François CholletVerified account

I saw @fchollet say sth similar. I _kinda_ agree. But just cuz I can spend next 10y applying convnets to probs I don't care abt doesn't (to me) imply summer. For me, winter is when funders (ind&gov&c) say "nah the cost/benefit for basic tech is too high, we have what we need."

(on cost/benefit: the infamous ALPAC report didn't destroy machine translation funding because it declared that MT was impossible... but rather, because, for what the gov't needed at the time, they could achieve more cheaply by just hiring human translators.)

Thanks for your breakdown on how you see AI summer/winter! My focus on benefit to funders ("if AI research froze right now") is as an AI winter for them essentially guarantees an AI winter for research. The funders are already seeing vast benefit that imho is unlikely to slow.

If it's AI summer for funders, it'll remain AI summer for researchers in proportion to the cost/benefit that further research brings. This is great as there are a set of trends within the AI ecosystem that will likely help ensure the flow of benefit from research to funders:

- The ML stack is becoming more standardized + shared, .: a strong summer in one domain (vision) still heats winter domains (NLP) - Teams are becoming better at the research ⇒ production gap = funder benefit - Prod scales hardware, research benefit seems to grow with scale

Trying to summarize w/o a billion bullet points: as the ML stack reaches more applications (+funders), the better the problem domains get at recording / processing their data in an ML friendly way (+funders, +researchers) which generally provides better grounding for +research.

We largely agree :). To pick nits (what's twitter for?! :P) I think that w/ standardization, "AI" becomes "default" and ceases to be seen as "AI" anymore (history suggests this is likely), and so funders find value in "import pytorch++" but that's no longer seen as "AI" -> winter

so it'll be just winter...