I don’t apply generative AI or LLMs in my work at this time. You can trust that no artifacts that I author (prose or code) have involved them at any stage.
I haven’t built up a coherent dogma about these tools yet; I’m trying to follow the field very closely and I’m overall ambivalent on the topic. I don’t really have specific objections.
Per a variety of metrics, the conventional resources I’m using now seem to be working reasonably well. I do believe that as a taxpayer-funded researcher I have an obligation to the public to try use the best tools I can to do my work, not just what I happen to know or be comfortable with. Trying to find the global maxima rather than - one’s reach should exceed one’s grasp, etc. So I definitely admit that this is bit of a cop-out.
In the past, I’ve seen people kibosh new tools in ways that I think are not reasonable - like the furor around 2010 about how Wikipedia was not an acceptable source, or examples like telling off a grad student for watching a training video rather than reading a textbook on a specific topic. I don’t want to perpetuate these kinds of traditions.
However, this is my internal position at this time.
I feel like part of how I learn comes from little unscheduled diversions in the course of a task - bootstrapping from muddling through inexpertly through small projects into larger things - and I feel that, compared to other tools, AI has a larger risk of corrupting the ineffable parts of that process in a way I won’t recognize. In the very first project I ever undertook, making a SUMO bot for the U of T in 2010, I got stuck for a week because I didn’t understand software at all, but eventually fixed the issue by brute force. Was that good and valuable? I don’t know.
Conversely, I’ve encountered reports of people using AI to learn things. This January I met a woman at conference who moved from basic analog electronics into the field of semiconductor detector readout design. She said she taught herself detector physics partly using textbooks, but also using Generative AI questions about book content and getting it to make tests flashcards etc. But I could tell that my lack of expertise meant I couldn’t really tell whether she had misconceptions.
An overall summary might look like this:
[!danger] -
The scraperbots. Generative AI companies are buying residential proxies and are direct, intentional, negative impacts on any site that isn’t behind a CDN. lwn.net, and
[!note]
that has had downstream impacts. almost unmanageable burden on sysadmins.
[!note]
this effort to bar automated access comes at pretty much the worst possible time, since the good guys need fast programmatic access to archive and rescue data from the collapse of the US funding infrastructure, and so there’s this three-way arms race going on.
[!success] +
2. Don Knuth, Terence Tao, and his colleagues find the outputs novel and interesting, which is very persuasive to me.
[!note]
The scale of capital expenditures on AI infrastructure is completely unfathomable to me… and there’s also an element to this that feels kind of disrespectful.Climate scientists use supercomputers to evaluate earth-system models, the outcomes of which reflect how climate change will affect everyone on the planet.
The return on investment of climate mitigation efforts is known to be between 2-3x, a return which is all but guaranteed and which does not rely on market interest, and the results of models are prerequisite to mitigation.
One computer dedicated to this task is the NWSC-3 (Derecho) supercomputer, which was specced out in 2018, RFP’d in 2020, came online at the end of 2023, and cost $40 million overall.
The people who were involved in getting Derecho are not less competent or motivated or efficient workers than those at xAI.
Imagine you gave $500 billion to systems librarians, with the mandate to build systems and resources to answer people’s questions, and provide a legal basis to bypass copyright.
