lansea
906 posts


前 Anthropic 研究科学家、现 Google DeepMind 研究科学家姚顺宇,在 @zhang_benita 播客「语言即世界」中首次披露了 Claude 3.7 的内部研发过程。他在 2024 年 10 月加入 Anthropic 后被分进一个名为 Horizon 的团队,当时整个团队只有 10 到 11 个人,却负责 Anthropic 强化学习的全部工作,包括数据、基础设施和算法研究。Claude 3.7 从启动研究到最终发布总共耗时四五个月,前两三个月做算法和数据研究,后两个月做训练和基础设施搭建。
Anthropic 押注代码能力并非一开始就有规划。姚顺宇透露,Claude 3 之所以写代码比 GPT-4 强,背后有一个他无法公开的纯技术原因,是某个团队自下而上做出来的。Claude 3 发布后 Twitter 上的大量正面反馈验证了这一优势,Anthropic 管理层随即把代码能力升级为公司级战略全力押注。他认为 Anthropic 能这样快速下重注,核心在于技术一号位 Jared Kaplan 和 Sam McCandlish 本身就是联合创始人,技术上服众的同时也有权拍板,而 OpenAI 做不到这点,Ilya 在的时候也许行,但后来失去了决策权就走了。当时的 Anthropic 在产品方面几乎没有意识,Claude 3.5 半年内发了两个版本却用同一个名字,最终靠外界起的绰号「3.6」才勉强区分开来。
中文
lansea retweetledi

Post shall be published soon, but for now please be prepared for a tweet storm.
Thread: why block merge is important milestone and how it brings horizontal scalability.
Spoiler: Some preliminary results of implementation in the end.
Nutzipper@nutzipper
Planning to write an article about this picture
English

@nutzipper @rchain_coop @leithaus Can we develop a compiler compile rholang code running in gpu, gpu have thousands kernel concurrent run the code?
English

@lansea @rchain_coop If I recall correctly, some initial PoC implementations were created last year. @leithaus please correct me if I'm wrong.
English
lansea retweetledi

Today at @rchain_coop debrief I'll try to talk about how RChain can scale and utilise all resources you give to it.

English

@nutzipper @rchain_coop amazing, this meant Greg begin to design FPGA chip to run rholang code directly?
English

@lansea @rchain_coop youtu.be/ne99laPUxN4?t=… recently this video has been circulating in RChain. Rholang is a language for exactly such kind of computing beasts. So you can imagine that Greg is not thinking very hard about how we will scale Scala implementation.

YouTube
English

@nutzipper @rchain_coop thanks the reply, I just confused by the number. It is a great demo, show the evidence of scale abilities, give us the strong confidence for the future of the rchain.
English

@lansea @rchain_coop I have to revisit what exactly Greg said, might be he meant 1k nodes. He was talking a lot about h/w acceleration, e.g. on FPGA. But any way, our current implementation has lot of room for improvement, even in pure s/w Scala node. The key point of presentation is that it scales.
English
lansea retweetledi
lansea retweetledi

@justinsuntron willing to donate 42 mil.. but cant afford a 35 k tesla. imo Shame on you. the clown of all crypto
English

@cz_binance @DerekGeorgeCox @Crypto_Bitlord @uzgaroth @justinsuntron @Tronfoundation @binance please let @justinsuntron give @uzgaroth a Tesla, I am sure @justinsuntron have make the promise.
English

To support @binance , I will personally deposit 7000 BTC worth USDT (40 million USDT) into @binance to buy $BNB, $BTC , $TRX & $BTT if @cz_binance agrees. No need to #FUD! Funds are #SAFU!
English

@inversebrah Probably, not enough liquidity for the smaller alts but that’s a good question
English







