Mayank Singh

PhD student at @UArizona | Interested in #NLProc, History, #Astronomy, and more
Part-time @GameOfThrones fan account

[Denny’s @DennysDiner]

@JensenHuang @nvidia

[American Mathematical Society @amermathsoc]

Congratulations to the 2026 Fields Medal recipients! Your groundbreaking discoveries are expanding the frontiers of mathematics and inspiring generations of mathematicians. The American Mathematical Society is honored to celebrate your achievements at #ICM2026 in Philadelphia.

[JT @jiratickets]

"claude fable just proved the jacobian conjecture is false"

"incredible this changes everything"

"do you know what the jacobian conjecture is?"

"no, do you?"

"no"

[(((ل()(ل() 'yoav))))👾 @yoavgo]

no no you don't understand its not "just training"! we had to try five different optimizers and play with the learning rate scheduling and we also did item weighting to get our best results!

[(((ل()(ل() 'yoav))))👾 @yoavgo]

why is it that influencing system behavior through a detailed prompt + relevant context is "oh, so your research is just prompting?" while doing some SFT or GRPO on some examples to achieve somewhat worse results is "wow cool research bro"?

[Phoenix Yin @Phoenixyin13]

如今大名鼎鼎、作为 ChatGPT 底层核心对齐算法之一的 PPO，在 2017 年竟然被顶级 AI 会议 NIPS（现 NeurIPS）给拒了。

更讽刺的是审稿人给出的理由：
创新性有限
对比基线提升不够显著

这个现象，直接扯下了学术界同行评审的几层遮羞布。

在学术圈，尤其是顶会审稿人的眼里，什么叫创新？

必须得有满大篇看不懂的数学公式推导。
必须得是一个前人从来没提过的全新复杂框架

而 PPO 的核心精神恰恰相反，它是极简主义的。
PPO 的前身是 TRPO，那篇论文数学推导极度复杂，虽然理论完美，但巨难实现，算力消耗大。
John Schulman 做 PPO 的初衷，就是想把复杂理论降维成一种写起来简单、跑起来稳定、工程上极度好用的算法。

结果，这种为了工程实用而做出的伟大简化，在傲慢的审稿人眼里变成了缺乏创新。学术界常常陷入一种自我感动的怪圈：

把简单问题复杂化叫水平，把复杂问题简单化叫水货。

2017 年前后的强化学习论文，充斥着大量的炼丹成分：
换个Random Seed，调个超参数，实验结果就能天差地别。

很多论文为了能在表格里刷出一个好看的最高分去应付审稿人，会疯狂地刷榜和调优，但这在实际工业界根本无法复现。

PPO 在某些特定的玩具任务上，可能确实没有那些特调的算法刷分高。
但 PPO 真正的恐怖之处在于它的泛化性和鲁棒性。
它是第一个能让你不需要怎么调参，丢给复杂任务就能直接跑通的 RL 算法。

审稿人只盯着眼前的Benchmark，却完全没有意识到他们亲手毙掉了一个未来能撬动千亿级大模型产业的工业奇迹。

如果当年 John Schulman 因为被 NIPS 拒绝就怀疑人生、放弃 PPO，或许后来的LLM对齐RLHF还要多走几年的弯路。

真正伟大的研究，它的生命力从来不在于评委给的 Accept录用。

[Mayank Singh @geekmarcus]

Molag Bal also goes hard.

[Greg the Sorcerer @gregthesorcerer]

Kid named Mehrunes Dagon

[Mayank Singh @geekmarcus]

From Google: "John D. Rockefeller became the world's first billionaire on September 29, 1916." So, it took ~110 years for the first trillionaire.

[Mayank Singh @geekmarcus]

Inflation will keep rolling on like entropy. What year do we get the first quadrillionaire?

[\/¡$H@|_ @__weeshal__]

This was really heartbreaking, & eye opening at the same time.

Don't want to sound negative but, what this showed us is:

Years of planning, hardwork, documentation, constant hustling for better career and life. Everything can be ended in 1 single snap.

All you got is 'Now' 😐

[Kyle Gawley @kylegawley]

I cancelled my $10/mo Calendly subscription and vibe coded my own with Fable for $12,000

[RJC @RJCcapital]

buddy needs to relax

[bayes @bayeslord]

They didn’t mean pause AI research, they meant pause *your* AI research

[The Witcher @thewitcher]

Medallion's humming... that can only mean one thing! It's time to announce The Witcher 3: Wild Hunt - Songs of the Past! ⚔️

This brand new expansion for The Witcher 3: Wild Hunt will take you to the Path with Geralt of Rivia once more. It’s being co-developed with @Fools_Theory and is coming to PC, Xbox Series X|S, and PlayStation 5 in 2027. Stay tuned for more information in late summer. ⏰

[Mayank Singh @geekmarcus]

@SwarajyaMag @astrokaran Loving these well researched articles.

[THE BOYS @TheBoysTV]

Last time we get to post this. End of an era, mates. Thank ya for making this admin’s job a fuckin’ blast.

[Mayank Singh @geekmarcus]

@ShashiTharoor Rasmalai >>>

[Birdcomixstudios @TheDylanRoca]

Do not mess with 100 Acre Woodites. They do NOT know how to read, and are particularly vengeful.

[NooB. @NooB_B3D]

"The Dark Lands"