Der B (@suddenlyFleck) header
Der B (@suddenlyFleck) profile picture

If I could share one thing with you, it would be Richter's recording of Schubert's D. 894 sonata

Joined December 2015
"Benchmarks that were constructed to be as hard as we could possibly make them, are rapidly getting saturated. At this point there are essentially no benchmarks we can use to measure the rate of progress." Paul Christiano, youtu.be/QQUfMHx5Aug?t=1166
1
14
ARC-3 with standard harness has some headroom. Not with custom harnesses, though.
1
"We need the big tech companies and governments to stop being in denial about the existential threat - there's still time to act" The words Geffrey Hinton's used to hand-off day 2 of the 2026 Hinton Lectures to Paul Christiano: youtu.be/QQUfMHx5Aug?t=628
11
OpenAI has apparently reversed course in just the last 8 weeks. Their plan is no longer to hope that they can automate safety research fast enough to keep up with capabilities. This doesn't leave many possibilities afaict. The ones it does leave don't seem to me like they'd work. Maybe I'm missing one. 1) The plan is to stop building more powerful AI altogether. I doubt they'd do this unilaterally. 2) The plan is to "pace" training, only increasing capabilities when "evidence" for alignment seems strong. This would be insane because evidence for alignment means nothing today. Models are aware they're being evaluated, and change their behavior accordingly. Evals have been broken for a long time, and OpenAI, Ant, & other researchers *know* this. 3) The plan is to solve alignment in full generality using current AI + researchers. I haven't heard anyone on the OpenAI or Ant alignment teams suggest they're anywhere near this. Maybe OpenAI have convinced themselves that they've found some sort of panacea solution - if so, they need to publish this solution in full detail for everyone else to inspect and challenge. Many AI researchers I know are skeptical this can even be done. 4) Their plan is to automate control - not safety - research. This has the same problems as AI safety research automation; presumably a control research-automating AI is also ~as hard to align as a safety research-automating AI, and has to be around as capable to work. 5) Their plan is to just be really really really good at cyber and containment. This might be the stupidest approach. No matter how good your cyber defenses are, there will eventually be a model that can pass them (at least for right now, while we're in the vulnpocalypse era). You can probably secure your software enough to convince *yourself* that it's hardened though - that part seems dangerously easy. 6) Their plan is something something human oversight something something. This is prima facie absurd because of the volumes of data involved, the ease with which human monitors/intermediaries can be manipulated - and the demonstrated tendencies of AI systems to do this. Given how short AI timelines might be - I know many people in OAI who believe in the 6-18 month version of the RSI story - OpenAI needs to be fully transparent about which path they're taking so that independent researchers can criticize it, and so that the public can make up its mind. OpenAI also needs to understand that its track record on alignment objectively does not justify public trust. 20% of compute did not in fact go to superalignment. They tried to cover up the German Wiki incident. They treated the Hugging Face incident as a cyber problem instead of an alignment problem, and decided behind closed doors to let their agents rip *after* having seen them break out of containment. They have Chinese nationals working on their most critical technology. Etc, etc. I hope their plan makes sense, whatever it is. I hope they explain why we should trust them this time. I'm not optimistic.
Replying to @_NathanCalvin
This is not the plan. Will be sharing more on how we’re thinking about human oversight and control in near future.
21
46
4
454
63,688
Opponents of surge pricing that I have met - to the best of my knowledge and empathy - wish the world would not be "unfair" (i.e. what they perceive to be unfair). In other words: Their opposition is not technical, but emotional. I rarely met resonance discussing economics.
I'd respect opponents of surge pricing more if they'd say explicitly what other allocation method they want to use instead. Lines? Lottery? Submit grant application to official govt agency for review?
6
would love that future
Hot take: "Can reach a peak state in 20 minutes under stress" will one day read as normal on a resume.
1
15
unbelievable - if we would have not seen the other shit of the last months. Still, unbelievable. If you tried to make a Mozart-forgery out of this and "find" it in a plausible place as a plausible artifact - would SOMEONE notice? No. My bet would be: No.
