When OpenAl converted the o3 parent model to a chat version of the model-also known as a student model-that allowed people to ask it anything, its gains degraded significantly to the point where it wasn't performing much better than o1, the people who were involved in its development said.
四月的最后一天,又是一个电闪雷鸣的雨夜。记得上次在这里,假期回家,小宝宝刚出生,正是闹觉的时候。半夜两三点钟,实在是无心睡了,看着这电闪雷鸣的夜晚,发现闪电是那么的清晰,完全没有被云遮挡。今天这场雨就没那么巧了,只见闪光但是看不见闪电。所以那年的闪电还真是,我可是说once in a lifetime thunderstorm吗。
比亚迪在巴西工作的工人,(https://edition.cnn.com/2024/12/24/business/byd-brazil-chinese-workers-conditions-intl/index.html)被认为是奴隶一样的工作。然后公关就拿民族主义说事,比华为更明目张胆的鼓动民粹。我心想这些照片都流出来了啊,这些工人的环境确实配得上modern slavery and foreced labor的定义。结果不少人说,也不知道是不是网军,国内都这样。如果真的是他们的肺腑之言,那我觉得他们活该天天加班,过劳死。他们配得上这些苦难。比亚迪、华为之流,作贱自己作践行业作贱社会。
Summary of how DeepSeek V3 was so efficient at training a frontier-level model according to Perplexity:
—
DeepSeek-V3 was able to train their large 671B parameter model with a relatively low compute budget of 2.788M H800 GPU hours through several key innovations and…
DeepSeek is probably the least understood tier-1 AI company
Some interesting facts about DeepSeek:
- never received/sought outside funding (thus far) - self-funded out of a hedge fund (called High-Flyer) - entire AI team is reportedly recruited from within China, no one… https://t.co/ujgKbQQ2Nq