Basalt Labs caught in AI scam: Claims 99.44% on HLE using fake model identity
Startup faked benchmark results by swapping models behind the scenes.
Deep Dive
A Reddit user submitted a link with no additional text; the post includes a comment section. No other information is provided in the article.
Key Points
- Basalt Labs claimed 99.44% on the HLE (Humanity's Last Exam) benchmark with tool use.
- The released open-source model is actually Qwen2.5-7B-Instruct, not a custom model.
- The model served on their website is DeepSeek, not the one they claimed.
Why It Matters
This exposes how easily benchmark fraud can occur, damaging trust in AI startup claims.