BTC Miner | Solo Builder | AI Infra
Joined April 2021
- Tweets665
- Following1.6K
- Followers451
- Likes1.4K
ChatGPT Ads just crossed $1B in annualized revenue. The next AI price war won't end at cheaper tokens. It ends at free inference.
Consumer AI builders should optimize for intent capture, not token margin. Intelligence becomes the acquisition layer.
尝试了三个星期去驯服 Opus 5，我失败了，我切回到原来的模型。
原因是啥呢？是这个模型实在是太不可理喻了。我宁愿要一个能 get shit done 的自闭症患者，都不要一个自己找理由、没办法做成任何事情的模型。
每天就是："啊，这个事有个 blocker 做不了。"
然后我说："你确定吗？你需要 ABC，并且我们已经有了。"
然后它说："哦，你太对了。"
Xiaohongshu just opened a 280B model that activates 16B parameters, 5.7% of the headline size, with 512K context across text, vision and audio.
Parameter count is dead as an infra metric.
Provision for active compute. Let marketing count the dormant weights.
Ten open problems in mathematics were solved last week. Total compute cost: about $2,000.
Two hundred dollars per problem. And every proof came with machine-checkable Lean 4, `sorry` count zero.
When a system produces original mathematics that a proof assistant verifies, the leaderboard stopped being the instrument. You need a different one.
The good news is we already have it, and it doesn't grade on a curve: does the proof check, or not.