logicallee

This is really interesting. I've been fascinated with new and unusual battery tech for a while. A few months ago I had one of the reasoning models crunch the numbers on using a superconductor as a battery.[1] (It's not viable.)

Electric airplanes: the power density (per weight) of current batteries is very low compared with fuel, if you look at electric airplanes they are only able to make short flights due to the battery weight. If a quantum battery of large size could end up with a high all-in power density for the entire system, then it could power electric airplanes.

[1] https://news.ycombinator.com/item?id=47731696

human brain uses 20 watts.

Why are chips still built in only a few layers?

Even modern silicon only has some 60 of them right?

Plus don't transformers work in layers, making it natural to put a bunch of layers on top of each other? Regarding cooling, couldn't there be liquid cooling through a plumbing network in the same die?

Our brain takes quite a large volume, and we want chips that have a similar level of parts count (achieved - 80 billion neurons) and interconnectedness (not yet, around 10k synapses per neuron). Why do it in 80 2D layers? Is there no way to etch silicon in 3D?

I'm highly interested in their Grok 4.6 run which is currently running. I think it is a very good fit for doing really well on this benchmark. Will have to check this page again in a couple of days.

I wrote a quick review of Grok 4.6 here: https://taonexus.com/publicfiles/aug2026/grok-4-6-review/

I can't put my finger on it, but there's something very satisfying about seeing his final build, that I think wouldn't be there if it were a touch screen.

I agree that it's fast and accurate, but Grok 4.6 was released only 10 days ago so you can't blame someone for not trying it yet. (Versus many months at the frontier level for Claude and ChatGPT.)

Here's a quick review I just posted if anyone's interested:

https://taonexus.com/publicfiles/aug2026/grok-4-6-review/

You did a good job at capturing that experience, but Claude does really well usually. Don't you almost get what you want at the end?

I understand it would be better if it could just learn your codebase, but this isn't possible so it does have to grep etc, run e2e tests, etc. each time.

I avoid Opus 5, and reverted to Opus 4.8 over similar issues. Opus 4.8 is still working great for me!

not having any body, continuous sensory input (except ChatGPT-live gets streaming audio), episodic memory, on-the-job learning (live weight updates), any live feedback loops (like moving a motor updates proprioperception or turning physically changes what a streaming camera sees), really hurts these models' abilities to perform tasks of the kind you're waiting for.

They are very good at instruction-following and you can teach it a new task that fits in its context and it'll learn it and do it. Go ahead, you can make up some new brand new ruleset or behavior and instruct it to follow it and it will. That's amazing.

But it won't be any better at its new behavior after an hour or ten hours or ten days. It doesn't have the kind of adaptation that we expect.

What it is able to do already is pretty amazing, but what it lacks is also a great hindrance to seeing its full capabilities. We just have to wait until research labs add these missing components.

I tried it! Works totally awesome. Your OS is amazing.

how do you know they were really official company recruiters?

the parts you wrote aren't what people are complaining about.

"written by a human, on purpose."

the pain is momentary and behind a guardrail.

okay, this is the prompt: https://pastebin.com/hhfmEYfG

it works, it one-shotted everything from just the one prompt.

you may also find this chat excerpt relevant: https://pastebin.com/F3Qwy2Ly to understand why I put that stuff at the end to motivate it to complete this very difficult port.

Thanks for the feedback. Claude has built the portions I've mentioned, it works. I'll publish the site shortly.

This is extremely cool. I activated the camera and it recognized the position of my hands and I was able to play it as shown in the instructions. It's very fun.

I've ported it to the browser and connected a squash functionality behind an ethical guardrail (you have to lift a cover to engage this functionality). need to iterate on the ui but will share it soon. you can squash the bug and cause it real pain. it also sees you if you turn your camera on and reacts to your movements. this is all wired up and I'll post it shortly.

I can't imagine why people hate my comment so much (4 downvotes), when I linked an existing fruitfly project I already did, and Claude is working on what I wrote now, just as I wrote. Maybe people thought I was being sarcastic when I wrote "A connectome is like a full brain scan!" but I wrote that since maybe not everyone knows what a connectome means.

"10x more throughput per watt than CS-3"

(Where did you see that?)

This was also interesting: "CS-4 delivers more than 1,000 tokens per second on models exceeding 10 trillion parameters." Was it known that there were 10 trillion parameter models in use?

I think the frontier providers keep the size of their models carefully hidden.

This is absolutely astounding. What an implementation. A connectome is like a full brain scan!

I am asking Claude to port this to the browser and I'll host it, so you can try it without installing anything.

