The astute among you will have spotted that last week's post didn't happen. My apologies for the gap in service! Some travel off-grid meant that the post didn't make it but normal service is restored. I also have a couple of longer posts in the pipe to make up for it.
Links from this week:
- AI Model training without copyrighted content. Fairly trained is a non-profit that certifies AI models to validate they do not use copyrighted material. The organization announced this week that it has certified its first LLM. The model is small but useful in specific applications (legal and administrative tasks). It's great to see this is possible, but it will have at least a couple of consequences: 1) The idea that a system can't train on widely available cultural knowledge means these systems will be less aligned to broad humanity than we want (do we want that?), 2) the practice of paying for data sources will accelerate (see Reddit's deal with Google for example) which may mean open source models get further and further behind big tech. I'm 100% in favor of a thriving data/culture market emerging where artists can get rewarded for the use of their works in AI creation (and maybe some training). Still, if the ultimate outcome is to prevent the use of large swathes of human culture in the training step, we're inadvertently creating non-aligned AIs and handing power to people with deep pockets.
- Apple shows signs of life on AI. Apple has been working on AI for many years, but observers have recently commented that it has released little. This week, the company published an extensive paper and open-source models including detailed documentation of the training process. One might ask the question: why they would suddenly be so open? Stratechery's Ben Thompson captures the answer best: Apple's interest isn't to compete on AI in the medium term. It is to be good enough to make sure people don't platform hope and to commoditize AI in such a way that the Apple hardware platforms remain the main thing people pay for. This is very logical, but I strongly expect that this strategy will only go so far. AI interfaces are one thing that may end up commoditizing device hardware.
- High-paying jobs in the age of AI. Thanks to a friend for pointing out Noah Smith's long post on why there may be more abundant, well-paid work in a world of AI than we think. It's a thoughtful post centering on the idea that even if AI is better at everything, it will also have its highest and best uses (comparative value rather than competitive value). This means that it likely won't make sense to replace every last human occupation, and a huge abundance will be created. The comments are also well worth reading, with the main pushback being that compute costs may literally go to zero (unlikely in my view, at least in the next many hundred years).
- Neuralink's first human trial. The brain chip implant in a 29-year-old quadriplegic allows the patient to control on-screen cursors and applications such as games. As scary as the idea of thought-controlled devices is, it's obvious how much quality of life they can add to some. We're moving from an era of text input to voice input. At some point, it seems inevitable that it'll be through input.
- Mustafa Suleyman, DeepMind and Inflection Co-founder, joins Microsoft to lead Copilot. In the weird AI deal of the week... Mustafa Suleyman and most of the team at Inflection.AI where he was a co-founder will be moving to Microsoft. Inflection has raised nearly $1.5B in venture funding and set out to build a responsive, human-friendly interface AI that would become a co-pilot for your life. This AI co-pilot market position will be hotly contested, and it looks like $1.5B was not enough (or the thousands of NVIDIA H100s the company bought may have been underused). In the deal, none of the IP moves, but Microsoft is licensing usage of Inflection's API. This deal appears to give investors their money back. In any normal world, this surely would be an acquisition, but perhaps the US's current aggressive stance against M&A by big tech may now be incentivizing a new way to keep corporate lawyers busy.
Sadly, the great Sci-Fi Author Vernor Vinge died this week. His books are an amazing mix of weird and familiar. If you're looking for something "new" to read, I highly recommend the Zones of Thought books, starting with A Fire Upon the Deep.
