Comment by lode

Just watch a few videos on Clawdbot. You'll invariably see some influencer's Anthropic key, and just use that. Wokka wokka!

If you have an old M1 Macbook lying around, you use that to run a local model. Then it only costs whatever the electricity costs. May not be a frontier model, but local models are insanely good now compared to before. Some people are buying Mac Minis for this, but there's many kinds of old/cheap hardware that works. An old 1U/2U server some company's throwing out with a tech refresh, lots of old RAM, an old GPU off eBay, is pretty perfect. MacBook M1 Max or Mac Mini w/64GB RAM is much quieter, power efficient, compact. But even my ThinkPad T14s runs local models. Then you can start optimizing inference settings and get it to run nearly 2x faster.

(keep in mind with the cost savings: do an initial calculation of your cloud cost first with a low-cost cloud model, not the default ones, and then multiply times 1-2 years, compare that cost to the cost of a local machine + power bill. don't just buy hardware because you think it's cheaper; cloud models are generally cost effective)

Huge pyramids are built of relatively small blocks, kudos to everyone contributed.

Yeah, I looked at Clawdbot / OpenClaw at the beginning of the week (Monday), but the token use scared me off.

But I was inspired to use Claude Code to create my own personal assistant. It was shocking to see CC bang out an MVP in one Plan execution. I've been iterating it all week, but I've had it be careful with token usage. It defaults to Haiku (more than enough for things like email categorization), properly uses prompt caching, and has a focused set of tools to avoid bloating the context window. The cost is under $1 per check-in, which I'm okay with.

Now I get a morning and afternoon check-in about outstanding items, and my Inbox is clear. I can see this changing my relationship to email completely.

That's the sad thing. There are so many millions of talented under-employed people in the world that would gladly run errands or set up automations for you for $200-$1000 per month or whatever people are spending on this bot.

Developers trust lobsters more than humans.

The other wild thing is that many of these expensive automations that are being celebrated on X can already be done by voice using Siri, Google, or any MCP client.

I had the same problem. Ask Clawdbot to optimize token usage. It cut my usage in half.

Can't you just point it at a local ollama? It'd be slower, but free (except for your electricity bill).

I think one thing these things could benefit from is an optimization algorithm that creates prompts based on various costs. $$, and what prompts actually gives good results. But it's not an optimization algorithm in the sense gradient descent is, but more like Bandits and RL.

There has been some work around this practically being tried out using it for structured data outputs from LLMs https://docs.boundaryml.com/guide/baml-advanced/prompt-optim...

I won't claim I understand its implementation very well but it seems like the only approach to have a GOFAI style thing where the agent can ask for human help if it blows through a budget

part of me sympathizes, but part of me also rolls my eyes. Am i the only one that’s configuring limits on spend and also alerts? Takes 2 seconds to configure a “project” in OpenAI or Claude and to scope an api key appropriately.

Not doing so feels like asking for trouble.

Would have been $68 on DeepSeek, which is also imho very good.

I still have Opus review the shit out of & plan my work. But it doesn't need to be hands on keyboard doing the work.

you can use your claude max subscription