logicallee

I read your proposal, I think it's great! Where will the navigator get the model if the user agrees to download it? For this demonstration I just serve the models on my own server, but for larger models it may be an issue as they may not have direct download links even if they are open weights.

You can try 7 different tiny LLM's in your browser.

This is really interesting. I think the Fireworks Serverless Training infrastructure they used to develop it is also unique and needed. Except if someone works at one of a handful of the largest labs, it is very difficult to set up or try any sort of training pipeline. The managed training infrastructure makes it available to more people.

Nice, I like it. The fonts are very interesting and captivating.

I'm not suicidal and have never been suicidal or thought about suicide. I'm happy and in a good mood every day. If someone is thinking about suicide please reach out to me: rviragh@gmail.com I'll listen to your issues and concerns and we will find a better solution for you.

Humans don't write things like: "Now, the popular way to tell this story is "the system is breaking," and I don't think that's right, and the distinction matters."

I answered requests to be a peer reviewer. (I'm not sure why I was selected, I don't have many publications or credentials.) I saw a lot of papers with hallucinated references. I also remember one paper that described a methodology that I don't think the authors really performed, I think it was just academic fraud where they pretended to have performed an experiment. At the time that I answered the journal requests, AI could hallucinate fake reports, but agents weren't powerful enough to run the experiments yet.

These days agents are able to really perform genuine experiments and write up the results. A prompt like this: "You'll work autonomously end to end to select a research task that meaningfully advances the state of the art in AI, is clearly defined and worth performing, that people would be interested in reading, and that you can perform on this hardware" (insert details) " in a week. Carefully log your steps so that your results can be replicated. Then, do a research review and write your paper about it up with correct, cited references. You must check all of your citations. Look up current lists of "Claudisms", (such as use of the word "genuinely", or "load-bearing"), and remove them from your writeup. After writing your writeup, edit it and pare it down, remove anything unnecessary, keep it fast paced and interesting. Also, try to tell a story, be engaging in your writeup. Don't use violent metaphors, remove references to killing, strangulation, etc. Your writeup should be ready to publish and accurately reflect a real experiment with a meaningful result that advances the state of the art and contributes to understanding. Be concise and focus on why it matters."

Okay, so there's the prompt. You can give it to any AI and have a journal-ready publication in a week. I guess you can ask it to add charts and stuff, if you want to be fancy.

If I gave my agent the above prompt, would I be one of the authors? Maybe it's fair to say I guided, facilitated, elicited, or advised it. But it's clear that the AI would be the one that is actually selecting and running the experiment and writing up the results.

Someone could probably get a publication without even reading the paper they wrote their name on. Their only contribution might be editing their name into the PDF.

How much crypto do you think the mentioned 30 GPU years would have produced at current exchange rates? They're not as efficient as ASICs but GPU's can still mine a lot...

Thanks for sharing your experience.

I recommend you try giving your AI an open ended task like, "it's at 4 fps even though it's a 22 year old game and the raw hardware performance is many orders of magnitude higher. Maybe due to running under Wine and Rosetta. Please figure out the simplest way you can unlock usable performance (30-60 FPS) short of decompiling, reverse engineering, and reimplementing the full game yourself natively. Maybe there is something else you can figure out for the performance unlock. The hardware is definitely there, so there has to be a way. The game is running so make sure you don't break it. Keep the same game behavior."

Regardless of which AI you're using, it will be able to figure something out in a few hours. 2004 pc's were 500-800 mhz, 128 MB RAM, and if they had a GPU at all it was 16-32 MB, 0.2-1 GLOP.

Your AI should be able to figure out a way to get it running in realtime.

(In case anyone remembers the compression post from yesterday[1], I checked and this one doesn't qualify for further compression - it's not zero-biased at all.)

[1] https://news.ycombinator.com/item?id=49732931

That seems to refer specifically to a provided history for it to read - presumably, this would just have the project it's running in. ("Variables (‘$node_dir‘, ‘$history_dir‘, ‘$baseline_dir‘, ‘$eval_program‘, ‘$problem_file‘) are filled in by the calling system.", meaning that a history directory for it to read is provided by the harness.)

This is a solid and very interesting paper! The authors were kind enough to publish the complete prompt for it (appendix B.1 on page 18), so anyone can try their approach with any LLM and see the results.

This exposition explains voltages and power as used in modern computing and memory, and the possibility of reducing this for power savings. Feedback and corrections welcome.

That's really interesting, thanks for your answers!

I've always wondered how mosaics were made, like, how could they know where to put which stone? Did they prepare a sketch on papyrus first and then put stones to match it? I'm having a lot of difficulty conceptualizing how it could be done.

yes this is correct. Thanks for writing that out, it's really interesting to read how you solved this without picturing it.

It's a bit surprising to me that it was unclear what I meant by horizontal/level, I thought this word is well-known and easy to understand. Here's what I mean by horizontal: For me, it means that with respect to the ground, the ray isn't rising or falling as it travels, i.e. it's level (like a plane flying in level flight with respect to the ground, not rising or falling). This constraint doesn't really matter I guess, except that I said that the cube is also level on the ground on one of its surfaces, so since you have to cast the ray in a horizontal/level way, you can't cast the ray at an angle through one of the corners so it goes through the opposite corner of the cube, since - if you think of the ray like a plane - you can only move the angle of the ray side to side not up and down, it stays in level flight = horizontal, so you can't cast it through both opposite corners if the cube is lying flat on the horizontal on one of its surfaces. As I just wrote this paragraph, I pictured what my description of casting through both corners the would look like and why it's not level with the ground anymore. I guess it's harder to follow my words without picturing it.