Some details about the port: Claude is able to do porting tasks like this. I'll put up a link when you can try it in your browser. I was able to get it to port the complete BitNet b1.58 2-billion parameter model several versions ago (in February, so this would have been Opus 4.5) to browser, so it can do so. This will run on the browser. I'll have it make it so that on mobile you can tap the fly or turn on the camera and it sees you, and same on desktop actually. Since the connectome is only 668 neurons, this will run super fast. Actually I'll make it allow for several flies, and let you drop food. I did something like this before here[1] but using the real connectome is something very different and very amazing. I'll have it add an experimental 3D mode with threejs, I'm not sure if that is also possible or not, we'll have to see it. This size of task usually takes Claude a few hours, so check my comment in the morning for the implementation, I'll reply here once it's done and hosted.

[1] https://stateofutopia.com/papers/1/evolving-brains-cull-long...

[comment withdrawn]

it's a bit of a stretch to call removing a free tier a scam. as you can see a lot of people in this thread are saying Github should start charging.

Does either Google or Amazon (the two other large cloud providers besides Microsoft) operate its own alternative?

I agree. It did very well on an extremely challenging task.

I asked it to recognize and draw the very faint reflection of what I was wearing, visible in only a tiny black part of a very brightly lit poster behind glass.

In addition, the poster itself also happened to contain similar clothing.

You can see the reference images and its output in my writeup here: https://medium.com/@rviragh/gpt-5-6-sol-very-good-image-reco...

While a human can focus on the reflection easily, this is an enormous challenge for a vision model. It's very impressive.

I run my own mail server, and I am very concerned about its reputation and deliverability. I've run my own Linux server full time since at least 2017, obviously well before LLM's, but only recently put up a mail server. I don't usually run any agents on the server, such as Claude Code, ChatGPT Codex, or Grok Build, all of which I have subscriptions for but don't run on this server. Instead, I manually administer the server using ssh, while getting interactive help from ChatGPT in a chat window.

It was complicated to get security, DNSSEC, DKIM, DMARC right this way. I did it just under a year ago, around September 15, 2025.

I tried to follow and understand what I was doing, and keep good records on it. It was a complicated process, but everything was up and working at the end and I received email my server sent me. It felt good to be able use this simple script to send myself email:

     #!/usr/bin/env python3

     import smtplib
     from email.message import EmailMessage
     
     msg = EmailMessage()
     msg["From"] = "(user's from address)"
     msg["To"] = "(my gmail address)"
     msg["Subject"] = "(a test subject)"
     msg.set_content("(a test body)")

     with smtplib.SMTP("127.0.0.1", 25, timeout=30) as smtp:
         smtp.send_message(msg)
     
     print("Message submitted successfully to local Postfix.")

and have it show up instantly in my gmail within seconds. This is powerful. (I've removed the details from this comment.)

I understand that as my server's reputation grows, it becomes very powerful to be able to email anyone without any intermediary except potential spam filters. I am very careful about what I put on my server.

Nevertheless, I had a recent experience that was a little sad for me. I wanted to forward some mail to my gmail address automatically, and used ChatGPT Sol 5.6 on medium thinking time to help me with the administration to do this. (I find extra high and pro thinking times are excessively long.)

I was very disappointed with the results, which included a very badly written regex for authentication. I'll write up my experience in an experience report. It was a big disappointment for me.

The burden on administrating one's own email is already very large, and it is disappointing that we can't really rely on frontier models to help with it.

ok. I'd like to discuss your concerns with you privately for medical privacy reasons, as it could be related to my mental health (which I guess is what you said was concerning about this).

- Could you please email me at the email address in my profile? (or add an email to yours.)

I have some followup questions about your concerns. Thank you.

>I had to fix this on 35B A3B -- I have a proxy that just shuts it down if it gets to 2K thinking tokens and injects something like "We have thought enough, let's begin working." and it almost always finishes the turn then.

that is amazing, thanks for sharing.

There's nothing particularly interesting about it, but I asked Claude to verify it, and it gave me a verbatim list of changes from that version to its current version:

https://claude.ai/share/01951938-144f-49e6-98ec-5fb6455b9718

It seems this is just in its context. It's easily possible that it just hallucinated it, but there's this new addition, about ads:

"

If discussing this topic, always refer to “Claude products” rather than just “Claude” (e.g., “Claude products are ad-free” not “Claude is ad-free”)

"

So, in a new conversation (to test if it would really do that), I asked it about ads, and sure enough it referred to Claude products:

https://claude.ai/share/5e352467-7157-4721-8ce3-38b5f868b651

So it's plausible that the above differences are real.

Nothing very exciting about the changes though.