Sociopath Community

My job has changed a little bit because of AI. I do less coding, but I've been orchestrating most of my research through other people already, so I've had a Physicsal Intelligence already, though now my minions have their own AI slaves. Ive got a bunch of mindless admin entirely automated. For example I asked it to make a cli to push agenda to our meeting page, and asked it to sort the annual funding applications to a centralised place.

I also got it to make me a tailored email filter script for aerc. My inbox has reversed its course, I suddenly get less email.

I tried to first automate tasks by giving them to automated AI but I had less success with it because I had to check the output and it sometimes acted non-deterministically. However I found more success in getting it to make a mechanical/deterministic script/cli to do the task, sometimes less efficiently but I know exactly what it does so I dont need to check the output after a few times. 

For coding, I use it sparingly. I still find that the code it makes looks spaghetti and illogical. For the code that is boilerplate and I dont need to take a look at, Ive used it a lot. 

I encourage my students and staff to use AI though. Whenever they have a question that can be answered by AI, I tell them to ask AI first, then group members, then me, and then go to the library to borrow a book, either University library or the commercial genesis library which lives closeby. I used to tell them to Google it, now its AI. Though I also teach them about the dangers of using AI, which is totally missing from this thread cause of the hype.

Good said: 

 

It wasn't the Claude ai model source that leaked; it was the Claude code harness the model uses. It's basically a set of tools and prompts for the AI.

You can use open code with Claude models too; that's another harness for AIs, which you can use with many different AI models. I used that when there was a hack to basically use it for free (legal hack), but they patched this a few months ago, so now I use Claude Code, as it is better.

I do have the source downloaded as well, but I don't really have any use for it now.

 

So you can't really run Claude without Anthropic giving you access to their models, because you need their models and they are on their servers; they have never been leaked.

What I am doing is making an infrastructure to run local models. Google, for example, lets you download its flagship models for free.

But they need a lot of RAM and GPU/CPU power. If I load a big model rn, it will take half my RAM and won't run as fast (tho i have a decent CPU and GPU, but I want to use those for other things; I would have to stop everything).

I can run them right now using open code, as it lets u use local models, and I can update the downloaded leaked Claude code harness to allow for local models too, but I need a better PC just for running AIs. And I like to code in C#, so the infrastructure I mentioned is that I am making a C# local model loader for C# projects.

And I will need to create a web service that lets me use my AI on the special AI computer over the internet, much like how current AI services do, so the local model will run on my power machine, while the code it creates is on my local machine.

That is the plan.

 

And I think Google's model is pretty good, + there are other free models too, and they will only get better. Claude is still the best for now.

Google released its model for free because its model runs very well if you use other Google products, so it's a marketing tool, while the other big AI companies do not have Google's environment, so they can't release it for free. And I think only Grok has the correct AI price; the other companies underprice their models to create a customer base, but that can't remain for long.

 We can get away with 8GB VRAM with 16GB RAM for smaller models but it's kinda slow. Pulled it off with a flimsy Celeron processor for something as small as trading. 

Friend is a Claude head and I told him not to use it for his Open Claw build, it doesn't always have to be Claude for everything but he's too much of a Claude Fan. At first things were fine cause it was using Sonnet, but then he wanted it to use Opus 4.6 and that's when all hell broke loose and it never wanted to work properly and Anthropic started eating his tokens alive, so I reversed it back to Sonnet. We never added anymore Ai to his build. He has a Ryzen 3200G 16GB RAM and 8GB VRAM. OpenClaw and Claude Sonnet ran fine once we got it up and running. The initial startup the agent would take a minute or 2, but after that it responded fine. This guy is so stubborn, he'd rather run Ubuntu in Windows and he wants Ai to do video editing for him, which I'd say is a really bad idea. 

I'm switching to Hermes 

Price of RAM is coming down slow so I dismantled the Ai build and will bring my PC back to 32GB RAM, with two GPU's bringing it to 8GB and 8GB of VRAM. Sadly they can't work together but they can be used for 2 agents.

i7-5820k 32GB DDR4, 8 and 8 GB VRAM. I'm using an X99 Deluxe Board so I can beef it up to 128GB DDR4 RAM, but damn, at that cost I'd rather get the most expensive GPU.

I'm eyeing the RTX 3090 as it's the cheapest of the RTX XX90 units and it comes packed with 24GB VRAM on board. Thinking getting it used retail is best, though I see people selling them for a little cheaper. Cheapest I've seen is 600 used. 

GPU prices should be coming down too so buying used would be best at this point. The Ai bubble is bursting. 

