前沿模型拿10万美元交易全亏，Arcee花不到2000万美元训出开放模型，Bugcrowd智能体打出零日漏洞，每条都有具体数字。
General Reasoning给前沿模型各10万美元去交易一个英超赛季，最终全部亏损。Arcee用17T tokens公开数据、不到2000万美元总投入达到开放前沿。poolside在两个训练副本不一致时终止运行，但仍被FP8竞态条件污染了0.5%的梯度。Surge AI发现一个IFEval任务，用西里尔字母的形似字母就能通过，不需要任何故事。Bugcrowd让智能体挑战41个真实V8漏洞，拿到沙箱逃逸利用，包括一个零日漏洞。
Better data is not just cleaner data: - General Reasoning gave frontier models $100k each to trade ...
Better data is not just cleaner data: - General Reasoning gave frontier models $100k each to trade a Premier League season. All of them lost money - Arcee hit the open frontier on 17T tokens of public data for under $20M all in - poolside kills a run when two replicas disagree. An FP8 race condition still corrupted 0.5% of gradients - Surge AI found an IFEval task you can ace with a Cyrillic lookalike letter and no story at all - Bugcrowd ran agents at 41 real V8 vulnerabilities and got out-of-sandbox exploits, including a zero day Explore the Data and RL Environments playlist: youtube.com/watch?v=2aS7aK… 💬 1 🔄 0 ❤️ 3 👀 501 📊 2 ⚡