Benedict's Newsletter: No. 549

The great AT&T hack, plus a16z's GPU cluster, and AI questions for the summer.

NO. 549   FREE EDITION   TUE 16 JUL 2024 SPONSORED BY TOKEN2049 Join 20,000 attendees at TOKEN2049, the world's largest Web3 and blockchain event, on 18-19 September. Industry leaders and global cultural icons will converge in Asia's leading economic hub to unveil what lies ahead. Secure your pass for a festival experience unlike any other: Use code BEN2049 at CHECKOUT now to get 15% off. REGISTER NOW Upgrade to Premium You're getting the Free edition. Subscribers to the Premium edition got this two days ago on Sunday evening, together with an exclusive column, complete access to the archive of over 500 issues, and more. SUBSCRIBE OR UPGRADE My Work The AI summer Hundreds of millions of people have tried ChatGPT, but most of them haven't been back. Every big company has done a pilot, but far fewer are in deployment. Some of this is just a matter of time. But LLMs might also be a trap: they look like products and they look magic, but they aren't. Maybe we have to go through the slow, boring hunt for product-market fit after all. LINK The VR winter continues Meta has spent at least $50bn on VR and AR so far, but we're still in the VR winter: the devices aren't good enough or cheap enough and the user base is flat. But no matter how good the devices get, how many people will care? LINK Another Podcast: The AI Summer As we go into the summer, we know a lot more about generative AI than we did six or nine months ago - or at least, we have better questions. LINK News Apple and Microsoft give up their OpenAI board seats Last week it was reported that Apple had secured an observer (i.e. non-voting) seat on OpenAI's board as part of its deal to link to ChatGPT access in iOS and macOS. This week we hear that Apple is giving this seat up - and that Microsoft, which has invested $13bn in OpenAI, is giving up its own observer seat as well (which it took after the board tried to fire the CEO last year). Instead, there will be 'regular meetings' to 'inform and engage key strategic partners'. This company continues to be a paragon of corporate governance. The backstory, presumably, is increasing scrutiny from competition authorities, in particular at deals like this. LINK A16Z has its own GPU cluster?  a16z has apparently built up a cluster of Nvidia GPUs, to rent out to its portfolio companies, and plans to have as many as 20k (which would be hundreds of millions of dollars at a minimum, depending on the models). This is an interesting turn on their 'entrepreneur service' model - no-one can get enough GPUs right now, and many startups don't have the money or bargaining power to to do this themselves, so a16z provides that as a service, and of course as a lever to get into hot deals. (Also, I wonder how this connects to the crypto fund?) LINK The week in AI Graphcore was trying to make AI chips (before the current surge driven by LLMs), but chips are hard, and now it's been bought for scrap by Softbank. LINK The sharable, interactive 'Artifacts' in Anthropic's Claude 3.5 are an interesting pointer to what products might look like when they're native to LLMs. LINK Odyssey might be the latest hot video-generator, this one focused on Hollywood. LINK Sam Altman and Arianna Huffington are backing an AI health coach. LINK A16Z invested in Hebbia, whose product enables LLM-based agents for financial analysis, with a lot of tooling around the use cases and error rates. LINK Everything is hacked - a lot  Someone accessed several months of call and SMS records for all AT&T customers for six months in 2022. This doesn't give them the content of those calls and texts (obviously), nor FaceTime / WhatsApp / iMessage, but it would reveal, say, who called which journalists and politicians. This seems to have come from an inadequately secured account on the cloud provider Snowflake, which has been in the news a lot for this lately. LINK, FILING Meanwhile, a ransomware attack in Indonesia took down large parts of the government. LINK 1 , LINK 2 And, someone reverse-engineered Ticketmaster's ticket generation algorithm so they can make their own tickets. LINK About What matters in tech? What's going on, what might it mean, and what will happen next? I've spent 20 years analysing mobile, media and technology, and worked in equity research, strategy, consulting and venture capital. I'm now an independent analyst. Mostly, that means trying to work out what questions to ask. FIND OUT MORE Ideas Bloomberg reports that OpenAI has proposed, internally, a 'level one to level five' scale of progress towards AGI, placing itself at level one today. LINK, CHART Goldman Sachs released a report questioning whether the current