Forget threejs demos, Opus 5.5 is *incredible* at building apps. This is Letters Abroad, a language-learning app I'm working on. You exchange letters with an AI pen pal, and learn a language through context (plus some brain science magic). (also built by Opus, of course) I've been tinkering with this the last week. I built a prototype with Astra, but wasn't happy with the design. Tried Fable, still wasn't good enough. Then Opus 5.5 dropped. In a few hours, I've made more progress on the visuals than in a week of burning tokens. This model is insane. It reads my mind and nails it every time. All the gorgeous 3D models and fluid transitions are thanks to Opus. Crazy idea, prompt, repeat. No more quota anxiety, just non-stop flow state. I'll be writing about my prompts/skills/learnings from building this. Lots to share; this generation of models is absolutely game-changing for design!

I had Astra analyze all my Codex threads and learned a lot about quota. Here's how to get the most Astra usage. (applicable to Fable too) I've spent ~$5000 on Astra since release*: - 70% is JUST cached input reads - 16% is uncached input - 13% is output - <1% is cache writes Till now, my Codex threads kept getting longer and longer, but Astra broke the trend (see chart). Long context is eating most of your quota. The biggest levers: - Smaller threads. New thread for each little task. Don't let context build up - Use smaller models that delegate coding tasks to Astra subagents. Don't let Astra see big outer thread - Avoid continuous polling. Have Astra set timers and wake back up for long-running tasks - Trim tokens wherever you can: skills/MCPs, AGENTS.md, custom instructions, etc - 3rd party tools like RTK that return more compact outputs might help (haven't tested this yet) Investigating that 70% input cost in more detail: - 40% is encrypted content (reasoning or other OpenAI stuff) - 20% is file reads - 10% is tool call arguments, schemas, etc - 7% is skills (catalog, loading instructions, output) - 5% is base instructions (system prompt, custom instructions, AGENTS.md) - 5% is browser/computer use input - 3% is command/tool output - The rest is minor or not actionable: compaction handoffs, assistant prose, user messages What you can do: - Use the lowest reasoning you can for your task, to cut CoT tokens - Make your repo as agent-friendly as possible. Small files/modules with clear folder structure, so Astra isn't grepping and reading large chunks. Try asking Astra to delegate file search tasks to subagents like Luna - Don't dump in tons of skills, instructions, AGENTS.md stuff; be intentional about it - Delegate browser/computer use to smaller models unless you need max performance And, probably goes without saying: - Don't use /goal with Astra. The few times I did absolutely obliterated my quota - /fast mode basically only if Tibo says he's resetting in an hour and you have 100% quota to burn Let me know if you have any observations to share! *I didn't literally drop $5k, this is mapping token to equivalent API prices

These alleged GPT-6 Sol leaks look even better than Astra. We know they have a much stronger internal model. Maybe they’re distilling it? Astra is 2.5x the price of Sol. This quality at Sol prices would be completely insane. Hope we see it this week…

Extremely proud of this article I wrote for the legendary @lennysan! This is the culmination of months of research into why most AI designs are slop, and my top techniques for fixing this. LLMs can be super creative, but as we post-train them more and more, we stifle that creativity. We train them to predict tokens that balance everyone’s preferences. But good design does the opposite: it’s risky, bold, and opinionated. You have to put your LLM in a different state of mind to get this kind of design from it. You have to break out of the average slop, into the fringes of the bell curve where unseen ideas live. This post is all about how to do that. Hope it helps, and let me know if I should go deeper on any of this for future posts!

I still don’t understand what ChatGPT Work is. I barely understand Claude Cowork and Grok Bot. This entire product category doesn’t make sense to me, yet everyone’s racing to build it. Can someone enlighten me? Codex, Claude Code, and Grok Build are useful because they run on my computer, with my files. I need them to do some stuff and they just do it. They use my apps, browser session, etc. “But you can’t close your laptop!” Yeah, that’s annoying. That’s why I have a Mac mini and a gaming PC that run agents too, so I can remote control them and they’re always on. As far as I can tell, all these new solutions are taking that solution and turning it into a managed service, with isolated cloud VMs. That defeats the entire purpose. Now to have it use my apps and files, I have to link a bunch of accounts and git repos. My agents are all juggling branches and PRs just to slop together a throwaway demo. Why? I don’t care! Just edit the file! What am I missing?

Is conspiring to commit armed robbery in Indiana a crime of violence under the US Sentencing Guidelines? It depends on whether the word "conspiring" should be read as "conspiracy". Nope, per Judge Taibleson. Sentence vacated and remanded.