You mentioned it's not clear how being horizontal constrains the ray, based on what I wrote and what I meant about it being level, is it now clearer how this constrains some of the rays that can be cast? (Like how you can't cast a ray through the opposite ends of the cube now)? Or is that hard for you to follow? For me I think maybe my words could be written in a bit confusing way but I think someone can understand them if they can picture what I referred to. Without that mental picture it could be difficult though.

>getting us off of commercial LLM

sounds fun, I had fun fine-tuning some open weights models for local use, and I run a free public inference engine here: https://stateofutopia.com/services/ai/ it's fun.

where are you planning to run your models?

Yes that's correct. How do you answer this one: a hollow cube lying with one face on the horizontal like a die on a table has a hollow pyramid inside with the same base horizontal on the bottom like a usual pyramid. Inside the pyramid is as large a sphere as will fit. A horizontal ray is cast from outside of the cube through the entire structure, passing through each surface it would hit. The location of the ray is not specified, only that it is horizontal.

1) State the order of the surfaces it hits/passes through if you choose to cast it at a point where it passes through the maximum number of surfaces in the structure.

2) Is it possible to choose a different point where it would still pass through the cube but would pass through a different number of surfaces? (such as by missing the pyramid or sphere or both). State the distinct counts of how many surfaces it could pass through depending on where you cast it horizontally, like how many different counts exist. (Being tangent does not count as passing through a surface.)

I would answer this question by just picturing it and moving the ray around in my mind. How do you do it?

It reminds me of a joke. It's better told in person but you get the idea:

A percussionist takes a day off and decides to see one of his orchestra's performances. The next day he's back and talking to one of the other percussionists. "How was it?" the other one asks.

"Unbelievable! You know the part where we go buhm buhm buhm buhm?"

"Yeah?"

"Well guess what the rest of the orchestra is doing?"

hums a spirited rendition of the William Tell Overture!

Can you complete tasks like answering the question "a square shaped piece of paper is folded once diagonally so two previously opposite corners are now on top of each other. Does the distance between the other two corners change during this fold?" I think most people would answer this by just imagining it, but I'd think the way I've described it is pretty hard to follow if you can't picture it. If you can answer the question, how do you do it?

all this while Cognition doesn't even make the top page of search results, that's some impressive stealth.[1]

(just to be clear, I am a far-left activist who spends most of my time working on funding Social Security Trust Funds (OASI & DI Solvency), which could impact my search results - this was while I was logged in.)

[1]https://www.google.com/search?q=what%27s+cognition+in+ai or https://imgur.com/a/UdxtnGg

Along the same lines, I asked ChatGPT-5.6 Sol to evaluate the feasibility of a neural network running natively on quantum computing, you can see the study here:

http://taonexus.com/publicfiles/sep2026/quantum_neural_netwo...

In short it's not really feasible and it suggested classical coherent photonics and in-memory compute as more viable approaches.

(Off-topic aside: these days I am more interested in funding Social Security Trust Funds (OASI & DI Solvency) - if anyone at OpenAI can help reactivate my account: rviragh@gmail.com it would let me do further studies that directly support this important goal, currently my chatgpt account was deactivated. I apologize for any mistakes I made earlier, it won't happen again. Please reactivate my account - thank you.)

My OpenAI account was deactivated on Sunday due to a claimed infraction of production of child materials, maybe based on a few words in a technical chat that clearly isn't about that. Can you take a look? rviragh@gmail.com - I was doing a lot of important work and projects and sharing much of my work with OpenAI. I also am a big proponent of funding Social Security Trust Funds (OASI & DI Solvency) so reactivating my account would let me do that as well. Thank you for taking a look.

The presentation will be livestreamed on my channel in about an hour (11 a.m. EST, 8 a.m. Pacific). Anyone is welcome to attend.

amazing, it's a huge achievement. can someone clarify, where the writeup says "The finished proof was checked by Lean; it uses just Lean’s three standard axioms" what does this mean? Aren't there a large set of standard axioms that are also necessary? (i.e. ZFC+)? if not, since it's only three axioms, can someone say what they were?

can the grandmasters still win with rook odds?

>exadurates

this word is usually spelled exaggerate (yes, with two g's.)

people very soon as going to look back at all of us and just think "they didn't even have robots yet! how did they even eat?" Well, future readers, it isn't easy! We either go to the store and bring back groceries ourselves or pay a helper to do it. No robot in the loop there. Then when it comes time to cook something we either have to cut it up and otherwise prepare it and then cook it ourselves, or another human can do it, or we can buy fully prepared foods that don't require cooking. And then when we're done we have to put our own dishes away into a dishwasher. We can't just say a few words and have a robot do it for us!

Congratulations on the launch! The demo in the video you linked looks awesome (though your video resolution is a bit low, some of the text is hard to read on screen, I suggest you record in higher resolution). I think you made a really good choice to have your connector provider handle OAuth credentials, since it is really hard to get security right on a multitenant app like yours, where some of your agents are doing work on customers' behalf and others are improving your product itself :). The design of your app interface also looks really great!

could you elaborate on what happened?