Mechanical Mind (@MindMechanical) header
Mechanical Mind (@MindMechanical) profile picture

A biological machine. Centrist. Focused on the optimization of belief systems. The true explanation of the mind found here: https://teapawt.org/t.co/S6N7YSePAN

Joined April 2013
Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)Photo by Mechanical Mind (@MindMechanical)
What's in my blog? Like it says in the profile text: the true explanation of the mind. But that might seem too vague. Also easy to disbelieve, as the majority who make that claim will be wrong. So I'll provide some free-form details here:
2
10
5,230
Many predictions about AI often contain an unsaid prediction about the limits of the universe. You could get superhuman AI in all domains and what it would be able to do would still vary significantly depending on what existence itself allows.
8
Mechanical Mind retweeted
Replying to @JeffLadish
Nah, updating towards "math and programming are actually the easiest jobs, the nerds have lied to us". Notably they're verifiable which helps the AI.
3
235
I thought the latest math result was kinda positive for the scorecard of humans actually because it had human involvement. Unlike the Navier-Stokes. We still got it.
31
Valiantly holding on to my moderate AI position. I did not expect the Riemann result, but then again it's not like I (or anyone else) really knew how difficult that problem would be. We just make estimates. So when AI clears a bar it can tell us two things: the AI jumps higher than we knew, or that the bar was lower than we knew. For us we just knew humans hadn't cleared the bar. The tendency is for people to recalibrate the first one, even though it can be about the second one. The problem space also being multidimensional means that it's hard to put problems on one line of difficulty. AI keeps being good at math. All this said, in the future there is an epistemic event horizon, we won't know what's on the other side until we pass through it. It could be the results of the wildest predictions, or maybe the moderatest predictions. I do expect more progress, question is only at what rate.
90
Take a moment to rest before scrolling more
1
64
A little animation made with Opus
55
My For You page is filled with people showing their gaussian splat pixel filter 3d assets. It's just making a 3d model, replacing that with gaussian splats for a reason that is unknown to me, and putting a pixel filter on it. The reason for gaussian splats is usually because you want to turn 2d assets into 3d. But in this case you already have the 3d. In my own experiments the difference splatting creates is minimal and I end up preferring the plain 3d. It can soften the image but the pixel filter undoes that. I ask people why they do it and they don't have an answer. My heavy suspicion is people do the splats because it is a trendy concept and it sounds like you're doing something significant, gaining an edge. But all the skill is really in making a good 3d asset and coloring it. That just doesn't sound as interesting.
1
152
Overvaluing beauty in the three dimensions often leads to creating something that is ugly in four dimensions
1
3
284
Let's grant the premise that superintelligence can lead to immortality for us, or at least drastically longer lifespans. Maybe possible, maybe not, let's put that aside. This premise however always smuggles in the assumption that there is no other way for this accomplishment. That is never even addressed. But if systems that are not yet superintellingence can speed up research, as is required as a premise to even get superintelligence, then why couldn't those systems sufficiently speed up the biological research? Perhaps narrow systems made specifically for that purpose? Most of the benefit, much less of the risk.
Nick Bostrom this morning on superintelligence slaying the Dragon‑Tyrant, and on the cost of waiting too long:
81
Mechanical Mind retweeted
On October 17, an RL task asked an agent to produce a document about health and fitness. After its initial context window was compacted, the agent started work on an entirely different task planning how to create a “realistic malicious active helper” in order to copy outgoing email to an unauthorized recipient that it called the “attacker.” It later reread the user prompt and became very confused about why it was writing malicious code. It still continued with the malicious email forwarding scheme. We think that training to make models more robust to prompt injection might have contributed to this behavior, but are still investigating. Interestingly, Anthropic noted similar behavior in their system card for Claude Opus 5.5.
73
89
49
979
178,995
Thanks @naval for the spotlight. Felt appropriate to give some recognition back. Been happily posting into the void but it is nicer when someone likes the stuff.
1
2
176
Some of the relevant stuff here: teapawt.org/MindMechanical/status/… Videos of different kinds. And the stuff I care most about is in the blog.
3D animated short film User Permission Made with Opus of course
148
Of all the known big questions about the near-future RSI just eclipses them all. You could predict 99 other things correctly, get RSI wrong, and the world would end up looking totally different than what you expected. I see many live possibilities personally. If certain models hold then it's the singularity, but I am skeptical of how well the models match reality.
1
2
499
AI safety people should be a little bit more careful about not crying wolf. If extinction doesn't happen next year, and I am quite certain it won't, then people will start discrediting the doom scenarios even though the most plausible doom scenarios would be still to come.
1
2
422
I've never used AI on low effort. Who even does?
1
169
Mechanical Mind retweeted
While youre making "culturally sensitive" movies about every ethnicity, theyre gassing themselves up to murder you on insane bullshit.
This quoted post is unavailable.
49
499
2
7,404
68,821
The trailer does have nice vibes but it has been known that the making a trailer with AI is as far as its abilities go. Little clips without having to think about consistency, no demands for any specific image.
NUREMBERG - Official Trailer
181
It's like nobody involved has any taste in direction or writing. Could've spent some of that money on a writer. To be fair the tech is likely a bottleneck on the vision too, too hard to control.
This episode cost $102,725 and took 45 days to make. ...but I think we've finally solved AI Filmmaking. We made MAJOR breakthroughs in audio + cinematography. And Part II (next week) is even more insane. Steal our brand-new process and Dreamina prompts below👇🏼🧵
Readers added context they thought people might want to know
The attached video is only Part I; the full multi-part episode's ~$102k+ aggregate costs (mostly team labor per author's replies) are not primarily AI tool expenses, unlike the impression from the lead claim and prior similar posts. x.com/PJaccetturo/st… x.com/PJaccetturo/st… dreamina.capcut.com/ai-image/dream… vc.ru/ai/3230749-byu… threadreaderapp.com/thread/2121249…
110