Google charging for AI? Most people in tech weren't born the last time there was marginal cost for a user pressing 'OK' - you'd have to go back to mainframes in the 1970s. But LLMs, for the time being, are so expensive to run that you can't give uncapped usage and you can't really make it free. In the next few years this will probably change, as the chips pile up and the models get more efficient (really, as markets function), but for now, what do you do? The FT reports that Google is considering a paid tier of some kind. LINK The week in AI Amazon says it will spend $150bn on datacenter capex over the next 15 years to support generative AI (it doesn't split out AWS capex separately, but probably spent about $30bn last year, so this isn't necessarily a huge change). LINK Apple's AI team published another very suggestive paper, this time about using on-device models to understand things like what's happening on the screen (what does the user mean when they say 'pick the third one'?). LINK The US and UK made a joint agreement on testing AI safety. I have very little concern about AI safety, nor interest in trying to regulate for it, so I post this purely for the record. LINK The NY Times will offer AI-powered audio versions of most of its articles. This is as notable as part of the NYT's multi-modal media strategy as it is for the AI aspect. LINK Perplexity says it's considering ads. It's not yet clear how much an LLM can really replace web search's 'ten blue links' or whether it might still drive traffic, but if you ask a question with clear intent ('what credit card should I get?') that can be matched with an ad. LINK Meta announced a policy on labelling content made with AI (but how can it know which is which?). LINK Tesla follies Tesla now says that its 'full self driving' (which cannot, however drive the car fully) is 'supervised full self-driving'. Elon Musk claimed 1bn miles driven with a cumulative chart that appears to show flattening growth, and also claimed that a 'robotaxi' will launch soon. He first claimed we would have full autonomy 'in two years' in 2015 - note also that the Roadster supercar was announced in 2017 and hasn't been seen since. The problem as ever: he says lots of stuff, much of which is a wild exaggeration or just not true… but then, there are the self-landing rockets and the electric car company. LINK, CHART Microsoft unbundles Teams Back in July 2020, Slack complained to the EU that Microsoft was bundling Teams with Office, which should have been a pretty straight-forward competition question. It took the EU until July 2023 to open an investigation, and Microsoft agreed to unbundle in the EU in October: now Microsoft will do the same globally. I doubt many big companies will switch at this point: everyone has made the choice and these are sticky, commodity tools. But either way, this is too long to take action in a market that moves this fast. LINK Tiktok isn't the only Chinese app(!) This weekend four of the top five apps on iOS are Chinese. TikTok raises special issues in its ability to decide what news and media we see, but Bytedance also owns Gauth, a maths app that's number 5 in the education chart. This is the drawback to the US writing laws about individual apps rather than trying to identify structural problems and then trying to solve them (by writing an actual privacy law, say). LINK Meanwhile, the US may be tightening scrutiny of the de minimis rule that means parcel shipments valued at under $800 don't pay import duties: at the moment this covers shipments from Shein or Temu suppliers in China that go directly to US consumers (there were 410m de minimis parcels in 2018 and 1bn in 2023). LINK Trump Media Truth Social, the Twitter clone part-owned by Donald Trump, went public last week. According to the associated SEC filings, it has pretty much no revenue, and it refused even to say how many users it has (almost none), or disclose any other operating metrics either. However, it currently has a notional market cap of about $5.5bn. This is a politics story, and perhaps a securities law story, not a technology story. LINK, FILING Amazon still needs barcodes Amazon has been scaling back its 'just walk out' grocery checkout technology, which used cameras and AI to work out what you'd picked up. The Information reported that Amazon has 1,000 people in India doing data labelling and error-checking, but that the error-rates and hence level of back-stop manual review remain too high for the system to be practical. (Note: some people thought (or decided to think) that this means the whole thing was just a con to offshore checkout workers, which shows a basic confusion about how data-labelling works.) LINK | 
No comments:
Post a Comment