i like making stuff founder @ husena
texas
Joined July 2026
- Tweets207
- Following4
- Followers11
- Likes65
think ive found a genuinely different approach to the human agent interaction problem that coincidentally allows the simulation of super intelligence. more soon (maybe).
1/
I gave a small language model 1,338× more parameters than it stores.
It got better, but not for the reason you'd think.
Here's what worked, what failed, and what it means 🧵
claude.ai/artifact/TaEZZJbE3…
8/
What it means in practice: roughly the quality of a 6–23% bigger model for ~3% more cost.
That matters most where memory is the bottleneck, like on-device AI and cheaper inference.
It's modest, not magic: about 1.2% lower perplexity.
9/
Full write-up with every failed experiment, the controls and the stats: claude.ai/artifact/TaEZZJbE3…
Still jobless and doing this research on my own in hopes of some exposure and a job. If you're hiring for ML/AI research or engineering, my DMs are open 🙏
made a PS1-style katana game 🗡️
parry the yellow. dodge the red. cut through a cult's neon skyscraper floor by floor. 3 bosses, 30 floors, then it goes endless
free, plays right in your browser 👇
dgresesjr.itch.io/skycutter
#indiedev #gamedev #madewithgodot #lowpoly #psx
For this test I went deeper. I wanted to see how models handle genuine reasoning and asset creation, not copying schematics or matching existing designs.
The prompt: invent an ORIGINAL mechanical walking machine, build every part from scratch in Blender via Blender MCP, then animate a 60-second sequence in Godot 4.
Key constraints:
- Legs only, no wheels. Every motion has to trace back to a visible motor through real gears, shafts, or linkages.
- No copying Strandbeest, Spot, or any known design.
- Carries a crate and sets it down. Fits in a 2x2x2 m box. Must look buildable.
- Design doc first, with real numbers (gear ratios, stride, speed) that have to match the final animation.
- Scripted 60s sequence: startup, walk, cross an obstacle, turn 90°+, drop the crate, shut down. No foot sliding, speed = stride × step rate.
- Final report with a designed-vs-actual verification table and an honest list of flaws.
They all did pretty well, ngl. Opus took the longest, but I think it also thought the hardest, even with every model on high reasoning. I prefer Opus for tasks like this. Astra and Sol were pretty creative, though, and their harness is easier to use for agentic tasks that need PC control.
Running comparison again. Looking back at the runs, Claude didn't pull reference material and Codex did. I'd also tried to control for default reasoning level, and I think that skewed things. This time: same reasoning level for all, plus references and more detail.
Prompt:
"Using blender mcp. Make me a realistic 3d model of the: Lockheed Martin F-22 Raptor (US only). Then create me 60 second mp4 of it flying around in Godot that I can upload to X."
Models: 6.1 Sol, 6 Astra, 5.5 Opus
Blender MCP (official): blender.org/lab/mcp-server/
Godot 4.6.3: godotengine.org/download/arc…
Demetrius Greses Jr retweeted
Wait, I can't believe Anthropic posted this.
Their new research post is a full red team report on someone else's model, GLM-5.3 from Zhipu.
The finding is that it builds working exploits almost as well as Claude Mythos Preview. You know, the model Anthropic kept locked away for safety.
Except this one you can just download.
making a new game called "Skycutter", working on game arch now. got alot of the assets done already. its gonna have ps1 graphics cause it easier to handle asset creation wise as a single person.
Demetrius Greses Jr retweeted
If your pancakes don’t look like these with the burnt edges they ain’t real pancakes
