Alexa is back (while Siri is MIA) Amazon has finally launchd a new LLM-based version of Alexa, which means it can answer questions that haven't been pre-scripted. $20/month, but free for Prime members (who must be most of the active installed base). While some of this uses Amazon's own Nova models, it also leans on Anthropic (Amazon has an equity stake). I wonder, though - does this make Alexa itself more valuable to Amazon? LLMs (and those Nova models) will be used throughout Amazon in dozens of ways, and AWS will make a new business selling them as infrastructure. But Alexa itself failed to produce any actual revenue or broader strategic benefit for Amazon - it failed to become more than a voice-activated clock-radio. Will LLMs unlock new use cases and revenue, or will this now be just an LLM-powered clock-radio? Meanwhile, Apple trailed a similarly rewritten (and potentially much more useful) Siri at WWDC last summer, but that still hasn't been launched and it's not even clear if it will make it out of the door before this year's WWDC (the steadily-increasing number of Apple features that miss the autumn launch is a story in itself - the launch cadence used to be metronomic). LINK Capex calibration It's pretty clear that OpenAI is planning to reweight its infrastructure use from the current Microsoft deal to Softbank's new building plans (whether or not that ends up looking anything like 'Project Stargate'). LINK Meanwhile Meta has said it will spend $60-65bn on datacenter capex in 2025 and will spend 'hundreds of billions' 'over the long term' - now, apparently it's in talks about a 'data centre campus' costing $200bn. LINK Conversely, Deepseek keeps releasing papers pointing to a quite different approach to capex - here it claims that it has theoretical margins of 545% to run inference on the R1 model. LINK New Models, feeds and speeds After trailing two weeks ago, OpenAI launched 'GPT4.5', its last model not to use the new 'chain of thought' (AKA reasoning) approach seen in o1 and o3, which most people in the field think is probably the path forward. This model seems to be somewhat better than the preceding 4o, though not dramatically so, but doesn't match o1 and o3, and is 15-30x more expensive to run than 4o, which puzzled a lot of people. LINK Meanwhile, Anthropic released its own new model, Claude 3.7 Sonnet, with a hybrid approach that can switch back and forth between one-shot answers and chain-of-thought'. LINK Apple re-announced $500bn of US investment Apple, with Donald Trump, said that it would invest over $500bn in the USA the next four years. As the WSJ pointed out, this appears to be a politically expedient re-announcement of plans that Apple already had, as it reweights manufacturing away from dependance on China, a project that began back when it was China that was the unpredictable bully that needs to be flattered. LINK, WSJ Sergey Brin back in the office Sergey Brin has no formal role at Google beyond board-member, but he's spent a lot of time the last two years back at Google working on AI, and he's now sent an internal manifesto: "stop making nanny products," "The race to AGI is afoot" - and everyone should be in the office at least five days a week and working 60 hours a week. That kind of commitment (not actually very extreme, either) has been rare at Google for a very long time. LINK The week in AI Chegg is suing Google over 'AI Answers' hurting the business - isn't the real problem that ChatGPT is better at homework? LINK Perplexity is thinking about a web browser. Not sure about this - the default search field is user acquisition, sure, but you've got to get people to install the browser in the first place and meanwhile, isn't this last century's model? LINK Meta is considering launching a stand-alone AI app - all the other model companies have their own icon on the home screen but Meta's models are buried inside tabs within the existing apps. Time for some unbundling? LINK The UK is running a consultation on changing copyright law to carve-out AI training uses. Newspapers and musicians are upset. Memo to Americans: other countries have laws too, and unlike the USA, they can change them. LINK New TV MrBeast is raising 'a couple of hundred million dollars' at a $5bn valuation. He's looking at expanding his food brands as well as video production, and is apparently profitable on $400m of revenue. LINK Apple moves on age verification There's a growing trend for laws requiring companies with adult content or services aimed at adults (eg gambling) to verify user ages. However, those companies really don't want the friction of imposing this on their users, nor the security, privacy and risk issues of collecting hundred of millions of passports themselves. This particularly hard for use cases like Instagram or Snap, where you want to know both if the user is over 13 and if they're under 18, and they probably don't have a passport to show you anyway. The obvious solution is for Apple and Google do this instead and provide it as a privacy-assured API to developers and websites, but they don't want to do collect ID documents either, for similar reasons. Now Apple proposes a slightly different solution: it will simplify the process for a parent to set up a child's phone, and allow a developer to ask for the age-range of the device's user: 4+, 9+, 13+, 16+ and 18+. Now Meta, Onlyfans or Draftkings can ask the device for an age range without needing any personal data, and you shift the responsibility to the parent (while hopefully making that easy for them). The catch is that this only works for apps, and Apple doesn't allow porn apps, pushing them to the web, so this doesn't help those companies unless Apple extends the API to Safari. LINK Indonesia's Apple 'negotiation' Last autumn Indonesia banned iPhone 16 sales because they don't have enough locally-made components. This is an interesting use of industrial policy, but it was hard for Apple to comply given that even the Indian operation, at much greater scale, is still really just assembling components made in China. Now, though, there's a deal, with Apple agreeing to build an R&D facility and invest in local plants to make components for other devices. LINK Citi's UI problem The FT reports that Citigroup mistakenly transferred $81tr to an internal client account, before catching and fixing the mistake a few hours later. This is the same bank that mistakenly wired $900m of Revlon's money to its creditors in 2020. In both cases, part of the cause was very bad UI design, where multiple people checked the transaction and didn't see the problem. Fixing this kind of stuff is what 'digital transformation' actually means. LINK RIP Skype Microsoft will shut down Skype this May. See this week's column. LINK | 
No comments:
Post a Comment