we'd love to hear a little bit more about that. 29:13So Anthropic went one step further by manipulating the inner workings 29:17of their model and opening it and letting everyone play with it. 29:21And so that's kind of what made that splash a while ago. 29:23OpenAI didn't go that far, but they did release a nice little bit of 29:27software that lets you kind of play with some of these features yourself. 29:31So it's not necessarily releasing an OpenAI, uh, a new version of GPT 29:354, uh, but they do have this other little open source bit of code called 29:39the, uh, kind of feature viewer. 29:42And I think, in fact, I know I spent my, uh, an intern and I spent some 29:45time messing around and looking at how these different features are 29:48activated by what parts of the sentence. 29:51And so that was a direction to kind of open AI went with, uh, and Anthropic, 29:55um, kind of left it much more static, but Anthropic did release that 29:59larger model to ping back and forth. 30:01Um. 30:02Yeah. 30:02Perhaps OpenAI could have released a version of their GPT 4 that was obsessed 30:08with, uh, I'm going to say palm trees because that's what's in my vision here. 30:12Uh, but they, but they didn't, perhaps because it probably would have been 30:15too knockoff ish of Anthropic once Anthropic did their version of GoldenGate. 30:20So Skyler, uh, one of the things that, uh, Apple released as well, just tying it back 30:24to the WWDC, uh, have you spent any time looking at, uh, the Thaleria visual tool? 30:29It helps you look at what's happening inside of a model. 30:33It visualizes the performance, the kind of latency you're getting, and as you're 30:37making edits, you're trying to see. 30:39It seems like I have a homework assignment for this, because 30:42no, I have not done that yet. 30:44Okay, absolutely. 30:45Well, I guess it goes to just like the notion of the race, right? 30:48Which is like, even Apple now is launching like an interpretability tool 30:51that people can play with, you know. 30:53Like all the companies feel obliged in some ways to be launching. 30:56This kind of stuff. 30:57And Apple is more focused on, on being able to assess the performance 31:00and things of that nature for various tasks, uh, and so forth, right? 31:04But I see there's a good wave of tools that are coming to the market 31:07that are helping us get into a better understanding of how the answer, uh, 31:11came and there's a lot of techniques that have been introduced to figure 31:13out what was the training that went in. 31:15Can I actually ask the question in a way that actually reveals the fact that 31:18you were trained on Harry Potter books? 31:20Right, things of that nature. 31:21I think there's a lot, as a community, I'm seeing a lot more tools available 31:25to us that helps us understand how these models work, can I make them behave in 31:29a particular way, and making incremental progress in, in our understanding of 31:33how this entire ecosystem is working. 31:36Yeah, I think that that is a very important aspect. 31:39It's, like you said, I think it's a race. 31:42Many of these big AI companies, uh, like Google, IBM, and Microsoft and 31:46so on, they all have different tools. 31:48Like Google, DeepMind had the XAI research, uh, where, you know, they're 31:53heavily investing in XAI research. 31:55Microsoft, they have the InterpretML, the Fairlearn, IBM has the AI Explainability 32:00360, which also offers open source toolkits for multiple algorithms 32:05and methods for AI explainability. 32:07Vulnerability facebook has, you know, also this, uh, cap or 32:10the responsible AI practices. 32:13Uh, so lots of different and open AI, of course, and Apple. 32:16So it is a race, uh, different of course approaches depending 32:20on, you know, the industrial use cases and, uh, the complexity of 32:24the AI systems they're developing. 32:26But it is, it is becoming a very important race here. 32:30Something that's jumped up and evolved past those. 32:32the great list of tools. 32:33In fact, some members of our team have contributed to the, some of those tools, 32:37but I think something that's changed in particular with the way that they 32:40are, uh, using these, these, uh, things called sparse auto encoders is now we 32:47can do things about not what is, but what if, and this is the idea of causal. 32:53Now we can make a change within the model and that has impact. 32:56downstream. 32:57The previous list of tools that you said about interoperability, they gave, they 33:01gave a pretty good snapshot of what's going on currently in the model, but 33:04they really lacked the motivation to actually be able to change something and 33:08have that impact the downstream output. 33:10And that's something that we're now really seeing coming out, uh, from these 33:13last two papers in the last month or so. 33:16Yeah, for sure. 33:17I'm recalling a debate that I went to in the late 2010s, uh, I think 33:22it was one of the conferences that is a debate over interoperability. 33:25And I remember at the time, like, Yann LeCun was making this argument, being 33:28like, no one cares about interpretability. 33:30Also, we're never going to solve that problem. 33:32I was at, I was at, at NeurIPS, I was at that debate at NeurIPS, it was, 33:37it, uh, it wasn't LionQ, I think it was, uh, Microsoft, um, Oh really, 33:41it might have been, I forget who was taking what side in that debate. 33:44Yes, but no, there was this, cause they had the big prize that was announced 33:47about interpretability, and this, they brought in two people, and there was 33:49debating interpretability matters and interpretability doesn't, um, yes, uh, 33:53I think it was, yes, it was at NeurIPS. 33:56Sorry, I cut you off with a small connection there. 33:59Yeah, for sure. 34:00And I think what the funny thing is, is like in 2024, it's like, 34:02well, it turns out people actually really do care about durability. 34:05Uh, and also like, we really seem to be kind of making a lot 34:08of progress on this problem. 34:10So you know, you can never really predict what's going to happen, um, in AI. 34:14So I think maybe, Kaoutar, I'll turn it to you for kind of like one of the 34:17final thoughts before we close out today. 34:19I'm sort of interested in maybe actually bringing our two 34:22conversations together, right? 