Banning Tiktok in the USA, for real this time? Remember when Donald Trump tried to ban Tiktok, or maybe get Oracle to buy it? That effort collapsed with typical ineptitude, but the underlying issue didn't go away. The USA has foreign ownership laws for TV, but Tiktok is as big as TV now, and a major vector for news in particular, and though TikTok's CEO lives in Singapore, Bytedance is Chinese, and like all Chinese companies it is ultimately obliged to do what the government there says. Does it put its thumb on the scale of the content recommendation engines? It could. So, after years of building pressure and concern, the US now has a bill that would let the US president designate companies controlled by 'foreign adversary countries' and force their divestment. Will that happen? I'm not a political analyst. How big would that deal be, and who could buy Tiktok? Bytedance is a private company with no public financials, but the last valuation was apparently at $225bn. LINK, COVERAGE, VALUATION New models OpenAI showed the way, but lots of people are following, and not just Google and Meta: there's a whole middle tier of startups with competitive LLMs now. Microsoft invested in Mistral (a French startup!) last week, and this week Anthropic released the latest version of its 'Claude' suite of models, which apparently benchmark ar par or ahead of GPT4, at least on some benchmarks (there are so many LLM benchmarks that I need an LLM to summarise them all). As is now the trend, it's split into three tiers, trading power against price: Haiku, Sonnet and Opus (presumably the most powerful of all will be Jean-Claude). And Inflection's 'Pi' model is also close to launch (this company is founded by Mustafa Suleyman, previously co-founder of Deepmind). CLAUDE, COVERAGE, INFLECTION Hiring biases Unsurprisingly, if you use ChatGPT and other LLMs to summarise and screen CVs for hiring, their analysis will reflect the patterns and biases that are in the training data, so they will tend to down-rank (for example) black or female candidates. Amazon found some similar issues with an experimental AI system back in 2018: Amazon had been more likely to hire men in the past, so the system inferred that men were better candidates (and inferred gender from subtle differences in wording). We had a long conversation about AI bias at around that time: there a naive view that 'maths isn't biased' - yes, well done, but are you sure you know what's in the training data? The inverse, equally naive view is that you can fix this by changing the diversity of the team building the system: if you add more women to the team they may be more likely to spot bias against women, yes, but no more likely to spot that your image recognition tool tends to look at images of grassy hills and say 'sheep'. The biases aren't about human diversity per se - the issue is much more structural than that. LINK, AMAZON The week in AI The Indian government issued a very odd (non-binding) advisory that all AI models need pre-approval from the government, and need to be certain they do not have any bias or discrimination. You cannot guarantee deterministic results from non-deterministic systems, and you can't require individual, one-at-a-time pre-approval for every new product and feature in the tech industry. LINK Sergey Brin spent an hour at an AI event in SF talking, amongst other things, about Gemini's launch hiccups. LINK AMD has designed an AI chip intended to be slow enough for the US government to let it be sold to China. The US said no. LINK OpenAI fired back at Elon Musk's lawsuit, producing emails that appear to show that Musk understood and was in favour of the principle that OpenAI would need to become a for-profit. Does anyone really expect this to go anywhere? LINK There's been some chatter recently that AI-generated spam is a bigger problem in search results, and this week Google rolled out a new search model update intended to tackle that. LINK Everything is hacked, French edition Apparently the French medical insurance system was hacked, leaking some level of personal data for 33m people, over half the population. LINK Apple's first EU fine! Welcome to the big tech club! The EU fined Apple €2bn over the long-running complaint that it didn't let Spotify sell a subscription in its app nor tell people where they could buy one. I've always been (mostly) on Spotify's side here - unlike Epic, Spotify didn't just not want to pay Apple's commission, it couldn't - and there was no way that this situation was good for Apple's customers. On the other hand, as Apple pointed out at bitter length, this didn't actually seem to hurt Spotify at all. LINK, APPLE Browser ballots 2.0 As part of the DMA, Apple and Google have introduced choice screens forcing EU smartphone users to pick a default web browser from a long and randomly-sorted list. (Note - this is separate to being able to install a third party browser, which was always possible, and from browsers being able to use their own rendering engines on iOS, which is another new requirement). Microsoft had to do this on Windows in the EU from 2009 to 2014, but it had little to no observable effect. Chrome was launched at the same time, but took off globally, not just in the EU: it turned out that you didn't need a choice screen to get people to change their defaults if there was a genuinely better alternative. It's hard to see why things would be different this time: the default browsers on each device are already pretty great (unlike IE!), which makes this most likely to be just an annoying waste of time, and a source of support calls from confused users who accidentally tapped on the Onion Browser. APPLE, ANDROID, WINDOWS Apple flip-flops on Epic Epic opened a new developer account to make an app store for games on the iPhone in the EU (NB: the iPad isn't covered by the DMA): Apple gave it the account, then withdrew permission, and then, when the EU expressed interest, allowed it again. On one level this was hilarious. Epic did lie to Apple and break the developer agreement, deliberately and openly, as a PR stunt around Fortnite, and so Apple said that meant it couldn't be trusted to run a third-party app store. Cue lots of blustering and quotes on both sides until Epic promised, cross their heart and hope to die, not to break its word this time, and Apple pretended to believe that. But coming after Apple's u-turn last week on web apps, this looks a bit chaotic. Is Apple floating balloons to see where the EU will draw the line? After all, the DMA does say that Apple has to preserve security and privacy. Or is Apple in denial? TERMINATED, RENEWED | 
No comments:
Post a Comment