Apple does AI Apple published a paper announcing and giving many details of 'MM1', its own generative AI LLM, which appears to be comparable to market-leading models on at least some benchmarks. I'm old enough to remember the last time that people said Apple 'couldn't do AI', and that time too this was fundamentally a public, openly-available technology. We should expect Apple to spend a lot of this summer's WWDC talking about AI: there will be a lot of AI-powered features and (like last time) lots of things that work better without being branded 'AI'. The obvious question of course is how much this runs on the device, and whether Apple does an actual chatbot, given all the challenges both in cost-per-question and reliability, but I expect a lot of discrete unbundled features - 'edit a photo with Siri', say. Remember, incumbents always try to make the new thing a feature. MM1 PAPER, DARWINAI America's Tiktok law? America's new law to force Bytedance to divest Tiktok is gathering momentum. See this week's column. LINK The US needs a privacy law: case 1 of ♾️ The US is debating a law about Tiktok, but meanwhile GM, Subaru, Kia and Mitsubishi provide personally identifiable driver telemetry data to car insurance data brokers, often hidden as part of 'click-through' agreements that drivers never see. This kind of thing would be all kinds of illegal under GDPR. (Note, tangentially, that China has placed restrictions on where Teslas can drive to prevent data leakage, especially from the cameras.) LINK The week in AI Amazon continues to deploy generative AI for Marketplace: this week it has a tool that will auto-generate product listings based on your website. LINK Robots are having a moment around AI and generative AI: Physical Intelligence did a $70m seed round this week. LINK Lots of excitement in software development circles around Cognition's Devin, which it claims as a big step forward in using generative AI to write analyse and deploy code for you, combining many different kinds of tasks. LINK The CTO of OpenAI, Mira Murati, gave a car-crash interview with the WSJ about Sora, its new video generation model, in which she replied to basic questions about what was used as training data by looking surprised and saying 'I don't know'. Even if you don't want to say, you should probably have actual answers (even non-answers) to such obvious questions, especially when you're being sued over this. LINK Deepmind worked out a method to steal private information about the internals of an LLM by crafting special queries into the public API. LINK Midjourney accused Stability of trying to scrape its systems. People who think that the entire field of image-generators are based on theft (I don't) will think this is hilarious, but meanwhile, Stability seems a bit… accident-prone. LINK India thought better of requiring pre-approval for all new AI models. LINK Th EU AI Act The EU finally passed its AI Act. The process for this began well before the emergence of Generative AI, which was hastily bolted on in the last year, but trying to regulate a space that's changing so much so quickly, and where so much is unclear, seems misguided. More fundamentally, though, I think that regulating 'AI' is the wrong level of abstraction: you would not try to prevent the UK Post Office scandal by creating a Database Regulator, and you wouldn't try to prevent a repeat of FTX by creating a Spreadsheet Regulator. A law to cover anything bad that might happen with AI - and there will be plenty - is trying to combine far too many quite unrelated kinds of problems. NEWS, LINK Starbucks gives up on NFTs A couple of year ago Starbucks decided to build an NFT-based rewards program - now, given the collapse in interest in the space, it's quietly shutting the whole thing down. You could make an argument that putting such a scheme on a blockchain would provide a new layer of trust (Starbucks can't mess with the rules, or at least you can see if it does) and 'composability' (third party developers can plug in) but not everything needs to be a platform - sometimes a database is just a database. LINK, STARBUCKS Who is Satoshi? Bitcoin was created anonymously, by someone using the (presumable) pseudonym Satoshi Nakomoto, and for years an Australian called Craig Wright has been claiming that was him, and suing people who expressed skepticism. Now a UK court, after a six week trial, has held that no, he is not Satoshi. LINK Texas versus age filters Texas is the latest US state to require adult sites to get an ID (or use a third-party provider) before allowing access, and the biggest site (you know the name - rhymes with...) has duly blocked users there rather than comply, arguing that such verification doesn't actually work and that it should not be collecting such information anyway. Meta and others have argued that the smartphone OS (i.e. Apple and Google) should play this role, verifying an ID in a secure on-device way and passing validation on to sites and apps with an anonymous API, which seems sensible to me. LINK | 
No comments:
Post a Comment