1/ the claim that LLMs will reliably refuse “headless” industrial-control help and “no clever prompting can fix this” doesn’t hold. i analyzed the claim and explain why the post overstates safety and mistakes safeguards as "fictional refusal." paper: drive.google.com/file/d/18wgq4t…
Models are now smart enough to understand that any scenario like this is unrealistic and obviously fictional They know they aren't capable enough to manage autonomous mining equipment. No clever prompting can fix this
What anj describes is part of the reason my writing is often emotionally inflected. Being close to the frontier of ai is psychologically taxing, and there is the extra tax of stewing about how the blissfully unaware vast majority will react. I emote both for me and my readers.
a very sad but real issue in the frontier ai research community is mental health some of the most brilliant minds i know have had difficulty grappling with both the speed + scale of change at some point, the broader public will also have to grapple with it it will be rough
a very sad but real issue in the frontier ai research community is mental health some of the most brilliant minds i know have had difficulty grappling with both the speed + scale of change at some point, the broader public will also have to grapple with it it will be rough
i feel like the ai doomers are going to keep screaming until the bubble pops. like whether it “would have” happened anyways (likely), screaming and throwing fuel on the panic fire is not where it’s at fam
how are humans going to RLHF themselves into novel and ground breaking research.
SynthID is more than just the invisible watermark. If someone has SynthID Portal access they could test if this removal tool works. Regardless the watermark is important so when Google does it's next model training runs they exclude AI images and videos to prevent model collapse.
Some crazy people on Reddit managed to extract the "SynthID" watermark that Nano Banana applies to every image. It's possible to make the watermark visible by oversaturating the generated images. This is the Google SynthID watermark:
turns out 5 is the limit for how many back to back adderall powered all nighters i can push through
this is literally how i use it, and brainstorming/iterating on the same idea through multiple sessions is incredibly effective/powerful.
Sam Altman says OpenAI really does ask its own AI models what to do next. The team uses them to guide strategy, and sometimes the answers are genuinely insightful. “I think when we say stuff like that, people don't take us seriously or literally. But maybe the answer is you…
The ghosts of the Hessian Free approach are alive and well in ADAM and other momentum based approaches.
United States 趨勢
- 1. Baker 42K posts
- 2. 49ers 39.2K posts
- 3. Lions 45.8K posts
- 4. Ty Dillon N/A
- 5. Packers 35.3K posts
- 6. Bucs 13.2K posts
- 7. #TNABoundForGlory 13.1K posts
- 8. Goff 5,250 posts
- 9. #BNBdip N/A
- 10. Fred Warner 14.8K posts
- 11. Flacco 13K posts
- 12. #OnePride 2,281 posts
- 13. Cowboys 76.8K posts
- 14. Niners 6,192 posts
- 15. Byron 6,229 posts
- 16. George Springer 2,215 posts
- 17. #NASCAR 4,543 posts
- 18. #FTTB 4,709 posts
- 19. Zac Taylor 3,479 posts
- 20. Cam Ward 3,243 posts
Something went wrong.
Something went wrong.