The week in OpenAI OpenAI showed off its latest voice synthesis model, which can generate a convincing replica from just a 15 second sample. OpenAI is not releasing this publicly for trust & safety reasons, but like most of what's happening in AI, lots of people are working on this and the 'good enough' versions will become generic commodities pretty quickly. Presume any voice you hear can be faked (tell your family, and tell your corporate treasury department). LINK Meanwhile, after showing off the Sora video generation model in February, OpenAI published some short films made by actual film-makers. You can quibble with some aspects of these (most obviously the continuity) but they point to the creative possibilities, and of course contrast dramatically with what it would have taken to get any of this onto a screen 10 or 20 years ago. LINK Next, after floating a story that Sam Altman was raising 'seven trillion dollars' for AI, OpenAI and Microsoft are now apparently planning a $100bn ML system, to be called 'Stargate'. For reference, $100bn is roughly the combined annual capex of Google, AWS and Azure last year. Two probably-relevant considerations: there is a lot of chatter about how the power consumption of such systems could overload some power grids, and, the challenges of interconnecting training systems across multiple locations. LINK Amazon's Anthropic play Amazon invested another $2.75bn in Anthropic, after putting in $1.25bn last September, saying at the time it would do up to $4bn (meanwhile Anthropic has agreed to spend… up to $4bn on AWS compute). However, the Verge reports that Amazon also has a team aggressively building its own LLM models, codenamed 'Olympus'. How many top-tier or 'second-tier but not far behind' models will there be by the end of next year (presuming no more giant breakthroughs at one company)? LINK, OLYMPUS This week's new model: Databricks On this question, the latest company to enter the LLM stakes is Databricks, which does enterprise cloud stuff (I'm paraphrasing) and developed an LLM that's comparable to most other things out there on at least some benchmarks, and did it with a budget of $10m. LINK, INTERVIEW The week in AI The FT reports that OpenAI's App Store doesn't seem to be getting much traction (any more than plug-ins did). A weak thesis: there is a trough between people using ChatGPT out of the box on one side (either to experiment or for specific use cases) and on the other side specialised vertical apps with dedicated tooling and UI (which are still only starting to emerge). LINK New York City has an LLM chatbot that can answer questions about local rules and laws. Unfortunately, as others have found out, LLMs are probabilistic, pattern-based systems that make things that look like a good answer - they are not databases, and at a minimum, they need a lot of careful productisation before you can use them for something like this. Hence, the NYC chatbot gives a lot of answers that sound right, but get the law wrong. LINK Google's Gemini LLMs can't run on some of its own Pixel phones because they don't have enough RAM (generative AI is very RAM-hungry). LINK Last week it was reported that Apple is in talks with Google to use its Gemini LLM tech on iPhones in some form; this week, apparently it's also in talks with Baidu, which would make sense whatever the underlying use case, since Gemini is (probably?) blocked in China. LINK The White House issued an executive order giving guidelines for US federal agencies in their use of generative AI. LINK This content creator was hired by an agency to make some generic videos - that were used as a source to generate (using Heygen) a whole range of new video ads using her image and voice. Very Gibsonian. LINK, HEYGEN EU, DMA and regulatory micromanagement In the last few weeks the tech companies that the EU's DMA designates 'gatekeepers' have all published proposals for complying with the new rules, and the next stage is that the EU has opened investigations into anything that it doesn't like the look of. There is a big culture clash here: to simplify a lot, US law tends to be based on very specific lists of rules where you can check the boxes and now in advance what to do, where UK and European law has more emphasis on general principles. This means that in theory in the EU there are no loopholes, and plenty of flexibility, but it also means, as here, that these companies are guessing what answer the EU will like, on complex questions based on trade-offs where there is no one correct answer, with the threat of a giant fine if they guess wrong. There's also a tension: the EU is asserting general principles, but also getting very deep into the weeds of specific product management decisions. NEWS, LINK Chinese hacking The UK and US accused China of hacking UK politicians and voter rolls, and US infrastructure. (First rule of espionage: don't get caught.) LINK Swipe fees When you use a credit card, the retailer pays the card network a fee. In the EU this was recently regulated to well under 0.5% (under competition law), but in the USA, while debit card fees have a regulatory cap, credit card fees do not, and they average about 2% but can be double that. The US National Retail Federation claims that costs them about $160bn a year, which is roughly the same as total retailer net income. (If you didn't know, this is where loyalty points come from, and it might also be why Apple's credit card hasn't launched in Europe.) For the last 20 (!) years, US retailers have been suing Visa and Mastercard under competition law, and now there is a partial settlement, but a pretty limited one: it's complicated, but headline rates will only fall by 7 basis points. SETTLEMENT, BACKGROUND | 
No comments:
Post a Comment