Saturday Links: Skills, Job Impact, and Adult ChatGPT

This week, stunningly, researchers optimized simplex and shortest-path route finding, which have been considered effectively near optimal for many years. In other news, China restricted rare-earth exports, which could severely affect global chip and electronics production.

Equipping agents for the real world with Agent Skills. Anthropic’s launch of skills got a fair bit of coverage this week. Skills are essentially a way to encapsulate task-specific knowledge in a bundle that can be used by Claude for a task, potentially also with the ability to access external resources. Other companies in the space are doing similar things, but then calling this an “agent”. You can also build “agents” with Claude. Claude Code is arguably one such agent, but here Anthropic is choosing to add a different label. My guess is that the aim is that they feel that, as humans, we’ll want to interact with one or just a few “agents” that can engage in many tasks. This can seem a lot friendlier than talking to many “Agents,” and my guess is that the pattern will catch on, so we’ll have both Agents and skills. Simon Willison has a nice overview of skills in this post.Italian news publishers demand investigation into Google’s AI Overviews. The drop in internet traffic to news sites and publications due to AI is really starting to bite, and this is one of numerous investigations and lawsuits. Ultimately, though, there’s really little logic that could compel the world to return to the pre-AI era. AI summaries, Perplexity searches, and ChatGPT answers are, in many cases, simply much more useful to users, and so click-throughs will drop. Forcing Google to remove them will simply push more people to ChatGPT and not click through either. Publishers are rightly upset that some of their content may have been used to inform the answer, but I suspect what this will boil down to is Google saying, “ok, set your robots.txt file to exclude Google, and we won’t do that,” which would also cause them to drop out of search. I’ve mentioned in previous newsletters that Google should separate its AI crawler from its Web crawler. However, it’s not clear whether it would help publishers in the long run. All this will do is cause their link not to be mentioned in the AI summary either. If users start to rely on that more, their traffic will still drop. The only real forward path is into content licensing deals or marketplaces. Sadly, larger players (like the New York Times) are far better placed to do this than small publishers. The aggregation of content that AI provides is so convenient for many queries that the diversity of content is likely to drop. One can only hope that a countervailing force develops a smart content discovery system to get really good at pulling out niche, relevant content that people want (and a way to monetize it).OpenAI will allow verified adults to use ChatGPT to generate erotic content. OpenAI has been on a tear in recent weeks, releasing new products that put pressure on almost every major tech company in one way or another (social networking, agent frameworks, tools for work, etc.). This week, the company added another business area to the list: erotic content. While I guess there is a libertarian argument to say “why not,” this is a head scratcher. Going in this direction does seem like it could risk adoption in other areas. It will also mean that OpenAI has to put very careful controls into models (maybe even have separate models) to avoid adult themes cropping up in work environments. Trying to cover too much of human activity in one product too fast might just be the undoing of much other good work.Evaluating the Impact of AI on the Labor Market: Current State of Affairs. A new widely quoted study seems to indicate that AI replacement has not had a significant impact on the labour force to date. The study finds that the occupational mix (an index based on changing occupational roles and job mobility) is changing quickly, but there is no indication that AI is driving this. The US currently has relatively low unemployment, so this seems consistent with the study. On the other hand, it seems unlikely this situation will persist. AI adoption by individuals has been rapid, and AI is being used across many knowledge work roles. What is taking much longer is company adoption of AI. As this happens, there is a strong likelihood that personal productivity boost will turn into corporate productivity boosts. Some of these will lead to “doing more with the same team”, others risk leading to team downsizing to produce the same output with fewer people. What remains unclear is if new companies will spring up to compete and keep employment high. Annecdotally, I don’t know a single person working in knowledge work roles who hasn’t seen some impact of AI in their workplace. In some areas (video games, art, marketing, sales), the pressure on available jobs has been much more negative than positive so far. Qualcomm acquires Arduino to supercharge the global maker movement. Arduino boards have been the basis for many amazing small electronic devices. The project has travelled a long road and garnered many fans along the way. Buying the company looks like a smart move for Qualcomm, which is betting heavily on a new era of embedded electronics. They also recently acquired the quantized model maker Edge Impulse. It seems highly likely that we’ll see more and more capable small AI models running on AI models and feeding data to more complex cloud-based AI. Entrenching itself from a chip maker into a compute platform provider for edge devices seems likely to play out well for Qualcomm. 

Have a great weekend, and a regular reminder to avoid using AI to defend yourself if you are in court.