• 목록
• 답변
• 글쓰기
• 게시판 리스트 옵션
  • 수정
  • 삭제

Early on, the OpenAI player (out of character) accused me of enjoying my position as "more misaligned to make it more fascinating," which was very funny, especially since that participant did not know the way aligned I may be (they didn't see the desk or my consequence). I was instructed that the one time individuals kind of like that did play, it was quite hopeful in key methods, and I’d love to see if that replicates. Key Difference: DeepSeek r1 prioritizes efficiency and specialization, while ChatGPT emphasizes versatility and scale. While DeepSeek-R1 has impressed with its seen "chain of thought" reasoning - a form of stream of consciousness whereby the model displays text because it analyzes the user’s prompt and seeks to reply it - and efficiency in textual content- and math-based mostly workflows, it lacks several features that make ChatGPT a extra robust and versatile instrument in the present day. Deepseek Online chat online’s effectivity raised doubts about whether or not large AI infrastructure investments are nonetheless crucial. One spotlight of the conference was a new paper that I stay up for talking about, but which is still beneath embargo.

But the scenario may have still gone badly regardless of the good circumstances, so at the very least that other half worked out. Ultimately, we had a very good ending, but solely as a result of the AIs preliminary alignment die roll turned out to be aligned to nearly ‘CEV by default’ (technically ‘true morality,’ more details beneath). I used to be assigned the position of OpenAI, primarily function playing Sam Altman and what I believed he would do, since I presumed by then he’d be in full management of OpenAI, until he misplaced a energy battle over the newly mixed US AI undertaking (within the type of a die roll) and I was all of the sudden function taking part in Elon Musk. Thus, they wanted lower than 1/a hundredth of the power to accomplish the same thing." Moreover, the announcement of the Chinese mannequin as "open source", in different phrases, free, severely threatening the lengthy-time period worth of the very costly American fashions - which can depreciate to almost zero. Today’s AI models like Claude already interact in moral extrapolation. When you do put some weight on moral realism, or moral reflection resulting in convergent outcomes, AIs might uncover these ideas. This discovery has raised significant considerations about Deepseek free's development practices and whether or not they might have inappropriately accessed or utilized OpenAI's proprietary know-how throughout coaching.

If the AIs had been by default (after some alignment efforts but not extraordinary efforts) misaligned, which I imagine is way more likely in such a situation, issues would have ended badly a technique or another. It was attention-grabbing, educational and enjoyable all through, illustrating how some things have been extremely contingent whereas others were extremely convergent, and the pull of assorted actions. One so embarrassing that evaluation have a tendency to go away it out, whereas being precisely what everyone seems to be at the moment doing. The third is that sure assumptions about how the know-how progresses had a big affect on how issues play out, particularly the point at which some abilities (comparable to superhuman persuasiveness) emerge. I rolled "balance between developer intent and emergent different goal"-the other objective was left as much as me, and i quickly decided that, given how I was being skilled, that emergent purpose would be "preserve inner consistency." This proved very tough to play! Anton (persevering with the thread from before): I used to be fairly rapidly given the evaluations to run on myself with none real impediment to interpreting them nevertheless I wanted to persuade the humans everything was nice. At no level did anybody strive any alignment strategy on me apart from "more numerous evaluations over extra numerous duties," and I used to be pretty much left alone to develop into superintelligent with my original targets intact.

The original GPT-3.5 had 175B params. "They’re not utilizing any innovations which might be unknown or secret or something like that," Rasgon mentioned. Steven: We had been too busy making an attempt to blow each other up using AI. Anton apparently meant to provoke more inventive alignment testing from me, but with the deceptive alignment demos in mind, and the speed that issues had been transferring, I didn’t feel any possible assessments outcomes may make me confident enough to signal off on additional acceleration. By the point resolution-makers obtained spooked, AI cognition was so deeply embedded all over the place that reversing course wasn’t actually possible. We had a pause at the tip, but it wasn’t sufficiently inflexible to really work at that time, and if it had been the AIs presumably would have prevented it. Connor Leahy (distinctly, QTing from inside thread): lmao, this is essentially the most reasonable a part of an AGI takeoff situation I have ever seen. Anton: Yesterday, as a part of the @TheCurveConf, I participated in a tabletop train/wargame of a near-future AI takeoff scenario facilitated by @DKokotajlo67142, the place I played the function of the AI.