At the same venue, nerdnite, as the last talk where I commented on being surprised by the AI-generated imagery not being half bad, a transplant surgeon gave a long talk, which exclusively used image model for basically all the slides, in a very narrative-driven way, generating a slideshow telling a story about an anonymized patient’s progression of disease.
This guy was super, super competent and I’m sure I would have enjoyed the talk no matter what.
I think the AI genuinely contributed something to his talk. His competence seemed to extend into knowing how to use this in a classy way. Not revulsed by this, I enjoyed the talk.
Talk at work by someone who had very little time to create a presentation, and so it was fairly clear that he leaned heavily into it. Very overworked dude, maybe could have improved time management a little and that would have improved the outcome a lot. but I don’t blame him, he has so many responsibilities to keep track of.
Anyway, this wasn’t very classy. Also, some things he said seemed to be guided by AI, and briefly verged into pseudoscentific. I think use of AI actively made this talk much worse than it could have been, like the “write what you know” guy above.
Had a fault on a pump controller.System Error 7. Troubleshooting tree in vendor manual said “contact vendor”
Someone googled this and the AI said “output short circuit”.
pointed at
then it shapeshifted, when I googled it in another tab same errorcode, DSP. tried again with slightly more different lhrasing
so it really caused unnecessary confusion and led on the wrong trqck.
The steps that it suggested for troubleshooting included “disconnect cable from pump and power up”.
It would be impossible to do so and it would be lethal.
I think the most intereting factoid of the whole Mythos thing is that the most recent High severity bug in curl was in 2023, even after all the AI scanning. that is, it’s still totally possible to have secure software if you follow a fairly airtight validation process like the creator uses.
in some students I was working with sent over their infromation and I was doing a peer review. it’s AI generated but very very sophisticated and they did a really really good job winnowing and isolating, getting rid of any false statements or disjoint stuff. There is one big issue which makes the project a non-starter which could probably have been identified earlier, but otherwise it’s self-consistent and good.
On the other hand, it took them a few weeks to do this winnowing process on the output of the AI and get a quality product. With my access to the university library, I was able to find some books on the topic, and I think I got a little less than parity with their work in about 15 hours of reading and writing.
This is encouraging to me, because it suggests that there’s not a huge factor of being “left behind”. Sure, the AI is much smarter than me, but it’s not incomprehensibly so - it can’t be that smart if I can catch up so quickly. And it takes very little time to write a synthesis, but it it
The biggest benefit of AI to the students I think is just popping those paywalls. One of the key books I needed to access was from CSIRO Publishing, $huge and wasn’t even from the l. I actually needed to get an ILL Request from another library.