Chat wtf is this curve?

Gary asks the doomers, are you “feeling the agi” now kids?

To which Daniel K, our favorite guru lets us know that he has officially moved his goal posts updated his timeline so now the robogod doesnt wipe us out until the year of our lorde 2029.

It takes a big brain superforecaster to have to admit your four month old rapture prophecy was already off by at least 2 years omegalul

Also, love: updating towards my teammate (lmaou) who cowrote the manifesto but is now saying he never believed it. “The forecasts that don’t come true were just pranks bro, check my manifold score bro, im def capable of future sight, trust”

it’s weird that Piper keeps getting paid to make content, but I’ve never once seen anyone claim to enjoy any of her work…

🍿

Reenforcement learning

Me and the gals having a totally normal evening at SBFs compound

(Bonus)

Only taste tester I trust dropped his verdict

The part about them sitting around crying made me big sad :(

The one big cope I’m seeing is in the METR graph ofc. Tiny bump with massive error bars above Grok 4 so they can claim the exponential is continuing while the models stagnate in all material ways.

Yeah, O3 (the model that was RL’d to a crisp and hallucinated like crazy) was very strong on math coding benchmarks. GPT5 (I guess without tools/extra compute?) is worse. Nevertheless…

Well, after 2.5 years and hundreds of billions of dollars burned, we finally have GPT-5. Kind of feels like a make or break moment for the good folks at OAI~~! With the eyes of the world on their lil presentation this morning, everyone could feel the stakes: they needed something that would blow our minds. We finally get to see what a super intelligence looks like! Show us your best cherry picked benchmark Sloppenheimer!

Graphic design is my PASSION. Good thing the entirety of the world’s economy is not being held up by cranking out a few more points on SWE bench right???

Ok. what about ARC? Surely ya’ll got a new high to prove the AGI mission was progressing right??

Oh my fucking God. They actually have lost the lead to fucking Grok. For my sanity I didn’t watch the live stream, but curiously, they left the ARC results out of their presentation. Even though they gave Francois access early to test. Kind of like they knew this looks really bad and underwhelming.

Guh wtf

cheers m8, ill drink to that

I’m ignorant- give me the lore drop.

Perfecting the art of getting sloshed is my 80,000 hours of meaningful work.

A nice long essay by Freddie deBoer for our holiday week: the release of GPT-5; I wholly recommend reading the whole thing!

https://freddiedeboer.substack.com/p/the-rage-of-the-ai-guy

Choice snippet to whet your appetites:

“With all of this, I’m only asking you to observe the world around you and report back on whether revolutionary change has in fact happened. I understand, we are still very early in the history of LLMs. Maybe they’ll actually change the world, the way they’re projected to. But, look, within a quarter-century of the automobile becoming available as a mass consumer technology, its adoption had utterly changed the lived environment of the United States. You only had to walk outside to see the changes they had wrought. So too with electrification: if you went to the top of a hill overlooking a town at night pre-electrification, then went again after that town electrified, you’d see the immensity of that change with your own two eyes. Compare the maternal death rate in 1800 with the maternal death rate in 2000 and you will see what epoch-changing technological advance looks like. Consider how slowly the news of King William IV’s death spread throughout the world in 1837 and then look at how quickly the news of his successor Queen Victoria’s death spread in 1901, to see truly remarkable change via technology. AI chatbots and shitty clickbait videos choking the social internet do not rate in that context, I’m sorry. I will be impressed with the changes wrought by the supposed AI era when you can show me those changes rather than telling me that they’re going to happen. Show me. Show me!”

Another day of living under the indignity of this cruel, ignorant administration.

💯

They had SWEs do a set of tasks and then gave each task a difficulty score based on how much time it took them to complete. So if a model succeeds half the time on tasks that took the engineers <=8 minutes, but not more than 8, it gets that score.