Andric

Update to this: having built two apps with it I think TanStack Router + Convex is a beautiful combination

What a simple yet powerful stack. Works great with LLMs and easy for humans to grasp

what’ll it do?

Postgres setup and deployment aside

I just watched the quickstart walkthrough and loved how easy it was to actually use on the client-side

Query and mutate almost like its synchronous state!

Feels worth it to go through the setup for that ease of use

The Postgres setup and deployment of the cache server seem tricky

Also not sure how much this will all cost at the outset or what it takes to scale gracefully

But I think it’s worth it to get going in dev for now.

Rocicorp’s Zero is pretty amazing from what I can see.

It’s in alpha so I was a bit cautious to use it but I’m very tempted now

It would replace 4 layers of my stack:

Prisma ORM
ZenStack (authorization)
tRPC
TanStack Query

And provide realtime multiplayer for free.

On the other hand I really need to rethink the stack I’m using to build my app (that’s authenticated, so I don’t need SEO features).

Maybe TanStack Start?

I’m currently using the T3 Stack with Prisma, ZenStack, Clerk, and Inngest.

Need to figure out how to get it working with all those pieces.

Okay, been a few months since this post.

Ended up going with TinaCMS. Pairs well with Next.js App Router.

I hated working with the App Router when trying to build an app (slowed me down so much), but building a content site with it is such a breeze.

Had so much fun building with this stack.

Sounds like a Facebook Page or Group from the early 2010s

AI will make it so we can organize around purposes more easily, because you don’t need to be an expert to operate these tools when an AI model can assist you.

Maybe this is too idealistic, but I think humans specializing around the ends and not the means leads to more meaningful work.

I don’t think AI will replace jobs. I think it will likely change the way specializations are re-organized

For example, in tech, specializations develop around different tooling expertise (UI designer, React developer)

But humans tend to organize around purposes (marketing, R&D, ops)

Much decision-making is tacit or qualitative. Making our OODA loops more legible is the first step towards observability & steerability. AI systems might help us codify this nebulous process. But how will it be trained to do this, and on what data? We don’t have good primitives or tooling for this.

Yeah I agree with you there: dialogue is important. What I want is dialogue that’s contextualized with the actions I’m taking.

A lot of what we do tacitly imply what we mean. If these tacit decisions and ideas aren’t made explicit, it’s impossible for an agent to really understand you well enough.

Current “AI agents” aren’t like this at all. They do things on your behalf in ways that their developers decided they should do.

An AI agent should instead do things exactly how I would do them.

Do as I do, not as I say.

Like my personal apprentice.

Just like training an apprentice, a true AI agent should amplify your abilities to the limits of how much compute you can buy, so that you can do a lot more with the same amount of time (which is true productivity).

*understand

I don’t really want to prompt an AI system.

Instead I want it to passively watch what I do as I do things, talk to me to understand what I’m thinking as I do them, and progressively learn from me.

Over time, it should understands tacitly what it’s like to make certain decisions, and then do them.

Statistical prediction of the sort that deep learning models do (where past data predict future data well, i.e. ergodic) isn’t quite the same as predicting socioeconomic developments, though.

The latter is non-ergodic.

The other thing is that the nature of said upside is more “downside protection”. You don’t really gain anything. Instead it prevents losses.

And the nature of this prevention almost always probabilistic.

In many ways it’s similar to insurance.

Which is hard for people to grasp.

The issue is that the upside is collective at the population level (which is hard to intuit) and the downside is at the personal level (which is more easy to imagine)

It’s a smaller quantized model (q2_k). The bigger ones don’t work on my computer.

Mostly I just want to see if I can use it for local coding. I have Perplexity Pro for chat, already

Managed to get it to work with @continue.dev (using @lmstudio-ai.bsky.social). Will use it more and see how it goes

Sadly I don’t have a beefy enough computer (which is an M4 Pro MacBook Pro with 24GB RAM) to run Llama-3.3-70b-Instruct.

But I was surprised it was beefy enough to run Qwen2.5-Coder-32B-Instruct at an acceptable speed!

9 tokens per second isn’t as fast as ChatGPT, but usable.

Not familiar with it!

What does it help you do and how have you found it so far?

Oh this is a neat one!

I do a similar thing with my hand drawn notes (on an e-ink tablet) and turn it into text.

I find scribbling notes on a freeform page much more conducive to creative thinking than typing

but text is more easily searchable and readable, so this closes the gap

What other boring but useful ways to use AI have you found?

Just tried this in Perplexity Pro (with GPT-4o) and was worried it might hallucinate text that wasn’t there.

But I double checked with the PDF and it got it!

I know, it’s nothing groundbreaking. But I was happy to be able to read a PDF on my phone now!

So I prompted it:

I haven’t read this PDF. I want you to help me summarize it and walk me through chapter by chapter.

First, list all the chapters with a short summary.

Later, when I tell you to continue, start responding me with each chapter’s text verbatim. Do not miss any details.

A boring but useful way to use AI I discovered today

I like to read on mobile but sadly PDFs have always been problematic

But then I thought, what about “reading” it with AI?

Maybe I can upload the PDF I wanna read and have it spit it back out to me

It takes a bit of work for me to get good sleep. If is slip up on sleep hygiene the night before then I’ll sleep poorly. It sucks.

Recommender algos don’t have a concept of “quality”:

If it uses popularity as a proxy for quality, then it’s at best uncorrelated, and at worse inversely correlated

Whoever chose music for GTA radio stations should finetune the Spotify algo. Since it’s their only good algorithmic playlist!

Listening to Spotify Wrapped

And I realize I’ve been using this product since 2012, yet the algorithm still doesn’t really know my taste well enough to recommend me songs!

I have an eclectic taste and I’m okay with 60-70% of genres. I just want the best of those genres.

Think: GTA radio stations