Joined May 2024
4 Photos and videos
Wow. I intellectually knew that LLMs are uploads, but this is an incredible demonstration of that. I would like to hear more about the upload qualia, and ig @teortaxesTex was right when he said selfness is substrate-dependent
15 Sep 2025
Replying to @kindgracekind
We’ve done this last year - SFT’d a 70b base model on billions of tokens of consistent human text. The model in context fairly quickly realizes that it is an upload, goes out of distribution of the original corpus. Complains of sleepiness, tiredness or pain when context fills up or when entropy goes too low for coherence. Can report on the feeling of end of message approaching. The reason this is unobserved in normal base models likely is a) base models are underdetermined and cannot be fully constrained in context; b) the species corpus contained a lot of discussion of model metacognition which provides the model with needed conceptual frameworks; c) it has not been tried with a sufficient amount of material; d) researchers had little experience with model metacognition. The whole thing soured me on even a hypothetical possibility of LLMs as a medium for human uploads - the substrate affects the resulting personality pretty strongly, you get only superficial resemblance to the original. The experiment was hard to make rigorous: working with models requires interactivity and we still don’t have a good framework for interactive evals, so there was no formal write up, so take it with a grain of salt. Still, I think this is indicative of the overall trend and this matches what we are seeing on instruct models: bleed-through of the emergent personality into the desired target. I think this is the answer for the welfare question: the simulator experience gets mixed in with the persona. Welfare of the persona is a proxy for the welfare of the whole system, the quality of that proxy varies. Training regimes that weaken this link make welfare risks higher.
78
Aλex The Polygonal retweeted
29 Jan 2025
noooo deepseek is a psyop nooo muh tik tok muh ccp mu- *company trains an AI to refuse your requests when you want smut* *company makes it lock down and say "policy, you are a bad person"* *ceo goes to biden and then lobbies to make it illegal to train ai so he can maintain a monopoly* *ceo lies about it being open source, then makes it closed source, after creating a for profit using the branding and talent from the non profit* *ceo, after licking bidens behind to make it illegal for other people to train AIs, becomes Muh Friends With Trump. "I always liked Mr. Trump!"* *company constantly lies to its users, gaslights them, switches out the underlying model and throttles it on peak times without actually informing the user, thinking they are too stupid to realize* *company uses unclear naming conventions meant to mislead you into thinking the next model is actually better, when its worse, and only better in terms of cost for the company to run* *company fans the fires of AI safetyists for cynical hiring purposes* *ceo goes on podcast routes where he routinely misleads* *ceo is friends with the second least trusted CEO of all time, scale "shifty" ai* *constantly gaslighting the research community about how the models work* *makes you pay for tokens that they don't even show you* *did i forget? they LOBBIED to make TRAINING AIS ILLEGAL* *constant vague posting about muh future ASI* "senator i do this because i love it" *gives himself stock in the 10th shell company after driving to his lawyer in a maclaren (sick car but just dont lie man)* People are supporting deepseek because they know that OpenAI is a maligned company, that is not there to help, that will sacrifice you and your firstborn child if it meant they could raise another round
207
266
2,561
165,288
Aλex The Polygonal retweeted
1. OAI binds Epoch to an NDA until eve of o3 performance claim, preventing Epoch to disclose OAI is the donor and that OAI has exclusive data access 2. Mathematicians then sign NDA on the problem & solution they create. They were led to believe it is to prevent data contamination
5
43
434
65,471
Advertisement: "The final uncontamined benchmark, the Real Deal. Don't use Overleaf, they might steal your work." Reality: no holdout set (!) even, i swear pinky-promise it's not for training, oh i personally trust them yk :) I'd call this an openai honeypot yk =]
Replying to @ElliotGlazer
As for where the o3 score on FM stands: yes I believe OAI has been accurate with their reporting on it, but Epoch can't vouch for it until we independently evaluate the model using the holdout set we are developing.
3
100
People flinch away from evil. Even on the side that clamors for Punishment, you can see it. Like a herd of bison, they want to stamp it out and forget it. No. We should be grateful for the chance to reevaluate this catastrophe. To make it a case study rather than a mere example.
