Neal Sanche

In Hamburg today, having a strange breakfast. Germany is roughly half the size of my home province, with 20 times the population. Sobering, really.

@botteaap Possibly, yeah. New AI users really don't understand that current LLMs are really just fancy autocomplete. They don't technically think. So what they tell us is really just a reflection of what we have put into them. They aren't people.

@botteaap Also, please stop initiating conversations with: Hey, do you have time to answer a few questions? I am going to start saying no.

@virtuous_sloth@cosocial.ca thanks. Yeah Ryan is quite the guy. He's done some really interesting things and I can't wait to see if what he builds here becomes popular. That would be pretty cool.

My brother has an inspiring pitch https://youtu.be/p533cIe6U_Y?si=H61mJmZXKU21kBeq I really wish Steve could see this. I too was very inspired by the openness of the original Apple ][ manuals.

Finally I think I have found a local LLM that works decently with my MCP, Devstral-small-2507. It borks on the odd tool use requests, but it is a very strong model for its size. Quick time to first token too. Now I can really work on figuring out why working with my MCP is so hard for these llms and optimize it some more.

Claude Code plan mode is so, so good. Nothing I hate more than an LLM just going off immediately changing whatever it feels like. Reading through the plans, making changes to the plans, then hitting the go button is quite satisfying.

Cursor has made me so angry with its random quota blocks. CC does this in a much better way, warning you when you are close to a quota, and telling you what time you'll have more quota to burn. Brilliant work Anthropic.

@alsutton@snapp.social @snappautomotive@snapp.social I assume the M4 Pro with 24Gb unified memory?

Lately I have been writing a lot of Flutter for an app that helps people manage a Eurorack module called Disting NT. This has led to building in an MCP to allow llms to help make presets for the NT. This led to trying to fine tune an LLM to know more about Eurorack patching. This led to the realization that local models are really not very practical for this sort of task. But wow, can Gemini Flash use MCPs well. I want something small with exceptional tool use pre trained.

@virtuous_sloth@cosocial.ca That is very cool, definitely up my alley. Thanks for sending that along.