In this range of code lengths, 400-1800 lines, lines of code does not correlate with effort imo. It only takes like 1 day to write 1800 lines of code by hand. The actual effort is dominated by thinking of ideas and huge hyperparameter sweeps.
Another note: i was curious what they had to do to reduce torch startup time and such, and it turns out they spend 7 minutes compiling and warming up for their 2 minute training run lmao. That does make it more realistic but is a bit silly.
There’s obviously no truth to that claim. Labs absolutely have better models.
Claude 4.5 sonnet is relatively bad at vision for a frontier model. Gemini 3 pro, GPT 5.2, and Claude 4.5 opus are better.
Saying you would 2 box in Newcomb’s problem doesn’t make sense. By saying it aloud, you’re directly causing anyone who you can influence and who would later have an opportunity to casually cooperate with you to not do that. If you believe acausal stuff is possible, you should always respond with “1 box” or “no comment”, even if you would in reality 2-box.
I support a magically enforced 10+ year AGI ban. It’s hard for me to concretely imagine a ban enforced by governments, because it’s hard to disentangle what that counterfactual government would be like, but I support a good government enforced AGI slowdown. I do like it when people shout doom from the rooftops though, because it’s better for my beliefs to be closer to global average average, and the global discourse is extremely far from overshooting doominess.
Yeah it goes out of its way to say the opposite, but if you know Nate and Eliezer the book gives the impression that their pdooms are still extremely high, and responding to the author’s beliefs even when those aren’t exactly the same as the text is sometimes correct, although not really in this case.
If you have a lump of 7,000 neurons, they can each connect to each other neuron, and you can spherical-cow approximate that as a 7000x7000 matrix multiplication. That matrix multiplication will all happen within O(1) spikes, 1⁄100 of a second. That’s ~700GFlop. An H100 GPU takes ~1 millisecond to do that operation, or 1M cycles, to approximate one brain spike cycle! And the gpu has 70B or whatever transistors, so it’s more like 10M transistors per neuron!
Note that since Paul started working for the US government a few years ago, he has withdrawn from public discussion of AI safety to avoid PR and conflict of interests, so going off his writings are significantly behind his current beliefs.
YC batches have grown 3x since 2016. I expect a significant market saturation / low hanging fruit effect, reducing the customer base of each startup compared to when there were only 200/year.
I’m surprised that’s the question. I would guess that’s not what Eliezer means because he says Dath Ilan is responding sufficiently to AI risk but also hints at Dath Ilan still spending a significant fraction of its resources on AI safety (I’ve only read a fraction of the work here, maybe wrong). I have a background belief that the largest problems don’t change that much, and it’s rare for problems to go from #1 problem to not-in-top-10 problems, and that most things have diminishing returns such that it’s not worthwhile to solve them so thoroughly. An alternative definition that’s spiritually similar that I like more is; “What policy could governments implement such that the improving the AI x-risk policy would now not be the #1 priority, if the governments were wise.”. This isolates AI / puts it in context of other global problems, such that the AI solution doesn’t need to prevent governments from changing their minds over the next 100 years or whatever needs to happen for the next 100 years to go well.
I would expect aerodynamic maneuvering MIRVS to work and not be prohibitively expensive. The closest deployed version appears to be https://en.wikipedia.org/wiki/Pershing_II which has 4 large fins. You likely don’t need that much steering force
I really struggle to think of problems you want to wait 2.5 years to solve—when you identify a problem, you usually want to work on solving it within the month. Just update most of the way now + a tiny bit over time as evidence comes in. As others commented, no doom by 2028 is very little evidence
I heard some rumors that gpt 4.5 got good pretraining loss but bad downstream performance. If that’s true the loss scaling laws may have worked correctly. If not, yeah a lot of things can go wrong and something did, whether that’s hardware issues, software bugs, or machine learning problems or problems with their earlier experiments
This is OpenAI cot style. See it in the original o1 blog post. https://openai.com/index/learning-to-reason-with-llms/
I can imagine scenarios where you could end up with more resources from causing vacuum decay without extortion. Like if you care about doing something with resources quickly and other agents want to use resources slowly, then if you cause vacuum decay inside your region, the non collapsed shell of your region becomes more valuable to you relative to other agents because it only exists for a short duration, and maybe that makes other agents fight over it less. Or maybe you can vacuum decay into a state that still supports life and you value that
Whether you can cause various destructive chain reactions is pretty important. If locusts could benefit from causing vacuum collapse, or could trigger star supernova, or could efficiently collapse various bodies into black holes, that could easily eat up large fractions of the universe.
No, AC actually moves 2-3x as much heat as it’s input power, so a 1500W AC will extract an additional 3000W from inside and dump 4500W outside
This overestimates the impact of large models on external safety research. My impression is that the AI safety community has barely used deepseek r1 and v3 open source weights at all. I checked again and still see little evidence of v3/r1 weights in safety research. People use r1 distill 8b, and qwq 32b, but the decision to open source the most capable small model is different than the decision to open source the frontier. So then it matters when 8b or 32b models can assist with bioterrorism, which happens a bit later, and we get most of the benefits of open source until then. It’s also cheaper to filter virology or even all biology data out of training for a small models pre training data because it wouldn’t cause customers to switch providers (customers prefer large model anyway) and small models are more often narrowly focused on math or coding.
What are your API costs, and how do they compare to the $ raised?
Redwood research did very similar experiments in 2022, but didn’t publish about them. They are briefly mentioned in this podcast: https://blog.redwoodresearch.org/p/the-inaugural-redwood-research-podcast.