Sunday links: GPUs on every desk, AI Hotels, and attack of the Perplexity clones!

A little late in the day due to my current timezone, but here are this week's links:

• NVIDIA Puts Grace Blackwell on Every Desk and at Every AI Developer’s Fingertips. In a rather unexpected move this week, NVIDIA released a $3000 desktop computer that essentially packages its GB10 Grace Blackwell chip into a box you can bring into your home office. The claimed performance is one petaflop (a quadrillion floating point operations per second). For context, that's 1000's of times faster than most desktop computers. By known estimates GPT3 took about 3600 Petaflop days to train. So it would take around 10 years on a box like this. With 30,000 USD, you could, however, buy 10 of them in just a year. (This is not the whole truth – since a lot of effort goes into the parallel processing, but if gives you a sense of the power of the systems. With most of the world's (and NVIDIA's business focus) on cloud deployments, there is a fair bit of head-scratching as to why NVIDIA would do this. My guess is that it's perhaps a bit of a side-show, but more than anything, a big billboard to advertise the fact that it's now possible to get GPU compute in a small form factor that other computer manufacturers should consider as a target architecture. Your next Tesla may come with two of these in the boot. Further analysis here.
• New research: What do gamers really think about generative AI in games? This survey (hat tip Amin Bakht) asked 6,400 gamers how they felt about the use of AI content in games. The answers were surprising to me. They show an already accelerating shift towards comfort and acceptance from a mood that was decidedly hostile just a year ago. 60% responded they were neutral on the use of AI as long as the game was good. 20% responded positive and 19% negative. Older gamers and some other groups also trended more negatively on average. The responses do hide some nuances, though: there is still a backlash against "uses of AI that affect employment in the industry" and also a general negative sentiment toward the use of AI that "makes games just bigger rather than just richer & more immersive". They also hide the fact that the negative camp can be very vocal online and this scares big publishers especially. The challenge going forward, though, is that the laudable wish of gamers to protect industry jobs will be hard to maintain. What will likely happen is that small indie studios will begin to outperform using AI, succeed, and put pressure on larger studios, which will then be at a cost disadvantage. If they then cut staff, they will risk a backlash. The result, at the very least, is likely to be a slow hiring environment for large studios. Perhaps the most attractive options will be small studios. Now if only those studios could get distribution for their games to reach an audience.
• AI Hotel Planned for Las Vegas: CES 2025. For the biggest fail of the week, I'd nominate this announcement of a data-gathering, client-optimizing hotel. The press release quote includes the line "“We create a virtual copy of the guest,.. There is an onboarding before coming to the hotel. We capture information and use AI to scrape the internet, and then we track behavior while on property.” which results in optimizing room choices and amenities. Apart from probably sounding extremely creepy to most users... this is a massive opportunity. The obvious way to go for an AI hotel is to model it on the Hendrix in Richard K. Morgan's Takashi Kovacks / Clatered Carbon novels. In the Netflix series, these lovingly became the equally cool character Poe, (the Endrix estate did not permit the name of the musician's name to be used). Given the level of loyalty and defense the Hendrix provides it might be worth submitting to some data gathering.
• VCs say AI companies need proprietary data to stand out from the pack. This Techcrunch article makes one of the most obvious statements in AI investing, highlighting the benefit of privileged data for AI startups. It's also far from the whole story, though. If there's one thing we're learning, it's that in many domains, synthetic data or even distillation from existing models can fill in data gaps quickly. In cases where the proprietary access is to customer data, those customers could just as easily provide access to their data to a competitor. So either data has to be truly unique (an exclusive license) and deep, or it needs to be continually refreshed by a process only the startup has access to. The latter case is the strongest position to be in: building the company so that it is in a data flow of effectively its own creation (pricing, transaction flows, relationships, sentiment, or other similar things). With this type of data, at scale, it becomes impossible for others to replicate the source, and the source is a continually moving target.
• Build a clone of Perplexity with LangGraph, CopilotKit, Tavily & Next.js. Perplexity seems to be everyone's target of choice this week, with demos being built that replicate the functionality of the growing startup. Dev.to provides a tutorial, and there is also this replit/gemini based opensource project by Ammaar Reshi. I still love perplexity as a search tool (it's mostly replaced Google for me – though Google Deep Research looks promising), but it's undeniable that every AI company layered on top of Foundation models needs to find a really solid anchor to defend itself against clones. My guess is that Perplexity will be able to do that with more features like model choices, shared research spaces, and memory, as well as a growing brand. One hundred clones doesn't mean their path is not viable. A bigger challenge is if players with existing audiences replicate the experience.

Wishing you a great week!