MIT Technology Review

We're on those other social media platforms, too. 🟦☁️: bsky.app/profile/techno… 🧵: @technologyreview" target="_blank" rel="nofollow noopener">threads.net/@technologyrev…

Watch a subscriber-only discussion exploring how AI might enter the physical world. trib.al/EW19P9I

As tools like Claude Code get better, more and more developers are happy to hand off coding tasks to them. The way software gets built has changed for good. trib.al/RJi2Un4

Plus: SpaceX has filed for an IPO expected to be the largest ever. trib.al/slw1IZC

Less focus on decarbonization, more on supply chains. trib.al/HWpdH8G

In a lawsuit, the Coalition for Independent Technology Research argues the government is using immigration policy to stifle free speech and tech regulation. trib.al/aSoj2po

Well @demishassabis if you ever want to go deep on AI and scientific progress, hit us up!

Ok, very interesting thoughts on scientific progress at the end now. A big motivation for him was looking at the stagnation in progress in physics after the mid-19th century. He had a sense we might be reaching the limitations of the human mind.

One thing Demis asks himself—would he still be working on AI if it hadn't had this explosive moment in 2022? Absolutely yes, he says—it's an obsession for him. More than anything else he's said today, that makes him sound like a scientist.

Says continual learning (i.e. AI that keeps learning after deployment) is one thing DeepMind is focusing on now. Definitely something to watch.

What are the big things to look at going forward? Demis says self-improvement. Coding is a small example of that, but what next? He (very correctly) highlights that coding is a very specialized area that's amenable to AI acceleration. Not clear what he thinks the next SI step is.

Demis says that the pace of science is picking up in all sorts of fields (implicitly due to AI), but there's a lot of research that science has slowed recently. IMO, whether and how AI will change the rate of scientific progress is a big open question.

That database is available for free to anyone and has millions of users. Probably the ultimate example of virtuous AI—an interesting thing to highlight.

Now showing a clip from the documentary The Thinking Game where John Jumper raises the idea to Demis that AlphaFold could just predict (and release) the structure of all know proteins. His response: "why don't we just do that?"

And immediately pivots to AI for science—the virtuous side of AI. My subjective take is that he thinks job loss is very possible but doesn't want to potentiate the backlash.

He won't say clearly whether his hunch is net more jobs or fewer jobs, but he "doesn't see why" there wouldn't be more opportunities. "There's going to be changes, but overall, it will be better."

Demis says he doesn't agree with his peers about this. It's super uncertain whether AI creates more jobs, destroys jobs—the future is yet to be written, he says. (So the implication here seems to be that he's blaming rhetoric from Sam/Dario/etc. for the backlash?)

Allen digging in about the AI backlash now. Interested to hear Demis's thoughts on this.

Demis's test for AGI, which I admittedly kind of like in principle, is whether a system with a knowledge cutoff of 1901 could re-discover the major physics advancements Einstein made in 1905. In reality, this would never work with current AI—not enough training data before 1901.

His current estimated AGI date is ~2030

What's making him feel that way? First think he mentions is his experience working with Google's own coding tools. Nice media training there.