"This is a massive shifting of the goalposts in “AI” defense. [...] now we’re just full speed ahead into slop-world. What if Fast Fashion but for everything?"
(Article title: On Moving Goalposts)
"This is a massive shifting of the goalposts in “AI” defense. [...] now we’re just full speed ahead into slop-world. What if Fast Fashion but for everything?"
(Article title: On Moving Goalposts)
@tante Simultaneously I'm seeing LOTS of people angrily insisting that it is fallacious to conclude that just because AI was used and then there was an outage that AI was responsible for the outage.
So we know AI was used, we know AI users are in a slippery slope of progressively less oversight over the AI with time, and we know there are surprising, surprisingly long outages at the org. And we lack insight into the org's inner working, so we can't do *better* than assuming. But DON'T ASSUME!
@tante Code can now be generated quickly. Human review is still slow. Answer: get rid of the human review!
🥴
Gawd, I hate this timeline so bad.
@tante really good article and spot on. Do you have any constructive proposal for how to get out of this situation? Would like to discuss. As I see it, short to medium term, companies and software projects which reject AI coding and review will not be competitive and might end up dying out. Rather easy to predict. Longer term, the picture changes, but it's unclear where it will lead. Even if we assume that AI coding and software design abilities will get better and better, the arguments regarding the impact on people, their cognitive abilities and societies remains, and of course, last but not least, the centralization of power into the hands of just few. Apart from pure rejection of AI and retraction into AI free zones/communities (or violent breakdown which we should avoid), there is just one method I know that is capable of reigning in AI, and that's AI itself. Use three agents to keep one under control. Fight fire with fire. Not a great solution,
just the only thing I tried that actually worked.
@bug TBH: I don't think anti-AI projects will "die", quite the opposite. LLM drive by contributions don't create community, shared values and shared knowledge.
Open source projects don't always have to compete. I don't think that that is a big problem
@tante agreed for many open source projects, but that's not neccessarily what I meant. I do think that important security fixes from LLMs should be accepted into popular projects after human review, though. Leaving a LLM discoverable security flaw in PHP simply because the discovery and fix were from an LLM would have been the wrong choice. So, I guess it requires a bit of nuance there. Accepting sloppily written feature contributions in is a different thing than point fixes or well (but LLM written) designed and documented concise code contributions. For me personally, the boundary boils down to whether I care intellectually about something. Do I need to understand in detail how the user interface code of a complex visualization tool works when I'm trying to visualize attention patterns and gradient behavior inside a biomedical AI model? My personal intellectual attention in that moment is elsewhere, the AI written visualization code is just a discardable tool. The problem is, the line is hard to draw.
@bug And "Agents keep agents under control" might stochastically work to a degree most of the time but given the other issues with LLMs I fund that a morally dubious solution.
@tante agreed, it's, as a colleague of mine put it, "a deal with the devil". At the same time, it can be perceived as the only option. That's the problem, at least if you actually are in a competitive situation (Startup, say) and others do use the tools. You are not really free to reject them. Similar to autonomous weapon systems, they can be forced on us by competition or adversaries, no matter how questionable, simply because the alternative is worse. That's what I meant by constructively getting out of the situation. Another aspect are the actual benefits. If the devil offers you a cure against cancer, is it morally acceptable to reject? I'm working in biomedical AI R&D, speeding up that entire field sustainably (including our competition) would be good for humanity. At the same time, we may never lose comprehension of the core scientific aspects of it. We might speed up development of user interfaces of tools we use through agents, but core scientific integrity and human peer review cannot be negotiable
@tante 🎵 She swallowed the spider to catch the fly / I don’t know why she swallowed the fly / Perhaps she’ll die 🎶
@tante Primitivism is futile.
The technology exists and will be used.
@tante Could not agree more. Thanks for writing it up.
@tante I'm seeing people brag about their Software Factory as though it was a good thing, and I weep