mindcrime

Solar panels? You can go down to Harbor Freight right now and pick up a setup that will run a laptop or tablet for a while, for a couple of hundred dollars. One wouldn't have to spend a ton more to get a setup that can run a decent PC with a GPU for at least a little while.

I mean, you're probably not going to run it all day, especially not in the part of year with fewer daylight hours, but anything's better than nothing. Plus the same solar setup can be used to power other useful things.

"They only do as a social contract, as long as power structures agree to enforce them."

is absolutely true. I've said many time "The Constitution is 'only a piece of paper' at the end of the day." Ultimately it's down to us, "we the people" to enforce this stuff.

Unfortunately the culture in the USA these days is sick. And not just a little bit sick... We are truly staring into the abyss right now.

Trump is dumber than box of brick bats.

You have sinned.

Your penance:

Three Hail Claude's, one Our GPT and an act of loop engineering.

Two minutes of silent prompting.

Disclaimer: I am not an economist, or a sociologist, etc. So this will be, at most, some random guy playing around with models with some help from an AI. And that's even if I find time to actually put into it.

I don't know. I was at my doctor's office a few weeks ago, and the topic of AI came up during our conversation[1]. I expressed the opinion that her profession should be fairly safe because it's so highly regulated and how doctors have the AMA to block new entrants. She said that her and her colleagues had literally just been talking about how they are worried about their jobs, and pointed out that with profit-obsessed private equity firms taking over more and more medical practices, she thought they very much would be trying to replace human doctors with AI. And she alluded to the AMA being a much less effective lobbying body than what I, as an outsider to the field, had perhaps believed.

I dunno. That's just an n=1 anecdote, but if nothing else it shows that (at least some) doctors are not just sitting back and resting easy, assured of a future of high paying career options. And this is a lady with a degree from Yale, who is a Doctor of Internal Medicine, and working for a very large and prestigious medical practice. Not some half-educated nitwit at a random back-country medical practice.

[1]: I mentioned that I'd asked ChatGPT about some test results, and told her what it told me. She replied "Yeah, I agree with that." Then chuckled and said something like "God, did I just agree with ChatGPT?" That's what started the whole sidebar about AI.

I mean, it can be that way. But it does't have to be. The Harness can be as much the brain as the LLM. You have nearly unlimited flexibility in what you do in the harness, and you can add as much of the overall "intelligence" of the system as you want in the harness.

To illustrate, you can:

1. Use multiple "brains" / LLM's, and do routing between them using a specialized model (a Jev like model, or something of your own design)

2. Use planning tools like pyperplan, jupyddl, pddl4j, etc.

3. Implement your own "memory" system using... pretty much anything you want. And I would argue that a lot of the apparent "intelligence" of a system involves how well memory works

4. Call classical machine learning models to solve specialized problems

5. Use reinforcement learning to improve performance in specific areas

6. Implement a blackboard for multiple knowledge sources to collaborate.

7. Use inference engines using forward/backward chaining, RETE, etc.

8. Use your own hand-rolled abductive inference strategy

9. Implement your own Bayesian reasoning techniques

10. Use custom causal inference code

... and so on.

Don't get me wrong. I'm bullish on what LLM's can do, but I posit that the overall expressed intelligence of a system can involve a lot more than just the LLM. In fact, I'd actually make the slightly stronger claim that, depending on what you're trying to do, an LLM might not even be required at all.

I do have it configured to use the Github MCP server and have given it access to 3 or 4 of my repos. Along with that, I put an AGENTS.md file in each repo and gave ChatGPT in general a custom instruction that says "at the beginning of each session, or when prompted by the user, consult the appropriate AGENTS.md file and follow the instructions found there."

I then keep various files in the various repos that catalog our discussions, decisions, speculations about future research directions, yadda, yadda, yadda. And I "incorporate by reference" by mentioning those files in the AGENTS.md. Setting all that up took some work, but it makes the AI much more reliable than trusting it to rely on its "memory".

1. I've been doing a lot of experimenting with neuromorphic and analog computing. ChatGPT has been my research assistant / lab partner throughout and has helped with doing experiment write-ups, analyzing scope waveforms (it's really impressive how good it is at parsing screnshots from the scope, using the graticule / divisions, and other on-screen indicators to figure things out), finding reference material, and providing general electronics advice. I've learned a TON of new stuff about using my oscilloscope, using some features I'd never used before.[1]

2. Along with that, I'd long wanted to get my lab all set up to use LXI for remotely controlling all (well, "most") of my test equipment. Once I did the work of physically connecting all of the instruments to the LAN, ChatGPT wrote most of the first cut of a batch of Python programs using PyVISA to control the various instruments (oscilloscope, function generator, power supply, multimeter, programmable load) and to download data from them.

3. I'd been "meaning to learn KiCAD one day" for about 15 years. About a month ago I said something to that effect to ChatGPT and it handily replied "I can help you with that." About 6 or 7 hours later I was sending Gerbers to JLCPCB with my first custom PCB design, and 4 days later I had my first boards in my hand. Soldered one up and it worked first time! Yay. Now granted, the circuit was VERY simple (a 555 timer blinking an LED) but the point was to learn the tool and the workflow, not the actual end product. About two weeks later I started on my second board, planning to do it with much less AI assistance (I worry a lot about cognitive abdication / brain-rot / etc). But right out of the gate, I ran into the point that KiCAD did not have a symbol for an obsolete component I was using (the 2N6027 PUT). So I turned to ChatGPT again and it helped me create a custom symbol. After that little diversion, I went back to the schematic and board layout. This time around, it only took about 85 "turns" with the AI to get to "Gerbers sent to JLCPCB" whereas the first time it took about 120 "turns". So I'm feeling pretty good that I am actually learning and retaining some of this knowledge and not "just following instructions from the AI". I do hope to get to the point soon of doing a board with NO assistance from the AI whatsoever.

Those are the ones that stand to me. There's other stuff of course, but those are ones that have really made a difference for me and/or really impressed me.

[1]: One thing I did that helped a lot, was creating a "Project" and uploading all of the user manuals and programming guides for my various instruments. So the GPT always has access to the actual documentation and doesn't have to rely on its "memory".

1. Since people are still going to find a way to buy items the FCC doesn't want them to have, what's that new outlet going to be? I saw somebody mention TikTok. OK... what else? Facebook groups? Craigslist? Private old-skool phpbb forums? Usenet? Mastodon? AliExpress? Something that hasn't been invented yet?