I also have a feeling these might be the last days of GPU sales. They've become something else now, and companies like Nvidia know they can just build warehouses and charge a subscription fee to handle everyone's graphical or Ai needs and probably make more money.  

when you run claude models, they dont run on your PC, so your specs can be like 1GB of ram and shittiest GPU and it will work just fine. Even with open claw and the leaked harness, you dont run the models on your PC, because the models were never leaked and were never part of the harness. Open Claw or open code or claude code, lets you query the Anthropic servers to use the claude models on their servers. The claude models then get access to your machine and your tools via the harness, but they run on the Anthropic servers, not on your PC. On your PC only the tools run and give data to the models so they can do their thing.

But a bigger model that you do run on ur pc can need 32GB of ram just to load in the memory to be used.

Idk what specs u need for the CPU or GPU to make it run fast tho, i havent looked into it yet.

Sonnet is bad at difficult tasks. When I update my projects each update requires a heavy investigation of the current code, of project and global rules, custom tools ive made. And opus is way way way better at it. Opus will detect when it can reuse code, it will use patterns correctly, it will find gaps in your prompt because u cant think of everything. And in plan mode it will go even deeper, it will find patterns and conventions you have established.

Jada said: 

My job has changed a little bit because of AI. I do less coding, but I've been orchestrating most of my research through other people already, so I've had a Physicsal Intelligence already, though now my minions have their own AI slaves. Ive got a bunch of mindless admin entirely automated. For example I asked it to make a cli to push agenda to our meeting page, and asked it to sort the annual funding applications to a centralised place.

I also got it to make me a tailored email filter script for aerc. My inbox has reversed its course, I suddenly get less email.

I tried to first automate tasks by giving them to automated AI but I had less success with it because I had to check the output and it sometimes acted non-deterministically. However I found more success in getting it to make a mechanical/deterministic script/cli to do the task, sometimes less efficiently but I know exactly what it does so I dont need to check the output after a few times. 

For coding, I use it sparingly. I still find that the code it makes looks spaghetti and illogical. For the code that is boilerplate and I dont need to take a look at, Ive used it a lot. 

I encourage my students and staff to use AI though. Whenever they have a question that can be answered by AI, I tell them to ask AI first, then group members, then me, and then go to the library to borrow a book, either University library or the commercial genesis library which lives closeby. I used to tell them to Google it, now its AI. Though I also teach them about the dangers of using AI, which is totally missing from this thread cause of the hype.

It will produce bad code, if you dont give it rules about it. I have been making coding rules over 5 years and it writes code similar to how I write it. Sometimes i have it refactor the code to check if its up to the rules, as it sometimes forgets a rule, but its rare and small.

Unit tests are vital to ensure the AI doesnt do regressions. It cant remember everything and it will forget what it was doing in a prev session, so I have everything in a unit test.

You've been coding AI for 5 years? 

I remember that you said that you hated AI like a couple of years ago.

I got copilot since almost the launch because I had people from Flatiron recommend it to me.

The reason I say it's spaghetti code is not so much because I don't like the way it's written. The problem is that most of the time I don't have a pre-defined 'try this and check' criteria. For example if I need to make a website, it's incredibly easy because I can specify how the website looks, there is plenty of training data, and I can specify how it needs to function and it can test that it functions in the way that I specify. Or, in a physics problem, I can ask it to code up a solver for evplving the hydrogen atom and I can ask it to ensure it can reproduce some textbook examples; it will have no trouble.

The problem is, say, I need to code up a solver for an unsolved physics model. If the problem is solved in the literature or can be validated against a benchmark, it can do it, but if not, it doesn't work. I need to break it down to very small pieces in order for it to do a good job, and even then it doesn't quite do it in the way I want it to, and usually messes up. By the time I've specified the problem in enough detail, I've already solved it. Reviewing code by AI in that case feels more like reviewing C- level student work. 

The reason is because the way LLM works is by interpolation between concepts, given a prompt. It's limited by Bayes:

P(word|prompt,training)

It uses transformers to connect concepts in a soup of training data. As long as you work in the domain of validity of the training data, it monkeys those who wrote the training data exceedingly well, including interpolation between concepts, which is where most of the current originality stems from.

I'd say that an AI is a super monkey, it's even better at copying people than I am. However, the main limitation is that monkeys cannot do what they cannot see. 

That's why I'm betting the next stage of AI must be robotics, so they can explore the world by themselves and test the nature.