wave of capex into generative AI (at least $200bn this year alone) will be supported by equivalent returns, given that 1: it could take a decade for real applications to be built and widely deployed (even cloud is still only 1/4 to 1/3 of workflows after 20 years) and 2: how much of the result will be cost savings and lower prices rather than new revenue and new profit (the airline industry is a case study of how a world-changing tech can produce no profits). LINK Andrew Ng, one of the pioneers of machine learning, on the problems with California's proposed bill to regulate LLMs. The narrow problem is that the bill is a other crude wish-list from doomers, but more broadly, as I've written a few times, trying to address problems from generative AI by regulating the models is like trying to stop fraud by regulating spreadsheets. LINK One of the core arguments around generative AI is how 'clever' the models really are as opposed to how clever they look - do they understand, or are they copying understanding? Here, a study suggests that ChatGPT is far better at coding problems that are likely to be in the training data (which suggests the latter). LINK Apparently, doctors in the US are now using LLMs to argue with insurers. This is pretty much exactly the joke from last year that half of LLMs will be turning three bullet points into three paragraphs and the other half will be summarising three paragraphs into three bullet points. LINK A technical post from Asana on how it assesses frontier models for incorporation in its products. LINK There is a market for smuggling sanctioned Nvidia chips into China. LINK A profile of one of the entrepreneurs using generative AI to create spam websites at scale. LINK Interesting little story about Meta and Vodafone cooperating to reduce network traffic from video. LINK Microsoft ordered staff in China to use iPhones for work rather than Android, for security reasons (inherent advantage of a managed system). LINK Bangladesh already has millions of electric rickshaws. LINK MG Siegler on the new spate of billion dollar AI acquihires, which look like a rather transparent way to dodge competition regulators (and one that's attracting scrutiny on that basis). LINK Outside interests The US and Germany accused Russia of planning to assassinate the CEO of Rheinmetall. LINK The empty newsstands of the New York subway. LINK DJI made a video of a drone flying to the top of Mount Everest. LINK The iPhone notes app as outfit generator. LINK The EconoMac index. LINK The ultimate in mindfulness-  a simulator of the Windows disk defrag tool. Just set it running and wait. LINK On that theme, Apple has approved a 90s PC emulator for iOS. LINK Data Coatue's annual presentation on the state of venture tech investing, and especially AI. LINK A guesstimate that OpenAI has $3.4bn of ARR (less than Accenture's generative AI business!) and that only 15% of that is from the API (= developer adoption). LINK YouGov did a survey of US public attitudes to AI, with some very vague questions and answers. What does it mean to ask people if they think it's ok to use 'AI' in the workplace? We are already, and have been for decades. LINK Spotting a surge in academic use of ChatGPT through excess verbosity. LINK Preview from the Premium edition Summer AI questions  It's the summer, and things are slowing down (see the news section above), but LLMs have been The Thing for  a year and a half now, and we're starting to work out what the questions are. Here are five things I wonder about.  Do these models keep scaling? Are they still scaling?  We got LLMs to work by throwing vastly more data and compute at them than seemed plausible, and then we made them better by making them bigger again, but for how long, how much bigger, and how much better?  On one side people like Dario Amodei at Anthropic airily talk about spending billions or tens of billions for better results, but on the other there are plenty of ML scientists who are hesitant, to say the least least, and while the benchmarks keep getting better, it's not clear that that   THIS IS A PREVIEW FROM THE PREMIUM EDITION - PREMIUM SUBSCRIBERS GET THE COMPLETE COLUMN EVERY WEEK. YOU SHOULD UPGRADE. Upgrade to Premium You're getting the Free edition. Subscribers to the Premium edition got this two days ago on Sunday evening, together with an exclusive column, complete access to the archive of over 500 issues, and more. SUBSCRIBE OR UPGRADE THIS WAS SENT TO abdulkerim.bosnia.business@blogger.com ON 16 July, 2024, BECAUSE YOU SUBSCRIBED AT BEN-EVANS.COM. You can UNSUBSCRIBE, OR READ THE PRIVACY POLICY. OUR MAILING ADDRESS IS Benedict Evans, 2 Golden Square, London, W1, United Kingdom.   COPYRIGHT © 2013-2024 Benedict Evans, ALL RIGHTS RESERVED.