Mystery model V4.1F-1017 still has '2025 LLM' failures in hands-on test
teortaxesTex · x · 2026-10-29
A hands-on test of the unreleased model codenamed V4.1F-1017 shows it still exhibits "2025-era LLM" failures: in its first run it coded itself into diminishing returns while recreating the 2004 game Ballance, hitting a ball spin-axis bug and spiraling into repetitive patching. After a handoff and another 500M tokens, it mostly recovered the original game.
The screenshot shows a 5-member Agent Team (lead/phys/vfx/audio/recovery) splitting physics, VFX, and audio tasks via shared task scopes. A replier teases "Bro is not on the team… curious," hinting the model's lab affiliation is unclear.
More from coding & agent
- Weaviate Podcast: MIT PhD's "Machine Studying" lets always-on agents prep without knowing the task — CShorten30 · 2026-10-29
- Dev rebuilds Slack as video with Cursor agents and computer use — mattyp · 2026-10-29
- Deepgram acquires Daily, spins up independent Pipecat Foundation — charles_irl · 2026-10-29
- Don't force multi-agent for enterprise: a one-night RPA script beats slow, costly agents — sujingshen · 2026-10-29
- After 30 approval prompts, 'yes' to everything sent a broken release to his whole list — latent_lurker8587 · 2026-10-29
- Nat Eliason builds a personal AI writing system in one Claude Code session — nateliason · 2026-10-29