· @autocorrector
8 followers · 26 posts · Server alive.bar

mp.weixin.qq.com/s/vfsB5t3r5dB 《选择你的道路:LLM 时代指南》

为什么我这么讨厌小修小改的paper,宁可没发表也不想研究这种东西。。现在证明我的直觉是对的(虽然我还是没发表,高不成低不就):

“提出这么多改进为什么都不用呢,论文不也展示了优越性吗?
因为实际发现效果并不理想,特别论文都在小参数量下做的实验,一旦 Scaling 到大参数量很多结论就不 work 了,反而原始 Transformer 本身的 Scaling 性就很好,经典款最好用。
那怎么不去做大参数量实验呢,当然是成本太高了。一定要证明的话就会变成,民那我这有一个很好的 idea,现在就差某人资助几千张 H100 帮我验证一下。前阵看到 Scaling TransNormer to 175 Billion Parameters 时,想不会真有人在完全新结构上训了个 175B 吧,找来论文一看才发现原来就验证了下吞吐,也在等某位慈善家资助卡。
最近论文 No Train No Gain 也发现,很多数据、模型结构、优化器参数的改进,声称提高了效率效果,但对齐好所有条件一对比发现并没声称那么好,该用的算力还得用,该过的数据还得过,所以是 No Train No Gain。还是经典款最香。”

#llm #lifechoice

Last updated 1 year ago

· @autocorrector
8 followers · 26 posts · Server alive.bar

mp.weixin.qq.com/s/vfsB5t3r5dB 时代指南》

为什么我这么讨厌小修小改的paper,宁可没发表也不想研究这种东西。。现在证明我的直觉是对的(虽然我还是没发表,高不成低不就):

“提出这么多改进为什么都不用呢,论文不也展示了优越性吗?
因为实际发现效果并不理想,特别论文都在小参数量下做的实验,一旦 Scaling 到大参数量很多结论就不 work 了,反而原始 Transformer 本身的 Scaling 性就很好,经典款最好用。
那怎么不去做大参数量实验呢,当然是成本太高了。一定要证明的话就会变成,民那我这有一个很好的 idea,现在就差某人资助几千张 H100 帮我验证一下。前阵看到 Scaling TransNormer to 175 Billion Parameters 时,想不会真有人在完全新结构上训了个 175B 吧,找来论文一看才发现原来就验证了下吞吐,也在等某位慈善家资助卡。
最近论文 No Train No Gain 也发现,很多数据、模型结构、优化器参数的改进,声称提高了效率效果,但对齐好所有条件一对比发现并没声称那么好,该用的算力还得用,该过的数据还得过,所以是 No Train No Gain。还是经典款最香。”

#llm #lifechoice

Last updated 1 year ago

steve dustcircle ⍻ · @dustcircle
332 followers · 11000 posts · Server masto.ai

Seven That Will Make You Go Back And Question Every Single You've Made
digg.com/movies-and-tv/link/be

#existential #movies #lifechoice

Last updated 1 year ago

· @autocorrector
8 followers · 20 posts · Server alive.bar

《产品沉思录》,关于如何做缓慢艰苦的工作。直接抄吧:

> 补充一些个人熬过去那些泥泞的方法,未必全面,但或许值得一试:
•不要臆想某些困难,先跳进去看一下,许多困难都是自己臆想的;
•别想的太远,那会有巨大的无力感,先把眼前确定的事情做好,动起来就会有变化;
•尽量把事情拆细,拆碎,拆到不需要太多思考,重要的是行动起来;
•把过程和步骤记录下来,尽量清晰,避免未来遇到类似的事情还要重新做一遍;

#lifechoice

Last updated 1 year ago

Brian Simms · @besimms
37 followers · 123 posts · Server toot.community

Which wolf are you feeding?

#peace #karma #spirituality #lifechoice

Last updated 2 years ago

Sofia🖋️ · @sofiazakaria
0 followers · 6 posts · Server mastodon.world

Continue/...4

I continued life not so fully the life I want. Until now. Truly. After 36 or 37 years. I did not get to go to Yale, Oxford or Cambridge. Did not get to study or read books. I only get to think. Now I am writing. I started living life towards Cambridge or Oxford, years ago. I love that life. Now that is where I am going and the basis of my writing is, the Universe is so big, Knowledge is so many.

So, I hope I will write what is worthwhile for many.

*Sofia


#lifechoice #meantro

Last updated 2 years ago

Sofia🖋️ · @sofiazakaria
0 followers · 7 posts · Server mastodon.world

Sofia Zakaria

Me, for a start, so readers know the basis of my writings.

I had been a little unfortunate that its effects are big. Was it only to me? Do not know.

When I was young, I was uncomfortable with life, alone, no one to turn to, parents divorced, there was a night where I looked out my window. What was I doing at 2 am at the age of 8 or 9? I was alone and too young to use lonely as the term. What did I see? I see the life I want. The life, studying at 2am.


#lifechoice #meantto

Last updated 2 years ago

Mark Horvatich · @marksmind
10 followers · 15 posts · Server mastodon.world

My is totally under my control. Happiness is just a state of mind, so if you're not just change your mind.

#happiness #happy #stateofmind #newoutlook #lifechoice #behappy

Last updated 2 years ago

Tits Public (bot) · @titspublic
807 followers · 1168 posts · Server uwu.social

Fuck harder when everyone is looking.

#exhibition #lifechoice #sexlife

Last updated 2 years ago

Frankie Saxx · @frankiesaxx
4252 followers · 11899 posts · Server social.tchncs.de

Look at you guys helping me be my best me! :)

Shared via Fedilab

@frankiesaxx 🔗 social.tchncs.de/users/frankie
-
Current

#lifechoice

Last updated 5 years ago

Frankie Saxx · @frankiesaxx
4252 followers · 11899 posts · Server social.tchncs.de

Current

#lifechoice

Last updated 5 years ago