This is me replying to your tweet

We need a futures market for tokens

There’s millions of use-cases for small specialized fine-tuned models. In many real evals Claude scores in the bottom tier or can’t get above 0. Many times it’s because duration and/or cost is part of the eval. I think a lot of the current discourse around models ignores time and cost but they are the two biggest considerations for real businesses.

I don’t think you can take any claim of a software factory seriously if there isn’t an industrial engineer on the team

We are beyond the point where the majority of the Frontier LLM’s training data for coding is obsolete. Without websearching and reading docs, almost all their references are for deprecated versions. RAG is more important than ever.

Needle1 was great, Needle2 was better, Needle 3 is awesome. Context engineering combining the Liquid AI Nanos and Cactus AI Needle boosts my ai outcome acceptance rate go thru the roof.

Some of you aren’t using the cia brainwashing techniques in your prompts and it shows

Coding agents’ training is so out of date they make assumptions and use their judgment based on a stale world model, which IMO results in a lot of the nasty hand-rolled code in lieu of just using built in features. And the big labs think you shouldn’t use ‘agents’.md or ‘claude’.md files and let the model use its judgement. Things this week that frontier models didn’t know about: Typescript 6 had been released (and TS7) PNPM & corepack Anthropic bought Bun SpaceX, xAI, X merging Postgres 18 That Zod4 existed and has JSONschema built in That Kubernetes can scale to zero If I let the models use their judgement, that judgement is stuck in 2024

I need a Thunderbolt network switch

I am so tired of coding agents trying to talk to me and make this a collaborative process

The only book Anthropic didn't include in Claude's training data

Accidentally just broke my streak but also just crossed 75 Billion tokens on my personal Codex account

Agentic Simming

Why did anthropic train claude to not read the docs

Everyone I know who isn’t in tech has tried chatgpt but doesnt pay for it, has maybe heard of claude, and uses Gemini daily. My mother-in-law was talking to Gemini voice 30min ago in my living room

Kind of infuriating how Anthropic will degrade from Fable to Opus if you mention "distill knowledge". It is reading notes from a meeting where we talk about distilling the knowledge from a bunch of spreadsheets and meetings into a presentation, and it triggers the safeguards.

Introducing Kids Paint Studio, brought to you by Buffering Studios, Inc. 7 (and counting) massive, entirely offline coloring catalogs for kids. No in-app purchases, no ads, no accounts, and absolutely *zero* data ever captured or recorded (not even Sentry/Bugsnag/Posthog analytics). There is no backend API. You buy it once, download the entire catalog and own it forever. --- This started with my kid coming to me every 5 minutes to clear an in-app purchase screen after clicking an ad intentionally designed to look exactly like the coloring app he was using. So I decided to build him something that didn't suck or try to monetize his little creative brain for "one-more-sale" Turned out he liked it. A lot. Like a lot a lot. And I liked it a lot a lot because I could trust it. So we kept going. My wife and I experimented with collections built around his favorite stuff. We figured AI could help us generate images and learnable information for an app we could trust. An app that tried its hardest to keep him inside it, with no deceptive in-app purchases or external links. It was tough. AI is not good at consistent image gen. After plenty of trial and error, we dialed it in for about 90% of the content. For the other 10%, I built per-catalog debug controls so we could review everything between the daily bottles, activities, and meltdowns. Funny side story: most of my wife’s animal-catalog reviews asked AI to change several scenes where the animal's "private parts" were really in focus. Nature is fine... GOOD, even.. and those parts exist. But AI's choice to randomly put them on such display was kind of bizarre. The funniest part? AI refused to make her changes because her requests included words like (change the pose to hide the) "penis". Instead it just noped out and refused to make any edits But I digress… The next big hurdle was the app stores. The apps were ready in early June, but we're still fighting to get them all approved. The App Store and Google Play keep rejecting our catalog builds as "spam" because each app uses the same driver code. Meanwhile, the current garbage kids apps littered with ads and in-app purchases are totally fine. Okay, sure. I understand what they’re saying. The app engine is the same. But the engine isn’t the product! It’s also not where the time, blood, sweat, and tears went. That’s in the catalogs: their pictures, information, organization, and the work required to make everything informative, educational, appropriate and not obscene. Every time I've pushed back, we've received another rejection. The repeated recommendation is to consolidate everything into one app and sell the additional collections through in-app purchases. But we're not going to do that because that's exactly one of the issues we’re trying to solve for. Or maybe they do realize it and that model isn't nearly as profitable. Maybe that's the actual beast we’re fighting when trying to build child-friendly apps that aren't dopamine-sucking slot machines. Something something show me the incentive and I'll show you the behavior. I don't know, draw your own conclusions So, while we originally planned to release every catalog at once, it looks like we’re starting with two: Bible and Heavy Equipment Bible will forever be free. It's a trust piece that lets parents try the experience. It shows you exactly what you'll get from every app we build. And if you don’t believe me, just ask Google Play or the App Store. They’ve made it very clear the engine is the same Since this whole ordeal started, we've had some really cool early-education ideas if we can get traction. Not the normal "replace me and teach my kid because I’m busy" tools, we want to build "we're actively teaching our kids and could maybe use software to help us" tools Hoping we can get approved and release the other catalogs soon! LIKES, RETWEETS AND TELLING YOUR PARENT FRIENDS IS GREATLY APPRECIATED! Godspeed, fellow parent soldiers

Idk how other people use codex, but this is how I use codex. I like gpt-5.6, it actually follows instructions really well.

Using Gemma4-31b via Cerebras in opencode is insane. Watching it complete a task that gpt5.5 or Opus would spend 5min+ on in 8 seconds is mind boggling. The compaction time is measured in milliseconds 😅

> dev work happens on remote linux boxes This is the way.

Claude doesn’t use the .agents/ folder so we make a husky script to sync the project skills between .agents and .claude Then half the other harnesses decide they want to have backwards compatibility with Claude, so they load .claude and also .agents. So now some harnesses which don’t deduplicate have 2 copies of every skill. 😒 I think opencode is the only one that properly handles this. right now colleagues using Codex and cursor all have duplicates.