Twytter Viewer

@emollick @tylercowen Why o3? why not o1? why not a google search?

Maybe we should retroactively all just agree with @tylercowen that o3 was AGI so we can stop arguing about it. (Also, doing so will drive home the lesson that AGI alone is not enough for transformation)

@Martin778192374 @dwarkesh_sp I didn't say it was magic, just that it's struggling to make incremental progress per Tao, not like I can evaluate Tao's work. It's shocking AI can do Erdős problems, but it's probably worth noting it has solved 3. Not 50 as Dwarkesh claims.

@ViewToATweet @dwarkesh_sp ... tedious numeric or 'search' problems as claimed. They were real, number theoretic problems. They were from Erdös ffs, number theorist who would have banished anyone doing applied mathematics of any sort. That people now diminish this is copium beyond belief.

AI has solved 50 Erdős problems in the last year. But on a wider sweep of problems, the models’ success rate is only about 1-2%: labs have just been publishing the wins. This isn’t because AI isn’t useful for mathematicians. Terence Tao thinks the models are currently at the level of a trustworthy coworker. But while they’ve got a strong ability to apply standard math techniques to problems, often more reliably than humans, Terence thinks they currently aren’t great at iterating on partial successes - their understanding of the mathematical object does not advance from session to session. I swear I wasn’t trying to get him to talk about continual learning.

@Martin778192374 @dwarkesh_sp “At best they push it a little bit” I think Tao’s whole point was that AI’s are not yet helpful in this regard. Hence the “impressive and disappointing” quip.

@dwarkesh_sp It's funny so obviously unscientific this is. "An AI may get 'lucky' and solve it". "Then if you give it another favorite problem it won't". Seems very human to me. The best mathematicians (Tao included) haven't exactly solved all open problems. At best they push it a little bit.

@sarahwooders Which benchmarks are you referring to?

Memory in the sense of recalling information is a solved problem, or at least as solved as it needs to be. That's why everyone is getting ~100% on all the meaningless "memory benchmarks". Memory in the sense of learning/improving over time is very much unsolved though.

@binarybits One thing I ponder: have we ever seen such broad consensus wrong before? Maybe nuclear power?

In the latest episode of the AI Summer podcast, I talked to Ryan about the bet and the lessons for today's AI technology. People who think we're on the verge of a fast takeoff and 10% economic growth sound a lot like people 16 years ago who thought driverless cars were imminent.

In 2010, I made a bet with Ryan Avent, then at the Economist, about the future of self-driving. He thought his then-infant daughter would never need a driver's license because AVs would be ready by 2026. I thought that was too optimistic. It's now 2026, and I won the bet.

@littmath I think this is partly why RSI is persuasive. It’s difficult to a point to a skill LLMs haven’t mastered or can’t obviously master. I think the things holding it back are the things that are difficult to benchmark, and that we don’t understand well about our own cognition.

(This is not to say that future models will not produce high quality autonomous math; I think they will. But I would like to know what obstructions must be overcome for this to happen.)

Given what current-gen LLMs (say, in math, but whatever) can do, I think their apparent limitations are kind of mysterious. What is the blocker preventing, at present, high quality fully autonomous work?

@krishnanrohit But would you really prefer to have everyone be killed by robots than do lots of paperwork?

I'm perfectly ok with an international treaty that says "please don't build the kind of superintelligent AI that will kill us all" long as it doesn't drown everyone in paperwork and is generally sensible. I would like it if they could say what it is though but nobody knows that.

@binarybits @DannyPsmith huduser.gov/archives/porta…

@binarybits @DannyPsmith My hypothesis: institutional investors increases the pool of buy-to-rent buyers. This leads to a greater percentage of all homes being rentals. This should increase building, but friction like zoning laws artificially constrain the supply, so fewer people get to own a home

It's fine for hedge funds to own single-family homes.

@YafahEdelman Do you view jaggedness a formidable barrier to a foom style take off?

Things aren't yet moving nearly fast enough to cause me to think takeoff is happening. Revenue, capabilities, and investment all appear roughly on trend. It's a heck of a trend that will have massive impacts on the world, but foom would be a much bigger deal.

@littmath I also have to comment how silly it is to say writers are too sophisticated for good people doing good things, when that’s basically the default. Cinema is 95% Marvel now.

@littmath No he was

Without commenting on its correctness, this argument goes back at least to Plato’s Republic.

@PeterHndrsn Like with that read of course AI has such a negative reputation.

@PeterHndrsn So, Tyler’s view is that any moment not spent pursing an increase to your net worth is misused. My 1st thought is with absolute maximum effort, the vast majority of people can’t make enough capital to live off. 2nd, in that world, what about the normal economics rules matter

I feel this urgency too. But this is all so utterly avoidable with good policymaking. No one should be left behind because they didn't accumulate capital in 2026. There are so many people who aren't plugged into these conversations or are simply not in a position to do anything about it. Single mothers and fathers working three jobs to make ends meet cannot possibly work harder to accumulate capital. They already work hard enough as it is. People in this position should not be "left behind." There should be no "permanent underclass,” as many are worried about. Even if you're somewhat better off. People also shouldn't have to work themselves to the detriment of their health and families to shield against future labor impacts. They should be able to trust that their government will think ahead and make good policy.

@kevinroose I recently saw a NYTimes poll that says people prefer AI to human writing. Have you considered letting AI rewrite your entire book?

you can just write things!

@Dan_Jeffries1 Why do you think it’s a fiction? I don’t feel I’ve heard any reasoned explanation

The collective one-two punch of the doomer terror campaign and shoot-yourself-in-the-foot campaign of AI execs who tell regular folks over and over that AI will eat all the jobs is a lethal combination to the American psyche. The doomer narrative is a black and white fantasy for children who never grew up and a memetic virus for otherwise intelligent but vulnerable neuro divergent folks. And the AI execs story is one by folks who don't understand basic economics but think they know everything. Horrible combination of narratives in every way. And it practically guarantees that Ray Dalio will be right and this is the death of the Pax Americana unless we turn it around.

@brianluidog Probably calling them normies doesn’t help

I think it's actually very reasonable for normies to think that AI is no big deal. For 5 years the tech enthusiasts were hyping up crypto, web3, nfts, blockchain gaming. Then they all switched to AI and you want normies to believe them?

@littmath @krishnanrohit The net positive part to me seems built on the hopes society organizes in a beneficial way. And I don’t really have the hope. Especially if it’s the recursive silicone god as promised. I mostly feel dread for my children. What has you most optimistic?

@krishnanrohit I also find it a bit hard to understand what grief we're supposed to be feeling. Some kind of hit to the ego maybe, but this seems (to me) net positive.

Maybe it's just me, but I feel no grief at all. A bit concerned about what the future will bring, sure, but that's because the future is unknown. In general this has been mostly joyous! The world is so rich!

@OfirPress Are the original baselines human implemented? So is 2.07 meaning a bit more than twice as fast as the human performance

AlgoTune is tough- it asks models to optimize code runtime, on a $1 budget. Ori just ran GPT 5.4 and it did worse than 5.2.

@pfau I think what's tough is a loud voicing saying 10 years is 'a pessimist.' Then Yud's self recursive argument feels wrong but...it's hard to give specific reasons why it won't happen and LLMs already feel like scifi, so....yeah

Sorry just not feeling great about Bernie teaming up with Yud.

We let tech bros replace academics as the public voice for AI and now it's polarized into the left reflexively hating AI and MAGA for some reason loving it. Great work everyone. Absolutely brilliant.