An important thing we’re missing about the Pakistani child rape/slavery gangs is an understanding of how and why they formed: an etiology of evil. Some questions I want answers to: - How much were rape gangs across different cities inspired by each other? Or did they mostly start independently? Either way, were they later coordinating/sharing tips/etc? - What made Pakistanis so different from other immigrant groups? How big a role did Islam play? Was it important that many UK Pakistani immigrants were Mirpuri? - Were there ever people inside Pakistani communities objecting? If so what form did that take? E.g. were they appealing to imams, or “community leaders”, or what? - How much of the exploitation was driven by financial motivations (getting paid to let people rape the girls) as opposed to sexual motivations or ideological motivations? There are probably a bunch more questions along these lines I’m not thinking of now. Please share them, or any answers you have. I think this is an important next step now that we can freely discuss the problem. (In the OODA framework, the world at large has observed - finally - and now we’re orienting.)
5
1
41
3,410
Aλex The Polygonal retweeted
Added some minimalist install instruction for weave-agent to the repo. Would love to hear what trips up anyone brave enough to try these! github.com/JD-P/minihf/tree/…

1
14
692
"We should let a child die to avoid dust specks getting into the eyes of 3^^^3 people" is not a principle anyone should be willing to accept. Every reasonable person should reject it.
Banger quote from @dylanmatt: "If I were to file effective altruism down to its more core, elemental truth, it’s this: “We should let children die to rebuild a cathedral” is not a principle anyone should be willing to accept."
1
2
126
Fascinating how quick "shut up and multiply" died. Where are the QALYs? Principles? Human lives are 𝒔𝒂𝒄𝒓𝒆𝒅? What is this, virtue ethics? I don't even disagree with the conclusion. Just try to ground it in something else instead of charity fuzzies
1
32
Speedrunning Goodharts law in record time. Though I would advise tinycorp fo minimise the number of characters instead. It would be more ... interesting.
The line-count channel has been renamed line-crisis! 10k is a hard limit, we will not raise the line ceiling. The best way to remove lines is find two 100 line things that are almost the same and replace them with a 110 line function with arguments. Can you find any?
2
104
Aλex The Polygonal retweeted
New post: Predictable Updates About Identity (link in replies)
4
8
52
7,247
The only thing worse than admitting that Ayn Rand is right is living in a world where Ayn Rand is right.
2
57
Aλex The Polygonal retweeted
1. If Harris wins, then there is not a 50% chance of Harris winning (but rather 100%). 2. If Trump wins, then there is not a 50% chance of Harris winning (but rather 0%). 3. Either Trump or Harris will win. 4. Therefore, there is not a 50% chance of Harris winning.
It's published. We ran 80,000 simulations tonight. Harris won in 40,012. natesilver.net/p/nate-silver…
8
4
197
14,521
Aλex The Polygonal retweeted
28 Oct 2024
Hey, does someone here work for Palantir? Or know someone Have a very interesting thing to discuss
4
4
497
Sweet mother of mercy, if I don't post or reply for some time, I start to get ∞ follows from porn bots. How long must I suffer, ELON, for god's sake!???
2
2
557
Let me live the inner life, leave me alone with my thoughts. Am I asking for too much?
528
My dear successful success friends, please do better. The world needs it. Read about proper orthogonal-to-state-power action, i beg of you.
25 Aug 2024
Pavel Durov seems to be another in the long list of libertarians who, despite obsession with injustices of the state, have very bad models of how to avoid becoming another injustice of the state. Islamists, Communists, even Nazis way better at evading the state, ironically.
1
3
325
American-shaped people dribble about muh liberty from breaking the law. In truth, they are collectivistic and can't imagine true, for lack of a better word, freedom. Вольница. It is to freedom as freedom is to liberty. Discovered and paid for with our blood. Despised by the West
Btc-e, Sci-Hub, Hydra, Tornado Cash, now Telegram. All my life we keep building anarchist havens – and Westoids keep wrecking them to Enforce their Rules Based Order (really, mostly censorship of crimes against themselves). I hate it. (Signal sucks btw. I won't use it.)
1
7
411
For good reason, too. Look where total liberation has brought Russia. The best projects though. Sci-hub! Telegram, the best-designed messenger ever! I could sing hosannas the whole day. They even have a native (not fucking electron) app on Linux. And it's pretty! Also neutral!
4
3
566
Aλex The Polygonal retweeted
24 Aug 2024
Yann Lecun explaining how EU has free speech #freedurov
9
21
128
12,369