Safety
The AI safety debate continues to bubble along happily. Google, no doubt feeling left out, said that back in May, Gemini systems in a cyber test managed to hack an external company’s systems, though in this case they were acting as intended, while another research group managed to bypass Anthropic’s restrictions to use Claude to break into an OpenAI system (um, because OpenAI’s opsec wasn’t good enough). A few more mid-level AI researchers quit Google, OpenAI and Anthropic, voicing the same sort of concerns we heard last week. And Anthropic, showing a hilarious inability to read the room, turned out to be setting up a biology research lab of its own, right when it’s claiming that AI could decide to make killer viruses and wipe out humanity. GEMINI, CLAUDE, BIOLOGY
China, on the other hand (like many people in tech and in AI), thinks the whole thing is nonsense and so it sees no reason at all to slow down, and that’s probably all that matters. LINK
AI numbers
Anthropic was expected to release its S1 IPO filing last week, but apparently it delayed to the end of the year for more feedback. Meanwhile, the FT says it told investors that it currently has 80% gross margins before payment to distribution partners such as AWS (should we call that ‘adjusted’ gross margin? ‘Gross gross margin’?) and positive adjusted operating income, which means before SBC. This reminds me of the mobile operator in ~2002 that reported ‘EBITDA before customer acquisition costs’, back when people used to joke about ‘EBBS - Earnings Before Bad Stuff’. Creativity aside, though, the real problem is that whatever today’s numbers look like, we don’t know how fast token prices will fall, nor how much pricing power a model lab might have, and we don’t know where mobile training costs will go. NUMBERS, IPO
The same issues apply to OpenAI’s revenue and cost projections to 2030 (also in the FT). It expects $278bn negative FCF from 2026 to 2030, and $350bn of revenue in 2030. These are big numbers, but they’re also illustrations, not forecasts. LINK
The week in AI
Meta’s consumer agent play, Muse, hit the top of the US app stores, and Meta also launched a Mac desktop app. It feels like the entire Meta growth cannon is pointed at at Muse, but I am very unsure whether this is a mainstream activity. Meanwhile, the super-hot Silicon Valley personal agent is Instinct, which remains invite-only but is apparently raising $1bn on a $10bn valuation with 100k active users. (Reminder: in 2012 Meta bought Instagram for $1bn with 30m active users and people said that was crazy.) LINK
OpenAI paid $300m to buy a hardware startup working on high-quality smartphone cameras. LINK
Disney hired its first ever ‘CTO’, picking the former head of Character AI, which was very buzzy last year (indeed, buzzy enough for Disney to send it a C&D). LINK
The FT reports that China is tightening control of foreign travel for anyone that it thinks has an important role in a strategic technology (recall the Manus fuss). LINK
Jev seems interesting: much faster, simpler decision-making AI to use within software. LINK
OpenAI is experimenting with allowing brands to run their own sponsored agents within ChatGPT as part of the ad product, for users to talk to after clicking an ad. LINK
AI kill chains
Bloomberg reports that the chain of errors that led to the US bombing a school in Tehran last year was, much as one would expect, a combination of out-of-date intelligence, pressure of time (both old problems), and over-reliance on AI, in this case, a system from Palantir, where automated analysis wasn’t checked properly before the button was pressed. LINK
US crypto rules
It looks like the US legislature won’t pass the Clarity Act, which is supposed to clean up the mess of conflicting ideas around how blockchains should be regulated. I have never found blockchain particularly interesting except as a piece of financial plumbing, but I do have sympathy with the complaint from the industry that US regulators acted as though this was only a financial instrument and that nothing about it required any new rules or new considerations. (It’s also worth noting that this was a big part of the reason for a bunch of people in tech to support Trump over Biden.) LINK
EU social bans for kids
The EU has picked up the vibe for banning children's access to the internet and social media, and is now proposing a general rule across the EU to shut off access to both social media and online games (always a slightly fuzzy distinction). As I’ve written before, I'm ambivalent about this, since the evidence of harm in generality is extremely weak and conflicted, no matter how loudly people insist that it's certain, and I’m pretty sure Minecraft isn’t bad for kids. LINK
Roblox unbundles the platform
Roblox wants to expand beyond the core social platform, allowing people to turn experiences and games built within Roblox into stand-alone apps. LINK
Merging Tesla and SpaceX?
One of Elon Musk’s superpowers is shuffling, trading and cross leveraging different promises, different capital structures and different sets of investors. That seemed to me the best explanation of why he decided to make SpaceX rather than Tesla the vehicle for building a foundation model lab. Now he might swap the cups around again suggesting (teasing perhaps) that he will join the two companies together. LINK
|
