NO. 528 PREMIUM EDITION SUNDAY 18 FEB 2024

In this issue: OpenAI does video, Google catches up a bit, Meta proposes a whole new thing, and Air Canadian wishes AI would go away.

Read this issue as a web page. LINK

News

OpenAI does Video

Amazing demo of the week: OpenAI has a new short-form video generator that blew everyone’s minds. It appears to be based on some existing open-source papers, but scaled up, and looks about as convincing as a high-end game engine. It struggles with text and fingers, as ever, and physics, and the lack of determinism makes it a little hard to use (as with Midjourney, you don’t exactly control what you get) but this makes it clear that we should presume video generation will make the same leaps as images and text, and that we should think about what that changes. LINK

Google catching up

Meanwhile Google showed off the next iteration of its Gemini model set. Again, cool demos, and this is multi-modal, so it can manage text, images and audio in the same session, plus input of up to 1m tokens (eg an hour of video or 70k words). There’s lots of arguing about how far ahead OpenAI still is, but it’s pretty clear that it’s not going to be the only company with ‘best’ models. LINK

Meta proposes a new AI approach

One of the basic science arguments in AI is whether LLMs will go all the way to ‘human-level’ intelligence (whatever that means) or whether a different breakthrough (or indeed an unknown number of further breakthroughs) is needed. Yann LeCun at Meta proposes a new architecture (‘V-JEPA’) which, very crudely, comes down to getting AI models to watch a huge amount of video and infer cause and effect and physics from seeing what happens - in effect, what infant mammals do. This is a return to supervised learning, but learning by watching video of the actual real world (billions of hours of it, perhaps, some of them copyrighted) rather than just playing Go or Pacman, as pioneered by DeepMind. There are old AI arguments about whether human-level AI (AKA ‘AGI’) is possible without physical embodiment and experience of the world, and perhaps watching video at massive scale is a proxy for that just as reading text written by people at massive scale has been on some axes for ChatGPT? LINK

AI and marketing

The two big 2023 use-cases for generative ML seem to have been coding and marketing. WPP and Publicis, two of the world’s largest advertising companies, both held investor days recently, and most of their time was devoted to AI. WPP in particular demoed its internal AI workspace, that gives employees managed access to a range of generative AI models, across both text and imaging. It also talked about training models on particular brands’ tone or voice, or target demographics. So far, it has almost 30k users, and millions of prompts. Meanwhile, Coca-Cola talked about using Generative ML to create thousands of marketing assets automatically (which must make WPP and Publicis... uncomfortable). WPP, PUBLICIS, COCA-COLA

The week in AI

In the USA, the FTC is banning AI impersonation, as seen recently in robocalls impersonating President Biden. LINK

An Air Canada LLM chatbot hallucinated a refund policy that didn’t exist, and the airline tried to deny responsibility. Nope. LINK

Every text box on the internet get an LLM this year, and every incumbent tries to make the new thing a feature. This week, Slack, with auto-suggest and auto-summarise. LINK

And not content with taking everyone’s money, Nvidia is also working a way to run LLMs locally on your own PC, pointing it at local data sets for analysis. LINK

Walmart buying Vizio?

The WSJ reports that Walmart is in talks to buy Vizio, the smart TV maker. Apparently Vizio is a best-seller in Walmart stores, but it’s probably more important as a connected advertising player: over 20% of US TV ad spending is now ‘smart’ and Walmart wants to build up its advertising business (where Amazon did $40bn last year). LINK

Remember Wish?

Wish was an early attempt at aggregating Chinese vendors to US consumers, and it was briefly valued at $14bn, but it was sold this week for $173m. Notes to Temu. LINK

Chip wars

ASML’s CEO pointed out that the EU’s plan to build world-class chip manufacturing is way behind schedule. Yes, this is ‘man who sells semis equipment says that people should buy more semis equipment’ but that doesn’t mean he’s wrong. LINK

TikTok wars

TikTok is banned on US government devices and plenty of people thinks it should be banned entirely, but meanwhile President Biden has an account now, because, well, that’s where the audience is. LINK

Apple blocks web apps in the EU

Apple will disable web apps in the EU as part of its compliance with the DMA. (A web app = adding a link to a website to your home screen so that it runs and look like an app.) The DMA requires Apple to allow alternative web browser engines: Apple says that will break the current model it uses to sandbox web apps from each other to protect user privacy and security, and it isn’t going to built a new one (at least, not by the EU’s deadline), partly because (it says) no one uses them.

This has ‘open web’ advocates very upset, but, well, welcome to trade-offs and unintended consequences, the drawbacks to lawyers trying to do OS design, and the consequences of lawyers thinking that you can resolve trade-offs by just adding a clause that says ‘and do it with no trade-offs!’ Hence, the EU told Apple: "remove lots of privacy & security controls - and keep the privacy & security!" Nothing Apple builds in response to that could make everyone happy.

More generally, I think the idea that Apple is doing ‘malicious compliance’ in the EU confuses three rather different things: what you might want from Apple, what the EU wants, and what the law actually says. In particular, you might think Apple’s privacy & security arguments are worthless, but the EU wants platforms to manage privacy & security. It invented cookie boxes, and thinks they work, and it wants to ban E2EE. This is not an apostle of ‘open’ computing. LINK

Apple's first EU fine!