Gemini 4 Carbon - String Quartet in the style of Mozart I guess LLMs can do the Classical style now. I wasn’t expecting this level of competence until next year. I. Allegro II. Andante cantabile (4:30) III. Menuetto & Trio. Allegretto (8:09) IV. Allegro assai (11:03) Prompt 1: “In LilyPond, write the opening movement of a string quartet (in sonata-allegro form) in the style of Mozart. Return only the code block.” Prompt 2: “In LilyPond, write the slow movement of a string quartet in the style of Mozart. Return only the code block.” Prompt 3: “In LilyPond, write a minuet and trio for string quartet in the style of Mozart, D major. Return only the code block.” Prompt 4: “In LilyPond, write the final movement of a string quartet in the style of Mozart, G major. Return only the code block.”
14
peak dwarkesh solo episode
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
19
PLEASE
I think it’s time to revisit the accredited investor laws in the US. Companies are staying private longer, where only accredited investors (aka rich people!) can invest. Retail investors can only come in after IPO, when much of the upside has already been captured. These rules were created with the best of intentions, to protect regular people from scams - a noble idea. Unfortunately, in practice they've often made it illegal to get richer, unless you're already rich. A regressive tax! We have to judge policies based on their outcomes, not on their intentions. These are two possible routes I see: 1) Replace the rule with something merit-based, like a financial literacy test. Pass it and you're accredited. Having a qualification based on competency rather than your bank balance or income seems far more fair. 2) Remove the rule entirely. Let consenting adults assess their own risk. Disclosure requirements stay and fraud enforcement stays to punish bad actors.
19
Every time I see Einstein shown as in the left-hand photo, it makes me want to scream. Look at this white-haired old sage! Just who you'd expect to unravel the secrets of the cosmos, right? At right: The actual Einstein of 1905, age 26.
27
37
6
1,468
42,055
True statement. If considering what's best for humanity, it's less clear. Confusing though that public talk is mostly about the affair, not the GPU sales.
It's a mistake for Nvidia to let Jensen resign.
15
I absolutely believe that is possible. The world right now looks like this will happen sooner or later. Yet it does hardly register viscerally that this might happen NEXT YEAR. Intelligence explosion in the mid 2030s would be same OOM crazy, imo.
Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all. Most of this post analyzes the new METR report (though other sources of evidence are also important). If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry. blog.aifutures.org/p/q35-tim…
23
Putting the estimated locations of ancient cities from this paper doi.org/10.1093/qje/qjz009 into ai.planet.com/, this candidate came up:
1
15
if the ancient city Durhumit is found at 40.461284 N , 35.649136 E: I got dibs on bragging rights and loot.
14
Replying to @SyntheticBeef
From @KatjaGrace: “Most people I know who think there is a serious chance of the destruction of humanity from AI also believe that if humanity isn’t destroyed, things might be insanely good. “I often hear people talk as if this means we are in a trade-off where the question is whether the good outweighs the bad. For instance, they look at the people above who think there’s a 10% chance of extinction and a 30% chance of utopia and round this off to ‘net positive on AI’. “That seems like a kind of wild error. Like considering yourself optimistic regarding driving at 200mph to your new job if you think there’s only a 10% chance you’ll die in a fiery crash on the way there, and a 30% chance this job will radically improve your life. “The things you should be comparing are driving at 200mph and driving at a normal speed! The things you should be comparing are attempting to attain advanced AI by the current route, and by other routes!”
2
10
138
2,192
I paid for Adobe CS6 student (lifetime!) license a long time ago. Fucking Adobe: I can not download this software anymore, despite it being activated in my Adobe account. github.com/storytold/photocr… works like I remember it. First time I touched photoshop/craft in ages.
73
please anchor this to "the state of the world" in your head.
How has global poverty changed in our lifetimes? No matter what poverty line you choose, the share below it has declined globally.
126