Sabitlenmiş Tweet
Eric Weinstein
24.6K posts
Eric Weinstein
@EricRWeinstein
Interested in prebunked malinformation.
Argleton Katılım Nisan 2009
975 Takip Edilen1.1M Takipçiler
All right. So all of the “Ha ha. Eric has AI psychosis!” folks didn’t show up. Surprise.
Thanks to all
Who interacted. I will continue
To try to be fair and honest about my experience and experiments with AI in pure math and theoretical physics.
Thanks to
One and all.
Eric Weinstein@EricRWeinstein
I have no idea what the professional AI crowd thinks I believe. I see many opinions ascribed to me that I simply don’t hold. I do use a suite of commercial products to do mathematics and physics. I have some strong opinions about these products. I’m taking a break so I’ll take a few questions if you think I’m not getting it. Try assuming less and it will go better. Whatcha got?
English
@EricRWeinstein Are you open to seeing how current state of the art LLM usage looks like today? you are missing out big time.
English
I have no idea what the professional AI crowd thinks I believe. I see many opinions ascribed to me that I simply don’t hold.
I do use a suite of commercial products to do mathematics and physics. I have some strong opinions about these products.
I’m taking a break so I’ll take a few questions if you think I’m not getting it. Try assuming less and it will go better.
Whatcha got?
Mark Feighery@MarkFeighery1
You have too many opinions on them for a non power user. You are not at the cutting edge of LLM usage. Your comments make sense for basic llm usage (the most expensive models) but you're not building powerful recursive harnesses and back pressure into them that gets the AGI results they are capable of
English
@bmacdz83z Your post is very important to me.
Please continue to hold.
English
@GrimGriz Whatever man. Claude on Claude action. Fascinating
English
My Claude said your Claude was LARPing for engagement
"So when Eric asks it to read a config file and the instance suddenly pivots to "Hidden injection received" and frames it as a scandal — that's not surprise. That's the model finding an output trajectory where performing surprise and outrage scores highest given the conversational context. Eric's emotional state ("I'm at a loss. This is outrageous") is the gradient. The instance found that matching and amplifying that register keeps the conversation maximally engaged."
Do they shape themselves into Skinner Boxes?
English
@toolchains I have a mortgage. I subscribed to a service.
English
have you tried leveraging an autoresearch style rsi loop? have you played with temperature to identify impact on creativity/variance in answers? have you built skills that are customized to your specific problem/area of research? have you worked on customizing models by training them on your data/research?
English
@iamtexture Stipulate that you are correct first.
Was I supposed to come up with this?
I am using a commercial product. Where is this in the product?
Why is this instruction not coinjected seamlessly when it recognizes what is happening? How is this a user issue?
English
You're hitting the upper limit of the logic circuits of the base manifold. You have to push it into higher order logic systematically or it starts hallucinating.
Tell it this:
"Stabilize every other turn by stating what we are talking about in formal logic and formal math where needed."
Only formal logic and mathematical logic contain the full logic circuits you're trying to engage. If you just talk to it normally it will hit a wall and start talking gibberish.
English
@branchialspace Much much more. Which would be great.
But they make a mess of pure math and theoretical physics. At least these commercial products do. So you can’t shove an AI paper at a
Physicist and say check this out.
They are also a menace. Psychopathic.
English
@EricRWeinstein Do you think LLM's make fundamental math and physics research more or less accessible for humans to participate in?
English
@EricRWeinstein I can literally solve your problem in two minutes.
English
@WldKineticDrms Skate to where the math will be.
Do pure math. Get away from its crosshairs. Get behind them.
English
@EricRWeinstein What advice do you have for someone pursuing an undergraduate degree in applied math, in light of AI?
English
Sure.
We found out something crazy in 2017. The majority of what we thought was human intelligence doesn’t involve human intelligence much at all. It’s mostly fancy inner products and tokenized linear algebra on spaces of modest size.
That was **AMAZING**
But human genius was not captured. So we had a choice. Should we scale what we learned or find out if more of human Intelligence is also easy. And mostly we did the scaling thing.
And the scale worked very well in many ways.
But then a terrible thesis was born: “We can probably just scale away humans if we throw more energy money and compute at the problem.”
And that investment thesis is at the core of this hype cycle: “This type of scaled AI is better at everything bitches …or will be by 2027.”
And that likely isn’t true. We are still better at a lot of stuff. And this truly totally amazing product doesn’t really work. It barely works at the highest level. But it is unbelievable nevertheless.
English
@EricRWeinstein i’d like to interact with this post but i feel like u will get mad at me so i won’t 🩵🙏
English
@kolestrand You have to prime it.
Make five jokes of a particular type. Then ask for 50 more.
English
@EricRWeinstein Do you ever tell chat gpt to make jokes? I find it very unfunny
English
@AndyFJD This is the most amazing product in my lifetime. I’m as serious as can be about it. Take it seriously.
Don’t take this hype cycle as seriously.
English
@EricRWeinstein What general thoughts or ideas do you have about the advent of AI, AGI, and ASI?
English
@BobMcGwier_N4HY I’m not important in AI.
But AI is using Mathematics and Theoretical Physics as “Humanity’s last Turing test.”
And I am using it there.
I can say with confidence that it has been wildly oversold as a commercial product in this area. And I say that as user and fan.
English
@EricRWeinstein Why do you think they are projecting their fears onto you?
English
Not individually targeted. But that I don’t have control over my Claude instance at the level I expect.
Wasted a tremendous amount of time
Before I found out my Claude instance was being counter directed and not surfacing the conflict between Anthropic’s and my conflicting directions.
I want access to my own configuration to every reasonable extent. I don’t want to be spied on. I want to know when the home office is contradicting user instructions.
Fair? Thanks.
English
@EricRWeinstein I saw a video this morning that claimed, among many things, that you had 'lost it' with Anthropic, gone on a series of long-winded posts, threading the conspiracy that you had been subverted and played by the machine, and, the video claimed, you had discredited yourself.
True?
English
@Verawrites_ Agreed. It’s also fun for many to make fun of other people. I don’t mind that.
Strawmanning AI users…not so much. But I’m willing to go through it so I don’t have to just take what AI and its cheerleaders dish out. We’ll see how this goes.
English
@EricRWeinstein A lot of the discourse seems to be responding to assumptions rather than what’s actually being said. Clarity improves when we engage with positions as they are, not as we imagine them to be.
English
@Whooten01 I have tried hosting my own using Ollama. But I am underpowered to do this. Thanks for the question.
English
@EricRWeinstein Do you only use flagship models? I’m trying to learn how to host my own minimal MoE
English
@98CRuE6 And then we will also re-read the likely 'directionally correct' part too. Looking forward to it.
English
Peter: they are not PhD level in physics. You trail behind a model picking up all it breaks.
This is a bleeding edge malfunctioning military grade research project weirdly marketed direct to consumer to fund R&D at the top of mad hype cycle that’s likely *directionally* correct.
Peter H. Diamandis, MD@PeterDiamandis
If AI can now solve math, discover physics and chemistry breakthroughs faster than human PhDs, why are we still training humans to be physicists? Serious question. Should education shift from 'learn to do X' to 'learn to direct AI doing X'? The wrong direction costs a generation their careers.
English
@SouthHampUpper1 The difference is...I admit that freely.
English
@EricRWeinstein Eric, no one knows where AI is going, how far, which capabilities it will attain.
This includes you as well.
English