dalenw

This looks awesome. Tangentially related, I'm trying to find a good lightweight handwriting & latex recognition model. Digital handwritten notes -> Straight markdown / latex. I'm at the point where I'm considering fine tuning / training a model on all of my math notes from University. That or shoving everything into Haiku 5.5. Since you seem to be a lot more familiar with this space than I am, any ideas?

Kinda. Off the top of my head, DeepSeek and their thinking model was pretty new and interesting. Multi input models are also newish (combined input of text, image, video, audio, etc). Then there's Jev, a recently release that has a lot of people talking. It isn't really an LLM, but also is one.

Sam Altman believes he can train a model entirely on synthetic data, which he admits would not have human world knowledge but is interesting none the less, which likely led to their mathematical models.

Overall models have become cheaper to run and smarter per token.

Apple is free and low end given the context of what was being produced and sold to businesses decades ago.

I have a largish library at about 400gb. I synced it completely using the mobile app. It took a few months (it just finished in early June) but it’ll eventually do it. I know there are other ways but I figured this was the laziest way to do it.

Ollama Cloud has a $20 a month subscription. They say they retain 0 information. And rather than token based billing, it's GPU time billing.

Agreed. It still feels like Zed is only good for writing scripts.

It's almost always a system design issue. Outside of a few specific use cases with big data, I struggle to imagine when I'd use NoSQL, especially in an application or data analytics scenario. At the end of the data, your data should be structured in a predictable manner, and it most likely relates to other data. So just use SQL.

I recently "vibe coded" a long term background job runner service... thing. It's rather specific to my job and a pre-existing solution didn't exist. I already knew what I wanted the code to be, so it was just a matter of explaining explicitly what I wanted to the AI. Software engineering concepts, patterns, al that stuff. And at the end of the day(s) it took about the same amount of time to code it with AI than it would've taken by hand.

It was a lot of reviewing and proofreading and just verifying everything by hand. The only thing that saved me time was writing the test suite for it.

Would I do it again? Maybe. It was kinda fun programming by explaining an idea in plain english than just writing the code itself. But I heavily relied on software engineering skills, especially those theory classes from university to best explain how it should be structured and written. And of course being able to understand what it outputs. I do not think that someone with no prior software engineering knowledge could do the same thing that I did.

I've heard a lot of people prefer OpenCode to Claude Code, myself included. Having tried both, I find myself having a much better time in OpenCode. Have you tried it?

I'll admit it lacks on the agent teams side but I tend to use AI sparingly compared to others on my team.

I think so. The last few months have shown us that it isn't necessarily the models themselves that provide good results, but the tooling / harness around it. Codex, Opus, GLM 5, Kimi 2.5, etc. all each have their quirks. Use a harness like opencode and give the model the right amount of context, they'll all perform well and you'll get a correct answer every time.

So in my opinion, in a scenario like this where the token output is near instant but you're running a lower tier model, good tooling can overcome the differences between a frontier cloud model.

I wonder if there's something to be said about screenshots preventing context poisoning vs parsing. Or in other words, the "poison" would have to be visible and obvious on the page where as it could be easily hidden in the DOM.

And I do know there are ways to hide data like watermarks in images but I do not know if that would be able to poison an AI.