Recently, major AI industry players (incl. a16z, Meta, & OpenAI’s Greg Brockman) announced >$100M in spending on pro-AI super PACs. This is an attempt to copy a wildly successful strategy from the crypto industry, to intimidate politicians away from pursuing AI regulations.🧵

Kind feel like there were pretty similar steps in improvement for each of: GPT2 -> GPT3, GPT3 -> GPT4, and GPT4 -> GPT5. It’s just that most of the GPT4 -> GPT5 improvement was already realized by o3, and the step from there to GPT5 wasn’t that big

GPT-5 didn’t live up to OpenAI’s hype, but it is *exactly* in line with extrapolations from prior AI advancements. Go ahead and discount future statements from OpenAI/Altman, but you should still expect the fast AI progress that we’ve been seeing to continue

If you’ve been predicting AI was about to hit a wall for years… and then progress continues to be exponential… you can’t really claim a win here. Sure, people predicting it was super-exponential also can’t claim a win, but, uhhhh

New SOTA results from Opus

Kinda weird that the same people who argue the most mild of AI reporting or transparency requirements will “enable China to outcompete the US in AI” also argue restrictive GPU export controls will just drive China to invent highly efficient algorithms to compensate

Agree with Garrison here. It’s probably the case that, if humanity sticks around for millennia, we’re ~destined to build AGI eventually. But there’s no reason humanity is sure to build it ~as soon as we can, and those saying so are trying to create a self-fulling prophecy

How much of industry opposition to the Chip Security Act is b/c reqs are “burdensome” to implement… and how much is that semiconductor companies want to be able to maintain plausible deniability about where their chips end up, so they can keep indirectly selling to China?

Looks like from a mental health perspective, you want to make sure to do at least ~2hr/wk of light exercise (eg jogging) or ~1hr/wk of vigorous exercise, and there’s not much benefit to going beyond that.

Coal baron arguing that instead of giving everyone UBI like he had previously promised, “what if we just gave them each a free steam engine?”

Three-sentence horror story: Your car swerves off the road and crashes. You wake up, floating on a cloud, and think, “Is this heaven?” You hear, “it’s not just heaven—it’s a personalized utopia!”

Woah Zuckerberg has a $300M estate in Hawaii?! That’s like 1/3 the cost of an AI researcher!

It’s been an entire 5 weeks - I think we need an update to this chart

Surprised and pleased to see this - Dario (Anthropic CEO) hints that he’s against the proposed 10 year state-level AI regulation ban:

Many claim AGI timelines are 2030 or bust, b/c compute scaling will hit production limits in ~2030. These limits exist, but I think the point is overstated. First, algorithmic efficiency improvements have been large & can allow for effective scale up. Second, “unhobblings” could take multiple years.

Guy on the right is sad he can no longer use a computer, wishes he had spent more time working on his b2b SaaS startup when he was younger

Every now and then I think of this quote from AI risk skeptic Yann LeCun

I feel like this is an unreasonable expectation for the product

Has anyone done a deep dive on the “actually ice cream is healthy” research results? What’s going on there?

[e/acc voice] “at long last, we’ve built ‘it’ from hit book ‘If Anyone Builds It, Everyone Dies’”

Incredible prediction market

Asking the AI for advice on how to socialize at parties, following whatever it says without any reflection and guided only by intuition - call that vibe vibing

Seems like a lot of people are getting annoyed at reward hacking and sycophancy from LLMs, to the point that they’re switching away from models that engage in those behaviors more. Man, if market competition actually does provide a race to the top on alignment…

Google’s AI summaries have gotten so much better. So of course people have stopped talking about it

Person who doesn’t entertain the simulation hypothesis because he doesn’t want to break the fourth wall

[EA who lives in a group house with a bunch of other EAs] “well… I guess you could say I’m EA-adjacent (my room is adjacent to the rooms of several other EAs)”

One thing a lot of people missed from AI 2027: Yes, the pause was necessary for surviving. But it only worked because it was ~perfectly timed - after ~AGI but before ASI. If it had instead been earlier, America would have ceded its lead for nothing and we would have all died.

“This sounds like scifi” is such a weird criticism of attempts to forecast effects from much more advanced technology. “This sounds like a story from a time that has substantially better technology than what recent history had” yeah okay

Still wild that telling the AI “use LW logic and give me your best guess” makes it better calibrated and gets it to bullshit less

MYTH: AGI will be like The Terminator Reality: it’ll be much more like Terminator 2: Judgement Day

New report: “Will AI R&D Automation Cause a Software Intelligence Explosion?”  As AI R&D is automated, AI progress may dramatically accelerate. Skeptics counter hardware stock can only grow so fast. But what if software advances alone can sustain acceleration? 🧵 www.forethought.org/research/wil...

AI won’t steal your job, someone using AI will steal your job - and then soon after AI will steal *that* person’s job

Here's what happens if you extrapolate the figure further

Anthropic researchers, trying to figure out why Manus is so good

If you want to get people to seriously engage with an idea outside their current worldview (eg “AGI could take over”), I think you need to both: • actually talk about that idea, instead of some adjacent idea • speak to them in their language, meeting them where they are now

Broke: preparing for AGI by buying TSMC & NVIDIA stock and shorting bonds Woke: preparing for AGI by getting Pakistani citizenship

Thinking about the homeostatic effect in politics, I guess one reason for it is: • if you solve the problem you championed, voters stop thinking it’s a problem anymore so it’s not a salient issue • if you fail to solve it, those who supported you over it grow disillusioned

At some point, some AI company is going to accidentally call their RSP a name that’s already been taken by another AI company. Imagine how embarrassing that’ll be!

Who called it “model weight self-exfiltration” when they could have called it “outside the box thinking”

[guy riding on Theseus’s ship] “I’m getting a lot of Theseus’s ship vibes from this ship”

The people who are not particularly surprised by recent developments in AI are by and large saying AGI could be soon. The people confidently saying AGI won’t be soon are the same people who kept being very surprised by advancements over the past years.

Interesting passage from Dario’s recent piece:

Reminder that during the Space Race, we had lots of engineers working on making sure the rockets would be safe, wouldn’t blow up, & could be steered. If we had instead only focused on “rocket acceleration” while ignoring “rocket safety”, we never would have been first to the moon

Hot take, but if the narrative from NYT et al had not been “lol you don’t need that many chips to train AI systems” but instead “Apparently AI is *not* hitting a wall”, then the AI chip stocks would have risen instead of fallen

Colombia is a major exporter of flowers, so tariffs mean we should expect flower prices to rise. Not financial advice, but now might be a really good time to speculate on some tulips

Splitting a meal with some friends in 2007 Zimbabwe:

People keep suggesting UBI as the solution to AGI-led unemployment. I don’t think society will accept that, and I suspect a job guarantee + make-work is more likely

Type of guy who one-boxes but picks box A

Good piece from Garrison Lovely arguing that, contra MSM claims, AI progress isn’t stalling but instead becoming invisible. Chatbots haven’t changed much, & that’s what the public sees. Meanwhile, AI has gotten much better at things like STEM research, which is a huge deal

Training a flexible, general-purpose reasoner that can succeed despite unexpected obstacles seems pretty hard. Worryingly, training a flexible, general purpose reasoner that can succeed despite unexpected obstacles *except when those obstacles are humans trying to stop it succeeding* seems harder.