Donti: GenAI works when generation is hard but verification is easy
Priya Donti · Live: How AI could supercharge climate tech · Apr 12, 2023 · at 35:45
Priya Donti (Climate Change AI co-founder and MIT professor) outlines an evaluation heuristic for LLM deployment in critical infrastructure and climate applications.
0:00 / 0:45exact quote · 45.2s
720p mp4 · rendered on demand · StarZero watermark
“As maybe thinking back to this barometer to evaluate whether an application kind of makes sense, I think a lot of these make sense because they satisfy this key property, which is something that is hard to do, but easy to verify. So ChatGPT and other generative models, what they're really doing is Kind of learning some big probability distribution over text or over images in the case of image models, and they're drawing from it. And so what that means is that there isn't necessarily kind of truth inherent in that or accuracy inherent in that, and people who have played with these models will have seen, right, there are lots of inaccuracies, there are lots of challenges with these outputs, and this is a fundamental challenge in the models. These models don't know how confident they are in an answer, how likely it is to be correct.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Priya Donti
Donti: Conditioned on AI or not, 'P-Doom' is high for power grids
“Conditioned or not conditioned on AI, P-Doom is high for the grid, right? We are dealing with climate change and extreme events. We are dealing with a lack of, you know, planning for these things as we saw in Texas, for example. So I don't know that AI increas…”
Donti: Generative AI Hype Risks Triggering Another AI Winter
“And with technologies like generative AI, there's a lot of, you know, flashiness and sort of jumping on the hype train, but I do sincerely worry that this will lead to an AI winter if a lot of those, you know, if we see a lot of, you know, hype and unscrupulou…”
Donti: Average-Case ML Fails on Grids Without Physical Boundary Guardrails
“The fact that you can get a machine learning algorithm that learns to do something nuanced from data and does it right most of the time doesn't help you in those times when it does something wrong that one time that really blacks out your grid. So this is wher…”
Donti: Utilities lack regulated rate of return incentives for software innovation
“Utilities and system operators can receive a regulated rate of return, some percentage of the money back on kind of investments in things like, you know, wires in the ground, but we don't currently see that for software innovation, and we don't currently see t…”
Donti: Attribution Science Galvanized Public Climate Conversation
“And I actually attribute a lot of that to
improvements in attribution science, right? When we were able to really start saying this specific event is due to climate change, I think that really kind of galvanized the public conversation.”
Donti: Open Climate Fix cut National Grid forecasting error by 33-50%
“For example, the non-profit Open Climate Fix, they used AI and machine learning in combination, you know, with satellite imagery, historical data, to cut the error of the forecast that National Grid, the UK power system operator, is creating by a third to a ha…”