Killer design. Proud of Chinmayi!

I'm judging the Claude Impact Lab hackathon in Seattle on October 25. It'll be good to get out of the AI builders building for AI companies bubble. The teams are building AI tools for small businesses. Looking forward to seeing what ideas they come up with.

My best rated @AITinkerers talk so far (4.8/5), almost perfect! This was my first time presenting anything to game developers. I knew that more than giving a demo about what I built, I had to educate the room on the basics of decision models first and then show them how to apply it to the gaming domain with Doom and Contra as examples. In my close-knit AI-bubble, decision models like Jev seem to be fairly common. But, I need to take a few steps back and realize that not every person in tech is necessarily up to date on the latest in AI. It can be hard and overwhelming, especially since AI moves at such a rapid pace, even I struggle to keep up. I rely on Twitter, LinkedIn and Reddit to keep me up to date when I inevitably doomscroll. I hope I was able to get people up to speed with decision models and why they're the right form factor for certain gaming use cases. Thanks to Craig Robinson for the opportunity and for doing a great job hosting the first ever Games and AI demo night.

I've replaced iTerm2, VSCode and all other tools for agentic development with Twine, an open-source workspace for coding agents. Have Codex implement and Claude review. Design the handoffs in a workflow graph, inspect what each agent did, and keep their project memories in view. Waiting on @aravind_ni for updates! GitHub - github.com/standing-desks/tw…

Can Jev-like models speed run games? Kinda? Here are some experiments that I ran where different decision models played Stage 1 of Contra. Here I have Jev, Laya and Qwen3.5 battling it out. I still need to try out more model types and sizes. I loved playing this game as a child on the knockoff Nintendo NES that I had. Comment and let me know what game to try next. Follow for interesting experiments and visualizations.

Proud of my wife for getting back into the industry after the baby.

Suffering from success.

I gave four Jev-style decision models the controls to Doom: Jev, Laya, finetuned ModernCE and Qwen3.5. A deterministic Python adapter turns ViZDoom's visible-object labels + HUD into text. Each model picks turn, move or fire. No extra Doom-specific training. Same starting seed, separately recorded games. Each model's actions change its view. The video shows action probabilities, kills and survival time. One preselected episode/model on each of two scenarios. Thanks to @typesafeai for Jev access. Inspired by their Doom demo and others on Reddit and LinkedIn. typesafe.ai/blog/introducing…

A Slinky Down an Escalator - GPT 6 Astra max vs Claude Fable 5.1 max I finally got access to GPT 6 Astra and Claude Fable 5.1. I wanted to see how far we've come. I was inspired by the "pelican riding on a bicycle" test but wanted to push it a bit more in the direction of inter-object physics, so I came up with this idea for a perpetual slinky going down an escalator. Here I've compared GPT 6 Astra max on the left to Claude Fable 5.1 max on the right. Some notes: 1. GPT 6 Astra max was really fast and didn't burn too much of my weekly usage quota. 2. Claude Fable 5.1 max used up the entire 5 hr usage window and errored out once crossing the output token max limit. 3. GPT 6 Astra max seems too simple? There's also an issue where the slinky crosses into itself which should not be possible. Is this AGI? Maybe? 4. Claude Fable 5.1 max looks more real than I expected. Seems like it passes the eye test, unless I missed something. Let me know if you spot anything or had better ideas for a test. Overall I'd say Claude wins on quality and GPT wins on speed. Maybe I can tune my prompt a bit better for a more reliable result. This was my original prompt btw: Make me a single HTML file of a rainbow slinky going down an up-escalator forever. No libraries, just canvas and code. The slinky should be a chain of springs, each coil a different color of the rainbow. It starts folded in half like a horseshoe draped over a step. When dropped it flips end-over-end down the steps and because the escalator keeps moving up it tumbles in place and never reaches the bottom. Include a drop button and a reset button.

Thrilled that Ubiklaw: Local Voice-to-Dashboard AI got featured in the @AITinkerers Community Spotlights! See it here: post-training.aitinkerers.or…

Qwen3.8-27B on one DGX Spark hit 36.59 coding tok/s with SGLang + DFlash2, 2.89× plain decoding. The fastest tool-safe run was vLLM + DFlash2 at 47.41 median tok/s with 18/18 native calls. 12 deployments. Full results and failure notes: morethanamachine.com/posts/q…

I really didn't expect Codex to close the gap so quickly. The difference between what Codex was 3 months ago to right now is staggering! I’ve moved most of my day-to-day personal coding from Claude Code to Codex. Claude Code is still excellent, but it seems like for the first time in a long time, it has to play catch up. Kudos @OpenAI and @thsottiaux! It's more that they've stacked a few different wins at the right time. Anecdotally Codex with GPT 5.6 Sol as compared to Claude Code with Fable feels better in the following ways: 1. Speed: Faster with respect to time to first token(TFTT) and tokens per second(tok/s), likely due to using the Cerebras chips. 2. Stamina: More suited effective across longer tasks with auto-compaction. I rarely see any context rot. Also it's very token-efficient without wordy paragraphs. 3. Hiccups: It seems to make sensible implementation choices without pausing at every step. I've rarely had to course correct. 4. Visuals: Diagrams, assets, and image generation feel like part of the same workflow instead of a separate toolchain. Also better integration with more SOTA image generation models. Also Codex+GPT 5.6 Sol made this GIF. PS: I know this is purely anecdotal. When I have time and tokens I can follow up with a more formal comparison.