According to the FT, the EU is finally getting around to fining Apple over Spotify’s complaint about in-app payment rules - €500m, it seems. I think Spotify is in the right, but I also think it’s the exception that proves the rule, with very few other real problems arising from Apple’s payment rules (ebooks are the other clear one), and meanwhile, as I wrote a few weeks ago, the EU has tried to address this structurally by redesigning the app store model itself, which has, inevitable, produced an unsatisfactory mess (as above, lawyers and economists should not try to design operating systems. Watch this space, but also, who cares? LINK

About

What matters in tech? What’s going on, what might it mean, and what will happen next?

I’ve spent 20 years analysing mobile, media and technology, and worked in equity research, strategy, consulting and venture capital. I’m now an independent analyst. Mostly, that means trying to work out what questions to ask.

FIND OUT MORE

Ideas

A survey of the state of LLMs. LINK

Stop trying to demo chatbots with ‘book a holiday for me’ - this is not a realistic use case. LINK

Tech regulators love ‘choice screens’, but they don’t work. LINK

Cyber in the Israel/Hamas conflict. LINK

A good write-up of the various efforts to give (very limited) satellite connectivity to mobile phones. LINK

Meanwhile, Ukraine alleges that SpaceX is letting Russia use Starlink. LINK

Mark Zuckerberg did a video review of the Apple Vision Pro (versus Meta’s Quest 3). Some of what he says is debatable (or incorrect) but notable that he’s going direct. LINK

Also interesting: Apple did a marketing post promoting the games on the Vision Pro, and not a single one of them is VR. Indeed, there really aren’t any actual VR experiences at launch, which is obviously deliberate. LINK

We went back to the office, but not every day. Office real-estate values are falling hard. LINK

Brian Weiser on the state of digital advertising measurement. LINK

Adam Wiggins on the (relative) failure of Muse, an ‘infinite canvas’ pen-based productivity app. New forms of work are hard. LINK

Outside interests

Living in the legendary Paul Rudolph penthouse apartment. 27 levels and no doors. LINK

There are so many e-commerce deliveries in Manhattan now that the city wants to try diverting them from trucks to barges and cargo bikes. LINK

Everything you want to know about manufacturing 155mm shells, as needed in Ukraine. LINK

And the growth of the Chinese navy. LINK

Data

The EU confirmed that iMessage and Bing have too few users in the EU to qualify as ‘gatekeepers’ there. LINK

Years after US sanctions kneecapped it, Huawei is back to the top spot in China smartphone sales (unclear how sustainable that is, though). LINK

Column

Abstracting Ai

The last wave of machine earning started working about ten years go. Suddenly, we had amazing demos. Almost overnight, broad categories of unsolved problems just… worked.

But what were we supposed to do with this? Machine leaning means image recognition now! Looks, here it is recognising a cat! But what do you do with that if your company doesn’t work with images? It solves speech and language too! But what if your company doesn’t work with speech?

Indeed, even if you did work with images or speech, it wasn’t obvious what to do with this. I had a meeting with a large sports broadcaster in about 2015, and they said that they understood that ML would let them index and analyse all of their recordings over the last decade or two. The problem was, they weren’t sure how that was useful. What would they look for? (There was a long pause in the room.)

It took us several years to works out how to understand, not what, technically, ML was doing, but how to think about it. Over time, we realised that this was not images or pictures but pattern recognition. And then we had to work out what problems might be pattern recognition problems, or what we could turn into pattern recognition problems, and indeed to find things that we’d never realised were problems before. We’ve had close to a decade of company-building around that, and there’s probably another decade or two more to come.

We’re now back at that point again with generative ML. It’s harder to understand what a Large Language Model is really doing, but it’s also harder to understand how to think about that. What’s the right level of abstraction? It’s technically correct but useless to say that this predicts the next word, but on the other hand this is not anything close to human intelligence. Is this a reasoning engine? Perhaps, eventually, but not yet (though do we agree on that?). Calling it a summarisation or synthesis machine seems closer, but also incomplete. And what would that mean, even so?

Talking about AI often becomes a hunt for analogies, and I always used to compare it to interns. The last wave of ML gave you infinite interns, who could read things or listen to calls for you (“listen to every call and tell me if the customer sounds angry”) - this wave lets those interns make things for you, again with the caveat that, like an intern’s work, you have to check every line. But again, what would that mean for your company? I have not yet worked out anything in my workflows where ChatGPT is useful, but then, I’ve never had an intern either.

As I wrote above, the two big Year One fields for generative AI adoption were coding and marketing - two fields where it was easy to see ‘out of the box ‘what you could do. The rest will take time, and companies. Companies invent problems, or discover them, and then point them out to the customer. As I wrote last year, I think that a lot of this will mean unbundling the general-purpose LLM and wrapping it in single-purpose UI, just as we unbundled Excel into hundreds of tools. I think the prompt itself will mostly be abstracted away. Of course, this is a paradox - it’s supposed to be general-purpose! And the prompt is natural language! Well yes, but SQL is general-purpose too, but at what level of abstraction? And just as it’s not the user’s job to know how the technology works, it’s not the user’s job to work out what it’s for.

Welcome

Was this forwarded to you by a friend? Sign up, and get your own copy of the news that mattered sent to your inbox every week.

SUBSCRIBE

Summarized Content