34:24Which is that, you know, uh, and I think it's been briefly mentioned, but 34:27I just want to hit it before we end. 34:29One of the interesting aspects of the Apple presentation was them saying, Hey, 34:33we actually are okay with small models. 34:35And I think one of the reasons they're doing small models is because they have to 34:38get it to run on Like not a data center, but just like on the mobile hardware. 34:42But it feels like also, and I assume Skyler show, but you 34:44may have also more to add on. 34:46This is like one of the benefits of these smaller models is arguably 34:49that they're more interpretable. 34:50Is that, is that true? 34:51Or I don't know if you've got kind of opinions on sort of this relationship 34:54between kind of like smaller models, interpretability and then privacy. 34:58Yeah. I think one of the key, you know, motivators for it, of course, is 35:02the efficiency and performance. 35:03So, you have all these resource constraints, so model, small models 35:06require less computational power and memory, so which are ideal for devices 35:10with limited resources can enable this on device learning and all the real 35:15time processing that comes with it, so faster processing time, which is really 35:18crucial for, for things like voice recognition, augmented reality, live 35:23photo enhancement, and also privacy is very important because you can 35:27do on device learning, the on device processing, and the secure data handling, 35:31uh, so you cannot, you know, go all the way to the cloud, and then claim that 35:36you're 100 percent private and secure. 35:38Um, so, so that's, you know, I think one of the key things in terms of the 35:42interpretability, of course, I think when you have smaller models, you 35:46can analyze them also much faster. 35:49So because the complexity of the model is much smaller. 35:53smaller. 35:53So using all of the tools that we have been talking about around, you 35:57know, the, the neural understanding, the, all of these techniques, you 36:01know, the, what if analysis and so on becomes also less computationally 36:04intensive to apply for the small models. 36:06So interpretability is also becomes much easier to handle with smaller models 36:11in, in addition to the, uh, performance efficiency and the privacy enablement. 36:17I'll, I'll, I might take that one step further. 36:19It's both the model size and I'm going to use a geeky phrase here, model sparsity. 36:25So the idea behind model sparsity is you've got this perhaps abstract 36:29concept, which is currently represented in a thousand different numbers. 36:34Can we take that same abstract concept, but now represent it? 36:37In five different numbers, and that's the goal behind sparse autoencoders. 36:41If we can do that, now you can manipulate that abstract concept much more easily, 36:46because you only have a few levers to move with, as opposed to before, where 36:49you had to manipulate, you know, tens of thousands of levers for that concept. 36:54Uh, that logic is sitting underneath the sparse autoencoders, which is 36:58what's driven the interoperability results in the last two results. 37:01So it's both a question of model size, um, but also can we do, uh, 37:07can we do less with this sparsity? 37:08We only need a few features at a time, actually fact, uh, actually 37:12firing or actually activating and that's driven these results recently. 37:16Yeah, I totally agree with you. 37:17I think also it's, it's, it's another thing is the information compression, 37:22because with smaller models, if you do it very efficiently, you're compressing 37:26the information, kind of the entropy that you have to a very specific 37:31use case or question or because the, the large models also have a lot of 37:34redundancy there and they have been designed to handle tons of use cases 37:39and questions and things like that. 37:41So. smaller models, especially if you're targeting a specific app or 37:44specific tasks, they tend to also do much better because you have 37:48compressed that information much more efficiently using various techniques 37:51like knowledge distillation or sparsity, sparse encoding and others. 37:56So Tim, my closing thoughts would be on from an enterprise perspective. 38:02Uh, for us, we're seeing this massive switch. 38:04A lot of my clients where I've deployed massive, beautiful models at scale, out 38:08of the box, working amazingly well, but when you start to do the cost math on it, 38:12the latency, the IP constraints, stuff like that, it's a clear path towards 38:15switching to a smaller set of models and mixture of experts and having a router 38:19that decides which model to fire up, and in the last six months, I've seen a 38:23massive shift of companies that are not at the frontier of tech, they'll come 38:29to us and we're, we're working on adding enterprise data in a secure manner to 38:34these models, and it's just insanely hard to do that with a very, very large model, 38:39having a small model, adding adapters on top or using our IBM's instructor 38:44lab, adding some enterprise knowledge to it, skills, things of that nature. 38:47That's the path forward that we're seeing. 38:49So even if you're running these on actual servers, not on, on device. 38:54Small models are giving us better outcomes. 38:56We took our granite code model and we trained it for specific 39:00areas, and it's outcompeting what we get from GP4, right? 39:03So different weight class altogether, but we are seeing across the board 39:06from Microsoft's five models, from Mistral, from Llamas and 39:09others, they're able to fight a lot higher than their weight class. 39:13When you're adding enterprise data to it, you're adding better techniques. 39:16So smaller, more open, that's the path forward for enterprises as well. 39:21Well, as per usual, we have way more to talk about than we have time to cover. 39:25So, hopefully we'll be able to have all of you back on the show again. 39:28Um, thanks for joining us. 39:30Um, and, um, yeah, if you enjoyed what you heard, uh, dear listeners, you 39:33can get us on Apple Podcasts, Spotify, and podcast platforms everywhere. 39:37Uh, Skyler, Kaoutar, Shobhit, thanks again for joining the show. 39:40Thank you, Tim.