this has been one of my helicopter moments too: astralcodexten.com/p/testing-ais-…
@sama It's pretty wild x.com/emollick/statu…
I think o3 can infer frequency/resonant harmonics in the tonal range if any given image's color palette. and it seems to make correllations between these harmonics to identify things...or at least to identify locations. total speculation based on observstion, but it does seem to be the case.
@sama Testing is always a helicopter moment 😋
@sama just wait for PIGEON-2... 🐦 @lkshaas @silasalberti @michalskreta @jmichaelkrause x.com/lkshaas/status…
@sama 4o and even gpt4 were pretty good at this, too! x.com/menhguin/statu…
@sama 4o and even gpt4 were pretty good at this, too! x.com/menhguin/statu…
@sama this would be a nice mini environment for tool use +rl/grpo training. Can even have a nice continuous loss based on the lat/long error
@sama Helicopter moment? Like a bird’s eye view, film style - seeing the bigger picture?
@sama Haha yeah, been there done that bro. Smooth moves tho 👀
“Everything is physically impossible when you’re 800 IQ points too dumb to figure it out. A chimp might feel secure that humans couldn’t reach him if he climbed a tree; he could never predict arrows, ladders, chainsaws, or helicopters. What superintelligent strategies lie as far outside our solution set as “use a helicopter” is outside a chimp’s?”
@sama absolute banger that it can do this really shows the intelligence of how far we have come and the reasoning too wow, i was blown away by this when i tried it... whats it say about this people?
@sama definitely had a few of those lately. wild how ai can just drop you anywhere and still know exactly where it is.
@sama There is even a benchmark for this: geobench.org made by @ccmdi0
@sama Fascinating breakdown — GeoGuessr is the perfect testbed for real-world reasoning, visual pattern recognition, and even a hint of intuition. If an AI can master that, it's inching closer to human-like generalization.
TLDR: In “Testing AI’s GeoGuessr Genius,” Scott Alexander explores the remarkable capabilities of OpenAI’s o3 model in the game GeoGuessr, where players deduce locations from images. Using a detailed prompt, o3 accurately identified diverse and obscure locations, including a remote beach in California and a high-altitude site in Nepal, based solely on visual cues. While the AI demonstrated impressive pattern recognition, it struggled with indoor scenes and ambiguous images, highlighting its limitations. Alexander reflects on whether such AI achievements signify a leap akin to a chimpanzee encountering a helicopter—suggesting that superintelligent strategies might be beyond human comprehension—or if they represent advanced yet understandable human-like reasoning. Ultimately, the article underscores the AI’s extraordinary abilities while acknowledging the boundaries of its current understanding.
@sama vpn can help throw off the scent, but holy shit
@sama This one seems pretty expected to me. The ability to correlate lots of data doesn't sound impossible. In fact it sounds like exactly what AI should be good at. Additionally the fact that humans already do this and now AI can do it better just shows how non-helicopter this is.
@sama Yes, last month, I tested this game with o3, gemini-2.5, pro, claude 3.7, QVQ, Doubao, and o3 won first place. See the specific test article: mp.weixin.qq.com/s/nApsHJlY_gK1…
@sama Helicopters are cool — until you realize most people can’t fly them, don’t trust them, and they crash a lot. Same goes for unexplainable AI.
