Yuchen Jin ✓

@Yuchenj_UW · another Galaxy · joined 22 Nov 2016

Cooking fun AI systems & products @databricks. Prev: co-founder & CTO @ Hyperbolic, OctoAI (acquired by @nvidia) Apache TVM, PhD @ University of Washington.

287 590Followers

836Following

14 229Posts total

3MViews on collected posts

โพสต์ล่าสุด

Yuchen Jin ✓@Yuchenj_UW12 ชม.

OpenAI’s “significantly more capable than GPT-6 Astra” model solving a Millennium Prize Problem is huge, an AlphaGo moment for math. AI is now smarter than a math PhD, but dumber than an intern in many tasks. “Jagged intelligence” needs to be solved. 9.2K views · 296 likes · 18 reposts · 22 replies Open on X →

Yuchen Jin ✓@Yuchenj_UW7 Sep

OpenAI’s comeback is the biggest comeback after Meta’s comeback, after xAI’s comeback, after Google’s comeback. AI is an endless cycle of “it’s over” and “we’re so back.” 46.7K views · 1.3K likes · 52 reposts · 79 replies Open on X →

Yuchen Jin ✓@Yuchenj_UW7 Sep

Sam and Jensen say “AGI has arrived.” But “AGI” means different things depending on who you ask. - Sam’s bar: outperform humans at most economically valuable work. - Dario and Demis’s bar: smarter than a Nobel Prize winner across most fields. These are wildly different bars. 54.5K views · 900 likes · 39 reposts · 112 replies Open on X →

Yuchen Jin ✓@Yuchenj_UW6 Sep

If you want to understand how fast AI is improving, look at intelligence per dollar. 1.5 years ago: o1 Pro: $150 / $600 per M tokens Today: GLM-5.3 Flash: $0.15 / $0.50 That’s a ~1000x collapse in price in under 1.5 years, while GLM-5.3 Flash is more intelligent than o1 Pro. 411K views · 5.4K likes · 314 reposts · 133 replies Open on X →

Yuchen Jin ✓@Yuchenj_UW2 Jun 2024

@Yuchenj_UW It’s interesting that you can 3X the LR. You’d expect the original paper to be well tuned near what is tolerable. 44.3K views · 138 likes · 1 reposts · 3 replies Open on X →

Yuchen Jin ✓@Yuchenj_UW1 Jun 2024

@karpathy I'm unsure why the evaluation score dropped near the end. It might be due to some bad training data batch. @karpathy, would love to hear your insights. It reminds me of the lossfunctions Tumblr though lol: https://t.co/9MQivsrCPI. 10.3K views · 27 likes · 2 reposts · 0 replies Open on X →

Yuchen Jin ✓@Yuchenj_UW1 Jun 2024

Outperform GPT-3 with @karpathy's llm.c using just 1/3 training tokens ✨ Another day has passed, and I trained GPT-2 (124M) with llm.c for 150B tokens, achieving 35.5% accuracy on HellaSwag. This surpasses the GPT-3 paper’s 33.7% accuracy trained for 300B tokens. It matched the

2.4M views · 1.6K likes · 137 reposts · 126 replies Open on X →

เทียบกับบัญชีขนาดเดียวกัน

4 โพสต์จาก 90 วันที่ผ่านมา เทียบกับช่วง 100K–1M ผู้ติดตาม แสดงไปไกลเกินผู้ติดตามของตัวเอง และผู้ชมก็ตอบสนอง.

ยอดดูมัธยฐาน50 598บัญชีนี้4 981ค่ามัธยฐานของ 100K–1M

การเข้าถึง, %17.59%บัญชีนี้1.73%ค่ามัธยฐานของ 100K–1M

การมีส่วนร่วม, %2.57%บัญชีนี้1.22%ค่ามัธยฐานของ 100K–1M

ตัวชี้วัด	บัญชีนี้	ค่ามัธยฐานของ 100K–1M	อัตราส่วน
ยอดดูมัธยฐานต่อโพสต์	50 598	4 981	10.2×
การเข้าถึง (ยอดดู ÷ ผู้ติดตาม)	17.59%	1.73%	10.2×
อัตราการมีส่วนร่วม	2.57%	1.22%	2.10×

บัญชีอื่นในช่วงนี้ → เปรียบเทียบกับบัญชีอื่น → ค่าอ้างอิงเหล่านี้คำนวณอย่างไร →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

2.4M1 Jun

10.3K

44.3K2 Jun

411K6 Sep

54.5K7 Sep

46.7K

9.2K8 Sep

Last 7 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.08%1 Jun

0.28%

0.32%2 Jun

1.43%6 Sep

1.96%7 Sep

3.17%

3.65%8 Sep

Reactions — likes, reposts, replies and quotes — divided by views. Median for 100K–1M accounts is 1.22%.

What the audience does

Likes75.6%9 646 in total

Reposts4.4%563 in total

Replies3.7%475 in total

Quotes0.8%108 in total

Bookmarks15.5%1 973 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.