Isaac Estabrook

If you're in the Austin area working in tech or startups, or if you want to talk about how automations could help your small business, please reach out.

In about a week I’m moving to Austin for the summer. I studied the economics of AI in school, but I want to see what happens when you go beyond the papers. The plan is to build quickly, meet people, and see what happens. Currently prototyping this research dashboard ↓

Can someone at @AnthropicAI please stop the model from saying "genuinely" in every single reply? Claude's obsession with the word "genuinely" is exhausting. Mine continues to say it even if repeatedly told not to in both custom instructions and in chat.

Codex computer use is pretty cool. Will do more extensive testing after finals but the fact that it can just figure out how my apps work is a big step in making useful agents that anyone can use.

@Dale_Chu @EdunomicsLab How much do you want to bet that higher spending is actually anti-correlated with better outcomes?

K–12 spending just topped $1 TRILLION. Per-pupil: ~$18K and rising. Outcomes? That’s a different story.

Private school and charter school enrollment are holding steady. It's not housing costs - it's bad schools.

@garrytan Bay Area schools stopped teaching algebra for 12 years. Grade inflation, "discovery-based" learning, "self-esteem;" all of it revealed during the lockdowns, which teachers lobbied to continue indefinitely because they didn't want to work. Maybe that's why.

California will lose 15.7% of its public school students by 2031. Nearly 3x the national rate. Idaho and Florida are growing. This isn't an inevitability — it'a consequence of what happens when housing costs too much and schools stop working. gli.st/2rmnc27p

Update: my ACT practice set generator is working! My Codex workflows autonomously categorize problem types from uploaded PDFs and generate targeted homeworks based on identified weak areas.

Using Codex to extract questions from ACT Math problem sets. My goal is to automatically generate the perfect practice set given a student's weaknesses. I'll start taking students soon :)

@FixingEducation Yes, lots of schools spend on useless edtech and useless edtech should be cut. But it represents a relatively small piece of the budget. Of course, the countless private schools operating at half the cost per pupil and scoring much higher prove that cost isn't the constraint.

Many districts don’t have a funding problem. They have a spending strategy problem. Smarter decisions with existing dollars would go a long way. Instead, they’re spending millions on things that don’t move the needle.

@BioLayne I like how people are saying "incredible strength at 44" as if it's not incredible strength at any age. Keep it up!

Squatting 625 for 2 at 44 years old PED free 4 weeks ‘til nationals. BRING IT

What's next: I think it would help to move off of Telegram. Trying to manage everything in one chat probably introduces tons of context that only distracts the model. I think I'll switch to Discord, with channels for each project. Heartbeats might restart a stalled model. (4/4)

Bad news: The results leave a lot to be desired. Qwen seems to have trouble remembering what it was doing, and it repeats itself a lot. Here's what my Telegram looked like in the morning. It also gives up. It worked for about 30 minutes and didn't have much to show for it. (3/4)

Some successes and failures building autonomous agents on local models (1/4): I've been trying to get my OpenClaw to work on tasks overnight using local models. I run a VPS that accesses my models via wireguard tunnel. I run Qwen3.6 35b for planning and Qwen 3.5 9B subagents.

@bathroomtile3 @0xSero Maybe - I tried to do that, but it seems like there's some global setting for the agent timeouts that's separate from the main model's timeout, and it's not letting me change it. Maybe something on Ollama's side 🤷‍♂️

@IsaacEstabrook @0xSero Remove the timeout? I'm new to this but I imagine that's what you're having issues with because I did lol

Actually useful local models to run on < 1000$ - 24GB vram (4090, 3090) - 24GB Mac Qwen3.6-35B & Gemma-26B = speed Qwen3.5-27B & Gemma-31B = quality Zeta-2 - Cursor tab style Parakeet - Speech to text Hermes-4.3-36B - No Refusals

@shaneparrish I've been looking for an easy startup - happy to burn tokens doing nothing on behalf of any business.

Countries that don’t want to get left behind should consider a 100 percent personal income tax write off for ai subscriptions and a 200-400 percent write off for businesses.

@AnthropicAI Surprised by how similarly most countries score in terms of AI sentiment. Obviously there's a selection effect - only Claude users took the survey, and those who took it probably skew even more positive than average users. Regional variation in the US is also interesting.

Last month, we published our look into what 81,000 people told us they want from AI. In new research, we’ve investigated the economic hopes and worries referenced in their responses. Read more: anthropic.com/research/81k-e…

This looks like bad news for AI Automation agencies, and maybe it is. But always remember substitution vs. scale effects. Yes, this could substitute for your skills in a significant way. But it will also expand the overall market - you may find you have more work than ever.