🦞 OpenClaw I wrote about OpenClaw a few weeks ago, when it was the craze of the week. Since then, OpenAI bought out the founder, and it turned into a consumer craze, apparently, in China. This week, China's cybersecurity agency issued a warning that it's totally insecure (undisputed), and Meta picked up Moltbook, a chatroom used by OpenClaw agents to talk to each other (maybe) and for people to troll each other pretending to be OpenClaw agents (probably). A cynic would call this the new Clubhouse, and when I wrote about this a few weeks ago, I talked about the pent-up demand to build AI stuff on your own computer with your own hands instead of relying entirely on cloud APIs from big labs, and of course, the desire to build the kind of autonomous assistants that big consumer companies can't ship because they're so unpredictable (ask it to tidy your inbox and it might tidy your inbox by deleting all your email). But another comparison might be desktop Linux. Engineers always love building their own systems, and they always said that with 'just a little more work' Linux would be ready for consumers - 25 years later, we're still waiting (no, not Android), but the web runs on Linux. CHINA, MOLTBOOK Meta cutting deep? Reuters claims that Meta is looking at cutting 20% of headcount. This would be partly for cashflow to invest in AI infrastructure (it plans to spend over 50% of revenue on capex this year and has already been out raising capital), but probably also a reset versus previous hiring (and perhaps more pruning of the Reality Labs stuff). This has already been really messy, though: Meta ended 2019 with 45k employees, hired 27k across 2021 and 2022, did an 11k layoff in 2023 but still ended the year up by net 15k, cut another 20k in 2023, added 12k back by the end of 2025 bringing it to 79k, and if it now cuts 20% (~16k) that would bring it back to mid-2020 numbers. It would also mean that Meta had laid off more people since 2019 than the company actually employed then. Everyone over-hired in the pandemic, but this seems careless. LINK After LLMs? Yann LeCun has raised $1bn at a $3.5bn valuation to build 'world' models, which he has argued are the next step after LLMs, since he thinks LLMs won't keep scaling and in particular won't get to structural understanding of cause and effect. Fei Fei li, another pioneer of machine learning, also raised $1bn to build world models in February (no public valuation but apparently $5bn). LINK Copilot Cowork Microsoft will integrate Anthropic's Claude Cowork into its 'Copilot' product/brand/distribution channel. Depending on your point of view, this shows either the strength for Microsoft's arms-length relationship with OpenAI, leaving it free to pick and choose the best tech to resell, distribute, and integrate - or that Microsoft is a product-taker and strategy-taker beholden to other people's roadmaps. But the risk is that Copilot becomes the new IBM 'Watson' - a meaningless brand name mostly associated with annoying additional fees other than specifically useful tools. LINK Google Maps gets AI Google continues productising AI by adding Gemini to maps. This lets it synthesise data that's probably already there but hard to find (My phone is dying — where can I charge it without having to wait in a long line for coffee?) while also enabling better use-cases (plot this route but also find a good place to stop for food and petrol). This is all valuable but also pretty obvious (and hard for Apple) - see this week's column. (Meanwhile, this is a painful contrast with Microsoft, which doesn't have maps, or AI.) LINK Grammarly flounders? On the other hand, this was a bizarre story - Grammarly, which clearly faces an existential challenge from LLMs, launched a new writing tool that claims it would make editing suggestions 'in the voice of' various named, living journalists and writers - without asking or even telling them. It's one thing to train a model on everyone's work, but quite another to sell a product on people's names. LINK Hacking McKinsey? Someone claimed to have hacked McKinsey's internal LLM-based knowledge-sharing system, getting access to a complete database of Teams chats, spreadsheets, and PowerPoints of client work. It's not clear to me how real this is, but it points to a much broader issue: if you want to use AI to find, analyse, and synthesise information across your company, then your traditional tools to compartmentalise and rate-limit access to data don't apply, and you'll need new ones. LINK Trump's TikTok fee The investors who bought TikTok from Bytedance in a forced sale will pay around $10bn as a commission to the US government. Really? LINK xAI wobbles After a couple more departures, only two of the original 13 co-founders of Elon Musk's xAI model lab remain, and he said he's rebuilding it 'from the ground up' (after selling to SpaceX shareholders). Setting aside the personalities, the challenges for xAI have always been why it was different. LINK Travis is back? In the nine years since he was kicked out of Uber, Travis Kalanick has been busy, but quiet, mostly working on cloud kitchens, but now he's decided to make a PR splash, launching a new company name (Atoms) and an expansion to mining and robotic transport. This is a funny one. The launch site opens with how terrible it was that an unnamed and shadowy 'investor' (which everyone knows was Bill Gurley at Benchmark) exploited the death of his mother to take his company away, but there's no acknowledgement of the escalating issues that led or at least enabled Gurley to push him out. He may just not believe it - some of his associates from then certainly think it was nonsense, and 'Uber Derangement Syndrome' was real, but some of the stories were true, too. Either way, he's an extremely capable and aggressive entrepreneur with a grudge, and grudges can be powerful, but also distracting. (NB: I worked at Andreessen Horowitz, an investor in Lyft, when Kalanick ran Uber; Andreessen Horowitz's founders have their own long-standing public grudge against Bill Gurley). LINK, COMMENT | 
No comments:
Post a Comment