🏆 3x AI Hackathon Winner | Ex-MSFT Cloud Architect building AI agents & benchmarks for Claude/Codex/Gemini

Joined May 2010
Pinned Tweet
I gave Gemini 4 Argon 8 reference images. One night later: a 3D Paris around the Eiffel Tower. Day turns to golden night, people walk, bikes and pigeons pass by. It runs in your browser. $48.97 · 3 h 24 min of agent work · OpenCode + Blender + Three.js Walk it yourself 👇
1
1
2
353
Same prompt. 3 AI coding agents. One 3D Lamborghini each, 100% code, no 3D files. 1 Gemini 4 Argon (opencode) 2 Claude Opus 5.5 (Claude Code) 3 GPT-6 Astra (Codex) Which car wins? Reply 1, 2 or 3 Full review with benchmarks on YouTube soon.
1
128
Day 0/ Anvilterm turn terminal agents ->chat ui Claude Code, Codex and OpenCode, as a real desktop app. Same official CLIs, same accounts. One right-click turns the terminal into a chat with live previews, and back. AnvilTerm 1.2.2, free for Mac, Windows and Linux.
2
2
114
I asked a stealth AI model to fingerprint its own tokenizer. 🐰 Space Bunny Alpha's token counts match MiniMax's tokenizer exactly: 18/18, zero error. Its reasoning about the result: "The model is Claude." Models can copy each other's style, and even identity. They can't copy a tokenizer.
134
AK retweeted
Nerfed? Today Claude Opus 5.5 added phone controllers, iPhone support and 4 jump tricks to my kart racer. Sonnet 5.5 built it. Opus 5.5 fixed what broke. GPT-6.1 Sol painted the tracks. 4-player split screen on my TV, runs in your browser.
2
1
1
183
Nerfed? Today Claude Opus 5.5 added phone controllers, iPhone support and 4 jump tricks to my kart racer. Sonnet 5.5 built it. Opus 5.5 fixed what broke. GPT-6.1 Sol painted the tracks. 4-player split screen on my TV, runs in your browser.
2
1
1
183
Claude Opus 5.5, one sentence with typos and an evening of complaints later. We got physics, nitro, drifting and six AI rivals. Version one had no physics and a black screen after 3, 2, 1. Made in three.js. The car started as one AI picture. Runs in your browser.
1
307
claude 5.5 created a dancing chef in blender from a reference image
109
what i noticed with codex and astra it start good but the more you keep iterating the more it gets dummer Astra high/medium used here, anyone noticing sth similar?
1
35
AK retweeted
Replying to @sama
your post got me thinking about the adventure games I grew up with. I used GPT‑6 Astra in Codex to build this Monkey Island-inspired playable prototype. Seeing Mara walk, talk and solve puzzles still feels surreal. Rough edges included 🙂 Play: lantern-cove.akartit.chatgpt…
1
1
171
Fable 5.1 was great for like 12 hours then seems to be nerfed again, anyone notice similar thing?
45
si c'est pour les hopiteaux pourquoi en publique le stock?
8
You said "proactive." So I gave it a mouse. Opus 5 drew this in MS Paint. 67 strokes, 753 drag events, on a Windows VM it had to ask permission to touch first. No image model. Just a cursor.
Introducing Claude Opus 5. It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.
147
AK retweeted
GPT 5.6 created @sama 's resumé in 5 minutes live using anvilterm Pro anvilterm.com/demo/gpt5.6sam…
1
1
1
196