Good said: 

when you run claude models, they dont run on your PC, so your specs can be like 1GB of ram and shittiest GPU and it will work just fine. Even with open claw and the leaked harness, you dont run the models on your PC, because the models were never leaked and were never part of the harness. Open Claw or open code or claude code, lets you query the Anthropic servers to use the claude models on their servers. The claude models then get access to your machine and your tools via the harness, but they run on the Anthropic servers, not on your PC. On your PC only the tools run and give data to the models so they can do their thing.

But a bigger model that you do run on ur pc can need 32GB of ram just to load in the memory to be used.

Idk what specs u need for the CPU or GPU to make it run fast tho, i havent looked into it yet.

Sonnet is bad at difficult tasks. When I update my projects each update requires a heavy investigation of the current code, of project and global rules, custom tools ive made. And opus is way way way better at it. Opus will detect when it can reuse code, it will use patterns correctly, it will find gaps in your prompt because u cant think of everything. And in plan mode it will go even deeper, it will find patterns and conventions you have established.

True. However, for my buddies build he was using the Ryzen 3200G, which is a pleb CPU and when Opus 4.6 was implemented it started sending OpenClaw instructions that would start bottlenecking the CPU.

[Claude Opus 4.6 Cloud] ---> Sends complex, multi-step instructions (JSON data)
|
v
[Local Ryzen 3200G CPU] <--- ❌ BOTTLENECK OCCURS HERE
• Parsing massive JSON token payload
• Launching containerized Docker sandboxes
• Running Chromium/Playwright browsers
• Compiling local code & tracking tool states

Only the Opus 4.6 would do that.

Gemini confirms....

"Heavy JSON Parsing & State TrackingModels like Claude Opus 4.6 utilize adaptive thinking, massive 200k+ token context windows, and deep multi-step planning. When Opus returns a dense structured text payload or calls up to 9 subagents simultaneously, your local OpenClaw gateway must parse that data. The 3200G's weak single-core performance struggles to crunch this data fast enough, causing the software to freeze or time out."

 Switch to Claude Sonnet 4.6: Claude Opus 4.6 is notoriously heavy, slow, and expensive for everyday agent workflows. Pointing OpenClaw to Claude Sonnet 4.6 cuts costs by 80% and processes tasks with significantly faster, less bloated local automation loops."

Yeah...  We learned this the hard way. When we switched to Opus 4.6, our agent got choked out. We'd see him typing but no answer. We keep testing and trying to fix it in the powershell, nothing worked. All those tests we did costed about $20 real quick. I just returned it to Sonnet and the agent came back to life.

While the brain is on Anthropic servers, the formatting of Opus's instructions require the CPU to be more beefy, my PC can handle it but I still wouldn't use Anthropic Ai for OpenClaw or Hermes cause they charge extra when we use it like this. I can see why too, cause OpenClaw can be used to turn Claude into a discord Claude bot where everyone in the room has access to an agent handing out paid Claude.

This is my friends build so he went straight for Claude, he's also a bit of a stubborn knucklehead and wanted to take over what I was doing for him, but the idea is to install multiple local LLM's for his OpenClaw.

Namely, Ollama....... Qwen 3.5 Coder 7B / 9B........ Llama 3.1 8B .........Gemma 4 8B. 

And yes, with the local models running in house, that's when the PC hardware will start ramping up. I prompt it and you hear the GPU fan starts shhhhhhhhh, then when it's done it goes idle until the next prompt/task.

My friend is such an ogre, in the end he got mad and he's pointing at his PC and he's like "All I know is this isn't powerful enough" cause he still insists on Opus being the #1 Ai for his agents, and I'm like dude, just let OpenClaw decide which brain is sufficient for the task, and probably ban it from using Claude.

Jada said: 

You've been coding AI for 5 years? 

I remember that you said that you hated AI like a couple of years ago.

He said he's been making coding rules for 5 years. Rules for himself.

This would mean Good created a prompt he keeps on a notepad he will paste into Claude to prime it for every coding session, basically telling Claude to follow his guidelines while generating code. Kinda makes me giddy. 

Ai wasn't this good at coding a year ago Legga, so of course Good is digging it now.

I got copilot since almost the launch because I had people from Flatiron recommend it to me.

Try Claude.

Cursor Ai. 2nd

Claude is better than Cursor. 

.

.

.

Grok 5 is going to be a coding beast. Coming out in June.

Grok is bound to be the best, since it's the native Ai for the Colossus super clusters.

I have both claude and cursor and most of my student use one of the two (or codex); I had ai in my terminal since a long while ago. I started with copilot, before chatgpt, but I've moved since.