Meta is still metaversing Meta had a big event around VR and AR, announcing an updated Quest headset and a prototype of AR glasses, called 'Orion'. People who were at the event say the glasses are impressive, with a 70º field of view (FoV) and the compute in a wireless puck. 70º is a lot better than the 46º for Snap's dev kit announced last week (which doesn't have external compute), and the same as the Magic Leap 2 a few years ago (which did) but less bulky. However, this is still bulky and heavy by consumer standards, the resolution is very low, the battery lasts 2-3 hours… and the Orion isn't for sale - Meta says the production cost is around $10,000, which would require a retail price of $15-20k. I haven't tried Orion, but I had a Magic Leap demo in Fort Lauderdale in September 2014, and compared the 'wow factor' to the iPhone - these devices feel like they must be part of the future, eventually. But here we are in 2024, with Meta showing us something still many years from being ready to sell. Apple is at least selling the Vision Pro, but it's effectively a dev kit that's also years for being ready for the mass market (which could mean Apple doesn't think it can get glasses to work). I wouldn't be surprised if Apple has something very similar to Orion in the lab, also for $10k and also years from being ready ready. Meta has now spent well over $50bn on VR and AR research, partly because it thinks this will be the next universal device after the smartphones and partly because if it is, Zuck doesn't want someone else (i.e. Apple) to control it. The same goes for its investments in LLMs. But though the tech has got better since 2014, the questions haven't changed: how long until we have something normal people can use, and when we do, will they care? LINK, ANALYSIS The week in OpenAI OpenAI seems to have enough drama in a week to last an ordinarily respectable startup all its life. This week it rolled out its 'advanced voice' model (that one that supposedly sounded like Scarlett Johansson). Amusingly, its ability to sound enthusiastic or sad would make it technically illegal to use at work in the EU, since the AI Act (effective for next year) bans any AI product that can infer emotions. But meanwhile.. LINK The CTO, Mira Murati, announced that she's leaving, along with both the Head and the VP of research. The board is considering ways to restructure the company into something more normal and rational, which might be a precondition of raising much more money. There was also a rumour that Sam Altman would be given 7% of the equity (he currently has zero), which he denied. DEPARTURES, ALTMAN EQUITY The WSJ points out that Murati is one of more than 20 senior execs, researchers and co-founders who have quit this year. A lot of people don't seem to like working for Sam Altman. Alternatively, lots of other people can offer them real stock options (not 'Profit Participation Units') at real companies (see the Noam Shazeer story below). LINK The NY Times has OpenAI financials, prepared as part of its current fund-raising. $300m monthly revenue in August and on track for $3.7bn in 2024, but $5bn of losses. It has 350m monthly active users (up from 100m in March) and 10m paying users. LINK Seperately, the NY Times also reports that TSMC thought Sam Altman's "invest $7tr in AI chips" plan, reported earlier this year but never officially announced, was so ludicrous that they started calling him 'podcasting bro'. LINK With all of this going on, it may or may not be a coincidence that the WSJ reports Apple is no longer looking at joining OpenAI's investment round or take a board observer seat. LINK And finally, Sam Altman published an essay about the future of super-intelligence, which reads like a cross between a TED Talk and a sermon about The Rapture. I wonder who the intended audience is? LINK California's governor vetoed its Doomer's Charter California's governor vetoed an AI Safety bill, 'SB1047', which was pretty much a doomer's wish list, creating all sorts of onerous and impractical restrictions on AI development, including making anyone who makes an open source model liable for anything bad that anyone else did with it. This was by design: the people who originated the bill would be happiest if LLMs just didn't exist. There are broader stories about what kind of regulation of AI is a good idea (see below), and also about the failure of doomers to persuade many people outside their bubble, but today most of Silicon Valley will be relieved that something dumb and destructive didn't happen. LINK Meanwhile, California passed a range of single-issue AI bills on things like deep fakes and using the voices of dead actors. In principle, I see this as the right approach. Trying to write one big law that controls anything anyone could do with AI is the wrong level of abstraction, just as it would be to regulate 'databases' or 'spreadsheets'. LINK The week in AI The WSJ confirmed that Google did a $2.7bn acquihire to bring back Noam Shazeer, a co-author of the transformers paper that kicked off the LLM boom, by 'licensing' tech from his new startup Character.ai. We seem to be repeating the war for AI talent of a decade ago, but at 100x the prices - he presumably made multiple 9 figures from this (?). LINK Reddit will deploy AI-powered translation in dozens of new countries, 'flattening the world' as the saying goes. It was quite recent that machine translation was not very good and terrible outside a few target languages that got all of the investment - now it mostly 'just works', thanks to machine learning, and we are quickly taking it for granted. LINK Microsoft is doing a deal to re-open a nuclear power plant to power a data centre. The power station in question is Three Mile Island, known for a famous though pretty minor accident in the 1970s. I am still getting used to hearing AI people talk about data centres in terms of gigawatts. Note, though, that while the models and the scale of their use are growing fast, the efficiency is increasing hugely - something between 75% and 90% efficiency gains in the last year or so, as the engineers took over from the scientists. LINK Alibaba launched a new version of Qwen, its open source (sort of) family of LLMs, and it's at the top of the leaderboard for some benchmarks. LINK US versus Chinese cars… Apparently, the US is considering a ban on Chinese software and hardware in connected and autonomous vehicles. The new Chinese EV companies have better connected software than most traditional OEMs, and are doing pretty well in autonomy (and certainly driver-assist), so I can see the national security implications - you don't want China to be able to activate the microphone in selected vehicles (or, more dramatically, tell all the cars to stop, or all drive somewhere and crash into a target). But the electronics supply chain is so complex and interlocked that there will be be a lot of nuance to what exactly is banned. LINK Huawei's triple-fold phone Huawei was pushed out of the handset market by US sanctions a few years ago, but it wants to show that it's back, at least in China, with its own chips and its own OS (mostly an Android fork), and part of that means launching a flagship triple-folding phone, which it announced the day after Apple's annual iPhone event a few days ago. At a price of almost $3000, this is more a trophy (for both the customers and Huawei itself), and a statement of intent, that a competitive product. It weighs 50% more than an iPhone Pro, but has 3x the screen space, which is potentially interesting, but the folding screen still has the characteristic creases. We are still waiting for the 'right' folding screen device, but we are getting closer. LINK | 
No comments:
Post a Comment