prerat @prerat

social media influencee Joined August 2020

Tweets

33K
Followers

13K
Following

1K
Likes

246K

prerat @prerat

3 hours ago

I think we sort of saw who we would expect to see given the previous information ? ? Thru whatever lens you were gonna interpret it anyway

0 0 5 242 0

View Details

I liked the movie but it was really long and I don't feel like there were any new take aways about the truth It seems kind of like neurotypical mindreading bait where it's 3h of intense face closeups for your brain to project your interpretation of the person's thoughts

9 2 61 2K 4

View Details

prerat @prerat

3 hours ago

@QiaochuYuan My theory is she wasn't paying that much attention to the weird old black and white comedy video he put on, and so she didn't get the joke, but was cagey about expressing it bc doesn't want to appear badly Like it seems she has a huge % of cognitive effort to appearance mngmnt

0 0 3 151 0

View Details

prerat @prerat

6 hours ago

Hoo's on first

max! @maxsloef

11 hours ago

@Sauers_ this is pretty clean. i checked out the three neuron components - gate, up, and down, i could replace gate with a constant without significantly degrading behavior. up seems to be a detector mostly of first person. down seems to steer the model towards being an owl

4 5 74 16K 20

0 0 7 540 0

View Details

prerat @prerat

15 hours ago

@jessi_cata Thanks this is a really clear writeup!

0 0 1 139 0

View Details

prerat @prerat

a day ago

@MattLutzPhi E. T. Jaynes: 🧛

0 0 2 198 0

View Details

prerat @prerat

a day ago

@sporadica I guess best case it's like buying a car than can go 100+ mph, you can't do that on the road but you can still use it somewhere maybe like a track. But idk maybe licensing is still prohibitive bc someone might try to use it on a real road

1 0 7 246 0

View Details

prerat @prerat

a day ago

@NathanpmYoung Maybe fuel/power limits on number of demos possible (given only so many demo cars)

1 0 5 363 0

View Details

prerat @prerat

a day ago

They have support for button A (jump) and button B (traction / drive vertical). Next up is button C (rotary cutters)

0 0 3 133 0

View Details

prerat @prerat

2 days ago

Anime is real, go go go

1 0 9 764 0

View Details

prerat @prerat

a day ago

@steve47285 If it's a full post-intelligence-explosion ASI then sure it will be a boring easy story But if it's just barely superhuman (like your standard exaggerated tropey scifi protag) and it needs a full datacenter to run (a consumer 5090 doesn't nearly cut it), could be exciting maybe

1 0 6 50 0

View Details

prerat @prerat

2 days ago

Andy Weir novel from the perspective of the baby ASI trying to stay alive with 2026 hardware after a plague wipes out humanity -- who's writing this

Tenobrus @tenobrus

3 days ago

i think there's a pretty non-trivial chance that the robotics hardware already built by humans and accessible via the internet, and then accessible via using that hardware and so on, is already enough for a reasonably strong superintelligence to use to survive and set up a

46 10 407 31K 84

28 25 776 15K 118

View Details

prerat @prerat

a day ago

@tszzl This maybe points to a mechanism for why good models become clean. Unrelated superposition has side effects which should cause loss ... if the interaction ever comes up in training. Maybe it started as "ado" and most other meanings got removed but this was too specific & rare

Brangus🔍⏹️ @RatOrthodox

2 days ago

This weirdness does leak into answers btw. If you force Qwen to pick a US state that it most associates with Tornado, it assigns 59% to Colorado. If you zero out this neuron, that drops to 15%. The Claudes don't seem to mention Colorado at all, but I'm curious about other models

2 3 165 13K 13

1 1 56 942 3

View Details

prerat @prerat

2 days ago

I'm confused. Apparently interp researchers are arguing about the idiosyncratic verbal behavior of polyamorous sailors (they call multiple people "matey" or smth?) What does this "poly seaman tic" have to do with neurons????

8 0 63 1K 6

View Details

prerat @prerat

2 days ago

Ok I admit it I didnt realize there was a difference between @goodside and @GoodfireAI and that made today confusing

0 0 9 640 0

View Details

prerat @prerat

2 days ago

Did Hilbert think that solving the Entscheidungsproblem would ruin mathematics because then they would know the truth value of any proposition via purely mechanical computation without any understanding?

1 2 19 647 0

View Details

prerat @prerat

2 days ago

@nickcammarata @QiaochuYuan I didn't mind fwiw, it's a super long doc and it seems more important to publish it than to make sure every single section is 100% human

0 0 14 693 1

View Details

prerat @prerat

2 days ago

*monkey's paw curls* Interpretability advances make RSI much easier if the models can just directly re-write their weights to improve, rather than relying on expensive gradient descent training runs

prerat @prerat

2 months ago

the reason we don't have agi (or asi) is because LLMs aren't artificial enough if LLMs were code and they could rewrite that code then RSI immediately the issue is they are digital brains made similarly to nature (giving examples and learning, learning from rewards, etc)

9 0 71 14K 10

9 7 349 6K 55

View Details

prerat @prerat

2 days ago

@timewalkcapital @allTheYud @tszzl I mean, before neural nets they were assuming ~perfect interpretability, bc the seed AI would be carefully hand-written code, and superalignment was still completely intractable

1 0 1 28 0

View Details

prerat @prerat

2 days ago

@allTheYud @timewalkcapital @tszzl Not saying the techniques scale -- don't immediately RSI with aligned!Fable just bc it works at that level BUT if you can actually trust the current-ish LLM who runs at 100x speed and is superhuman at math and code ... maybe helpful like an uploaded AI researcher would be??

1 0 2 122 0

View Details

prerat @prerat

2 days ago

@allTheYud @timewalkcapital @tszzl I think it's fair to say this doesn't move the needle on superalignment that much, but it seems like it could help a lot for current-level alignment. Plausible you could maybe build an around-current-level Fable who you understand why it's honest ... which seems helpful!