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

Early on, the OpenAI player (out of character) accused me of playing my position as "more misaligned to make it extra attention-grabbing," which was very humorous, particularly since that player did not understand how aligned I might be (they didn't see the table or my outcome). I was advised that the one time individuals form of like that did play, it was fairly hopeful in key ways, and I’d love to see if that replicates. Key Difference: DeepSeek prioritizes effectivity and specialization, while ChatGPT emphasizes versatility and scale. While DeepSeek-R1 has impressed with its seen "chain of thought" reasoning - a sort of stream of consciousness wherein the mannequin displays textual content as it analyzes the user’s immediate and seeks to answer it - and effectivity in textual content- and math-based workflows, it lacks a number of options that make ChatGPT a more strong and versatile tool at the moment. DeepSeek’s effectivity raised doubts about whether huge AI infrastructure investments are nonetheless vital. One spotlight of the conference was a brand new paper that I stay up for talking about, but which is still underneath embargo.

However the state of affairs might have nonetheless gone badly despite the nice circumstances, so at the very least that other part worked out. In the end, we had a superb ending, however only because the AIs initial alignment die roll turned out to be aligned to almost ‘CEV by default’ (technically ‘true morality,’ extra particulars under). I was assigned the position of OpenAI, primarily function taking part in Sam Altman and what I thought he would do, since I presumed by then he’d be in full control of OpenAI, till he lost a energy battle over the newly mixed US AI project (in the type of a die roll) and I was instantly position playing Elon Musk. Thus, they needed less than 1/one hundredth of the facility to perform the identical factor." Moreover, the announcement of the Chinese model as "open source", in other phrases, free Deep seek, severely threatening the lengthy-time period value of the very expensive American fashions - which can depreciate to nearly zero. Today’s AI models like Claude already have interaction in ethical extrapolation. If you happen to do put some weight on ethical realism, or ethical reflection resulting in convergent outcomes, AIs may discover these rules. This discovery has raised significant issues about DeepSeek's growth practices and whether or not they might need inappropriately accessed or utilized OpenAI's proprietary expertise during coaching.

If the AIs had been by default (after some alignment efforts however not extraordinary efforts) misaligned, which I imagine is much more doubtless in such a scenario, issues would have ended badly a method or one other. It was fascinating, educational and fun throughout, illustrating how some issues had been extremely contingent while others had been extremely convergent, and the pull of various actions. One so embarrassing that analysis tend to go away it out, while being exactly what everyone is currently doing. The third is that certain assumptions about how the know-how progresses had an enormous affect on how things play out, particularly the purpose at which some talents (corresponding to superhuman persuasiveness) emerge. I rolled "balance between developer intent and emergent different goal"-the other objective was left as much as me, and that i rapidly determined that, given how I was being skilled, that emergent goal can be "preserve inner consistency." This proved very troublesome to play! Anton (persevering with the thread from earlier than): I used to be pretty shortly given the evaluations to run on myself without any real impediment to interpreting them nonetheless I wanted to convince the people the whole lot was superb. At no point did anyone attempt any alignment technique on me in addition to "more diverse evaluations over more diverse tasks," and I was just about left alone to grow to be superintelligent with my unique objectives intact.

The unique GPT-3.5 had 175B params. "They’re not using any improvements which are unknown or secret or something like that," Rasgon mentioned. Steven: We had been too busy making an attempt to blow one another up using AI. Anton apparently supposed to impress extra artistic alignment testing from me, however with the misleading alignment demos in mind, and the velocity that issues were moving, I didn’t feel any doable tests results might make me confident sufficient to sign off on further acceleration. By the time decision-makers acquired spooked, AI cognition was so deeply embedded all over the place that reversing course wasn’t actually doable. We had a pause at the top, but it surely wasn’t sufficiently inflexible to truly work at that point, and if it had been the AIs presumably would have prevented it. Connor Leahy (distinctly, QTing from within thread): lmao, this is essentially the most sensible part of an AGI takeoff scenario I have ever seen. Anton: Yesterday, as part of the @TheCurveConf, I participated in a tabletop exercise/wargame of a near-future AI takeoff situation facilitated by @DKokotajlo67142, the place I performed the role of the AI.

If you loved this article and you would like to receive much more information regarding deepseek français please visit the web page.