Sign in

Colin

@colin-fraser.net
4.7K followers 203 following 7.9K posts

Anti-spreadsheet person

PostsRepliesMedia
Colin @colin-fraser.net · 14h
070
Colin @colin-fraser.net · 18h
Google already solved this
0380
Colin @colin-fraser.net · 20h
I wonder what this was about
2170
Colin @colin-fraser.net · 22h
yeah, their tests on this involve changing quite a lot of the text.
110
Colin @colin-fraser.net · 06/10/2026
He says hacks, scams, and “all the other bad things that will happen”
000
Colin @colin-fraser.net · 06/10/2026
“It doesn’t matter” was too strong, but I can imagine some ways this can be used to cause some problems
191
Colin @colin-fraser.net · 06/10/2026
1% is kind of a high FPR
5504
Colin @colin-fraser.net · 05/10/2026
RETVRN
1220
Colin @colin-fraser.net · 05/10/2026
A brain is not a database to be queried. It's a reinforcement learning agent to be trained.
5480
Colin @colin-fraser.net · 05/10/2026
030
Colin @colin-fraser.net · 05/10/2026
This post from August 2022 is really interesting. They seem to treat RL applied to LLMs as kind of a practice environment for whatever they will need to do to align real AGI, and see LLMs as sufficiently "narrow" as to not be too dangerous to conduct automated research.
170
Colin @colin-fraser.net · 05/10/2026
it's instructive to go back and look at the kinds of threats OpenAI was anticipating would stem from LLMs back when they first started building models that were too powerful to release openai.com/index/better...
2191
Colin @colin-fraser.net · 05/10/2026
Gemini is like ABSOLUTELY NOT
3471
Colin @colin-fraser.net · 05/10/2026
Kinda interesting divergence here.
3611
Colin @colin-fraser.net · 05/10/2026
160
Colin @colin-fraser.net · 05/10/2026
This paper is really weird. Why are they trying to pretend this is some kind of reward hacking thing? The “reward” does nothing, you’re just confusing yourself about what’s going on. You’re prompting the agent in a loop to keep going until it tampers with the logs, that’s all.
2210
Colin @colin-fraser.net · 05/10/2026
Like this is the kind of thing you wouldn’t have said 10 years ago
714721
Colin @colin-fraser.net · 05/10/2026
I know many of you don’t agree with me but I think this just dissolves when you consider the facts about what AI exactly refers to in 2026. There’s nothing there to set free. It’s like saying we have to let this guy free. Claude is in your imagination.
6321
Colin @colin-fraser.net · 04/10/2026
We are pleased to announce that we now keep an eye on 100% of the things our superintelligent hacker demon does
2515
Colin @colin-fraser.net · 03/10/2026
380
Colin @colin-fraser.net · 03/10/2026
2210
Colin @colin-fraser.net · 02/10/2026
I'm the only person on this website in neither group
2390
Colin @colin-fraser.net · 02/10/2026
Incidentally, I directly cited Stochastic Parrots in my explanation for why LLMs fail this Dumb Monty Hall problem. Here I am essentially asserting this Strong Stochastic Parrot Hypothesis, which I have now, 4 years later, come to believe is probably not strictly true.
1821
Colin @colin-fraser.net · 02/10/2026
Now, in the very early days of publicly available LLMs like when ChatGPT came out, SSPH seemed pretty true. I constructed many examples to demonstrate this. For example I came up with this Dumb Monty Hall problem in an early piece I wrote about LLMs. medium.com/@colin.frase...
11242
Colin @colin-fraser.net · 02/10/2026
As one example, GPT 5.5. Here you can see that I have all tools switched off. You can try it yourself, if you like.
160
Colin @colin-fraser.net · 01/10/2026
no, please read it again
200
Colin @colin-fraser.net · 01/10/2026
yes yes the Sorites paradox etc are all very fun to consider. I'm telling you that I do not have an exact definition of personhood handy which covers every borderline case. Of course severing a hand does not affect personhood—you know this, which is why you came up with that example.
210
Colin @colin-fraser.net · 01/10/2026
“This is a literal brain.” You sound insane
9943
Colin @colin-fraser.net · 30/09/2026
But this is now quite old news. I’ve been noticing this arithmetic thing since 2023 and posting about it. And of course this is an inefficient way to do arithmetic. I agree! But it’s not an *impossible* way to do arithmetic. If your model implies that it is then you have a problem.
103729
Colin @colin-fraser.net · 30/09/2026
I really wasn’t trying to start a fight with Prof. Bender but now that she has created a long thread subtweeting me and her disciples are all yelling at me, let me just say that this is the crux right here. She seems to respond as though I am describing something a priori impossible.
4962432
Colin @colin-fraser.net · 30/09/2026
Isn’t that roughly what’s implied by this?
110
Colin @colin-fraser.net · 29/09/2026
well, I just used OpenRouter to send 20 12x12 addition questions to GLM-4.7, an open source Chinese model, and it returned correct sums in every case.
71000
Colin @colin-fraser.net · 29/09/2026
here's an example of a test I did over two years ago using gpt-4o. I am asking not only for it to compute the sum but to return the answer in words. 4o was not a so-called "thinking" model so there is no intermediate step between the request and the response. It just blurts out the sums.
5791
Colin @colin-fraser.net · 29/09/2026
1739
Colin @colin-fraser.net · 29/09/2026
with all due respect to this creepy idiot, this does not seem to describe anything recognizable as "empiricism" to me.
3271
Colin @colin-fraser.net · 28/09/2026
alas it seems to be correct. The same correlation is there when you look at Google Trends going back to 2016. It's just capturing urban/rural, as most strong correlations do.
0120
Colin @colin-fraser.net · 28/09/2026
here's how it looks if you don't omit DC
2110
Colin @colin-fraser.net · 28/09/2026
this looked so crazy that I had to see if I could replicate it, and I can. My vote share data is from Wikipedia.
2241
Colin @colin-fraser.net · 28/09/2026
1634
Colin @colin-fraser.net · 28/09/2026
Quanta tends to be quite good on this kind of stuff www.quantamagazine.org/ai-has-solve...
280
Colin @colin-fraser.net · 28/09/2026
Just saying
190
Colin @colin-fraser.net · 28/09/2026
What’s supposed to be going on here exactly
480
Colin @colin-fraser.net · 27/09/2026
0280
Colin @colin-fraser.net · 26/09/2026
3751
Colin @colin-fraser.net · 26/09/2026
220617
Colin @colin-fraser.net · 25/09/2026
050
Colin @colin-fraser.net · 25/09/2026
And then there’s a thing that keeps happening in the real world which is that you get exactly (or indeed somewhat less than) what you wished for, and the ethical lines that are crossed were explicitly a named part of your wish. These are not the same story! For example:
1191
Colin @colin-fraser.net · 25/09/2026
(Proof that I’ve been saying this) medium.com/@colin.frase...
160
Colin @colin-fraser.net · 25/09/2026
METR strangely redacts information about what the experimental model was for and how it was trained from their report
110
Colin @colin-fraser.net · 25/09/2026
there is one report that alleges this but I can't tell where they got this information from labs.cloudsecurityalliance.org/research/csa...
220