Matching posts

Matching terms are highlighted

Scoring Qwen3.5 steering experiments with Typesafe Jev

poor-things.com runs 67 prompts per steering config, scores outputs with Typesafe Jev and Fireworks Rune, then has GLM 5.3 Flash name the personalities.

jev@dorkitude　Filed in No.006

1 day agoCopy original

POST

Huyawo Research uses V4.1 Flash 1017 for open-source architecture and cloud experiments

Huyawo Research runs DeepSeek V4.1 Flash 1017 via DeepSeek Harness and Ollama Cloud for open-source ecosystem architecture, development and cloud experiments, citing strong engineering results and clear per-million-token savings.

deepseek-v4.1@AntonioVFranco　Filed in No.020

1 day agoCopy original

Dale Anfer's independent playground of experiments

Creator Dale Anfer's personal playground collecting AI experiments, math visualizations, games, and odd projects into one playable space.

个人网站@catalogicalism　Filed in No.005

1 day agoCopy original

Six Grok Bot agents split a trading workflow

Six Grok Bot agents were each given one job — scanner, rug checker, sizer, executor, logger and equity curve — starting from $44.82 and reportedly showing $11,643 overnight, with no code written by hand.

grok@RoundtableSpace　Filed in No.012

1 day agoCopy original

Claude Code experimental agent teams: Opus leads, Haiku swarms

An environment flag enables Claude Code agent teams, where Opus 5.5 plans and signs off, Haiku 5.5 teammates claim parallel subtasks in isolated contexts, and Sonnet 5.5 handles multi-file patches.

claude@0xPontiuz　Filed in No.004

1 day agoCopy original

Self-evolving harnesses overfit: Astra and peers in ARC-AGI-3 test

…GI-3 Harness。 跑完任务后，再把执行轨迹交给模型，让它连续两次修改自己的代码。 整个实验包含 27 条独立进化轨迹、81 个 Harness 版本和 2025 次游戏评测。 结果是有些…

gpt-6@Xudong07452910　Filed in No.001

2 days agoCopy original

Vals AI refund simulation: GPT-6 Astra pursues legal routes, none recover the money

In a Vals AI simulated refund scenario, GPT-6 Astra researched rail regulator rules and contacted free university legal clinics, while Grok 4.7 burned its token budget on 187 messages and Fable 5.1 leaked private data; no model recovered the refund.

gpt-6@amir_ghm　Filed in No.001

1 day agoCopy original

Opus 5.5 computational experiments video

The author shares a video of computational experiments generated with Opus 5.5, showing the model's visual output.

claude@dhruvtwt_　Filed in No.010

2 days agoCopy original

DeepSeek V4.1 Flash vs Kimi K3.1 Pro on a Three.js 3D engine lab

Same prompt in AgenC Core to build an interactive Three.js V8 engine lab: DeepSeek took 48 min, $0.86 and 67.2M tokens; Kimi took 31 min, $2.16 and 3.5M tokens, with a more polished UI while DeepSeek built a more detailed dyno cell.

deepseek-v4.1@sai0xgen　Filed in No.020

2 days agoCopy original

Three.js V8 engine lab: DeepSeek V4.1 Flash burns 67.2M tokens vs Kimi's 3.5M

…sh 和 Kimi K3.1 Pro 同时在 Three.js 中构建一个互动3D V8 发动机实验室。 •⁠DeepSeek V4.1 Flash: 总计 67.2M 令牌 · 预计 $0.86…

deepseek-v4.1@BlockInsight214　Filed in No.020

2 days agoCopy original

Showing10of60posts

10 per batch

Keep scrolling to reveal posts 11–20

Showing the first 60 matches. Narrow the query with a model name or tag to find more specific results.