Weekly Links: AI and Weapons, Jobs, and Abstractions

This week, Uber previews Air Taxis in Dubai, Nano Banana 2 arrives, and Perplexity launches its own agents. Sadly, out in the real world, another military attack is also happening.

On to the longer stories:

• Anthropic boss rejects Pentagon demand to drop AI safeguards. Anthropic's terms of use explicitly prohibit the use of its model for population surveillance and fully automated weapons, leading to a very public fight between Anthropic CEO Dario Amodei and US government officials. Later, in a post on Truth Social, US President Trump said he was directing all US Government Agencies not to use Anthropic's technology. Not only would this be a huge loss for Anthropic, but it would also likely be a huge loss for many Federal Agencies. OpenAI's Sam Altman said in a statement to employees that the company shares Anthropic's usage redlines, yet somehow has not been caught in the President's ire. In the short term, this might seem like a win for OpenAI, but the situation seems very fluid. It seems unlikely that the US government could sustainably ban the use of all frontier models from all operations. A counterargument is whether companies can expect to sell technology to military customers and retain control over its use. The real question is just whether model labs will be able to hold their red line on usage, or whether this would ultimately result in appropriation of the technology in some form. For military equipment, the US military would be able to choose suppliers that do not insist on such clauses. For AI, it isn't clear whether such suppliers will exist.
• We're making @blocks smaller today. In an announcement on X.com earlier this week, the CEO of Block Inc (the publicly listed payments provider) announced that it would be reducing its staff from 10,000 to 6,000 in response to accelerating efficiencies from AI. In his post, he outlined his thinking that such reductions in force were an inevitability, and it was better to carry them out all at once rather than executing them as a drawn-out set of adjustments over the next few years. The layoffs led to a lot of media commentary, and there is plenty of other job displacement commentary in the air. Square is a highly technical company, and hence it is perhaps unsurprising that change happens at such organizations first. However, it seems clear that significant job losses will occur across white-collar sectors over the coming years. Society really does have to prepare for this.
• The accidental hacker: how one man gained control of 7,000 robots. You may want to check your DJI Romo home-cleaning robot and patch it. Earlier this week a Spanish engineer realized that his cleaning robots server communications were insecure enough to allow him to see camera feeds, maps and other information from devices all over the world.
• The Persona Selection Model: Why AI Assistants might Behave like Humans. Anthropic has an interesting post on its blog about its research on instilling personas into its Claude series of models. Users already experience Claude as having a certain personality, and OpenAI allows users to customize the tone of their ChatGPT model. This work goes further and digs into how such training could be done deeper in the model. The big takeaway is that by training personas more deeply into models, it is likely to make interacting with them in an anthropomorphised way more effective. It seems to me that, as humans, we will get used to the characters of our AI systems just as we'll get used to their capabilities. Almost all our evolutionary apparatus is designed to interact with other humans; we can't help but apply that to these new human-like interactions. What seems key is keeping the model's underlying actions consistent with the persona it displays.
• A neural network for modeling human concept formation, understanding, and communication. I think one of the key things missing in current LLM architectures is the formation of better abstractions and higher-level concepts. It seems that this currently only forms when they are also explicit in training data, and LLMs are getting better at relating the abstract to the concrete. To actually form new associations, though, more work needs to be done. This paper makes some progress towards abstraction learning by layering in an explicit abstraction learning component to training. The early results show the ability to (and the utility of) learning parallel, highly compressed representations of multiple concepts. There's still a long way to go since humans often infer such generalizations in one shot, but interesting progress.

Wishing you a safe and productive week. Hoping for peace in the various conflicts around the world.