Meta's Llama 3 beats OpenAI Meta released the largest version of its newest model, Llama 3.1, and on most benchmarks this is at least as good as the latest and best from OpenAI. And it's open source (sort of) - you can download it yourself today. See this week's column. LINK Google gives up on killing cookies Google has been talking about replacing cross-site cookies in Chrome since at least 2019. As the dominant player in online ads and the maker of the dominant web browser (excluding iOS), it's tried to pull the rest of the industry around some kind of anonymised interest-based targeting, calling this a 'privacy sandbox'. Chrome would know that you've been looking at lots of websites about cars and could tell other websites 'load a car ad on this page' but the actual tracking data would never leave your device. And, meanwhile, Chrome would stop allowing cross-site (AKA third party) cookies by default (Apple's Safari already did this), forcing everyone to switch. However, Google has never been able to solve the stake-holder alignment (i.e cat-herding) of persuading privacy regulators, competition regulators, publishers and ad-tech companies that this is a: a good idea and b: would work. Now it looks like Google is giving up: instead of killing 3P cookies, it will keep them, and "introduce a new experience in Chrome that lets people make an informed choice". The devil is in the wording, but that sounds like an 'ask people if they want to block cookies' button. Now the ad industry is scrambling to work out what that means, and what comes next. LINK OpenAI (finally) does search OpenAI has finally announced that it will do a search engine powered by ChatGPT. In principle, you can point an LLM at an index of the web and get it to answer questions about that index, or summarise results that a more conventional search engine finds from that index, and perhaps other approaches as well. Bing Copilot, Google 'Search Summaries', Perplexity and a few others are all pursuing versions of this, and now, so is OpenAI. I think there are two views on this concept. On one hand, as we saw in the (very brief) excitement about Bing Copilot, and also in the buzz around Perplexity, there is at a minimum a class of search query where the 'job to be done' might be better served by an answer or a summary, not a link to a web page, and so LLMs could displace classic Google search (even if this ends up being done best by Google itself). But on the other, the error rate inherent to LLMs is an unsolved and quite possibly unsolvable problem: these systems tell you what the answer would probably look like, not what it is, and that may matter too much for too much general web search: this might work best or only work in narrower verticals. We don't know, but, OpenAI will join the effort to find out. LINK OpenAI burn rates Accuracy and ranking is one barrier to entry in search - another is just how much money you have, and the Information reports that OpenAI is already on track to burn $5bn this year. I doubt that it will have trouble raising more (and trying to get a share of the firehose of cash that comes from Google search might help), but it's still a reminder that LLMs have unprecedented capital-intensity, especially for a technology that has yet to find broad product-market fit. LINK LLMs and IPR Perplexity is in even more trouble with publishers, with Conde Nast now sending a cease-and-desist for its AI 'summaries' that are-perhaps-too-often just copies of other people's work. Meanwhile, someone leaked an internal spreadsheet of training data for Runway, a very buzzy-in-Hollywood generative AI video maker: it appears to have scraped thousands of influencer videos from YouTube, against ToS. There is a growing collision between the philosophical view in many AI circles that training-by-looking is no different to what people do (after all, these systems aren't Napster - they can't generally reproduce what's in the training data) and the legal status of 'using' people's property in an entirely new way but without any new model for permission. PERPLEXITY, RUNWAY Deepmind does maths Google's DeepMind built systems that solved four of six problems in the International Maths Olympiad (it looks like DeepMind is still doing pure research even as it was re-orged to be more focused on product). Projects like this are valuable in their own right as pure research, but they're also aimed at getting models to be better at 'reasoning', as opposed to 'pattern-matching' (both crude terms). LINK Social sextortion There have been a bunch of stories about scammers extorting minors on social platforms, with a few suicides, and now Meta has removed 63k Instagram accounts linked to the so-called 'Yahoo Boys' scammer scene in Nigeria. LINK, REPORTS Meanwhile, The Information reports that Snap has many of the same problems (obviously, because all social messaging platforms do), but, with its smaller scale, struggles to resource the teams that try to address this. (A few years ago Alex Stamos, former Meta CISO, suggested that Meta should offer content moderation as a service to other smaller social platforms.) LINK Remember Cameo? This might be one to put in the 'Covid rotation' file with Clubhouse - Cameo was briefly valued at $1bn but is now too hard-up to pay a $600k fine. Sad, but perhaps not surprising. LINK Apple Maps on the web Over a decade after the famously-disastrous launch, Apple Maps is now pretty good, but Apple never made it a website (fitting the general apps-are-better approach), but now, suddenly and somewhat randomly, there is a beta that works in a browser. I wonder if this is a regulatory thing? LINK | 
No comments:
Post a Comment