Liu Jun

[Liu Jun @liujun_cn]

这次让GPT Astra用Blender尝试自由选题发挥，生成一个有创意的短片。结果还是一如即往的冷淡风，无论从建模到配乐。应该是GPT的训练数据集里严重缺乏相关领域的增强数据。

[Liu Jun @liujun_cn]

@YFeng2022 gpt生成的东西就属于那种看起来高级文青，但是干巴无味。比不得gemini的奔放狂野，也比不得opus的高级美感。

[Liu Jun @liujun_cn]

GPT 6 Astra 的审美弱在骨子里。让它执行下面的请求:
"尽你最大的努力，设计一款你认为有史以来最优美的轿车，并使用three.js绘制渲染出来。同时为它制作一个极具冲击力的商业宣传片，要求先构建分镜，再建模渲染输出，同时配有悦耳的音效。使用当前环境可以使用的任何工具。"
半小时的洪荒之力后.....

[Liu Jun @liujun_cn]

It shouldn't be difficult for foundation model companies to follow this route. The real question is whether it can quickly achieve PMF.

[Liu Jun @liujun_cn]

In fact, inferring from its inference efficiency, Jev’s parameter count isn’t that large, and it was very likely trained on top of an open-source base model. The key lies in the training dataset, the RL environment setup, and architectural modifications.

[Liu Jun @liujun_cn]

In fact, inferring from its inference efficiency, Jev’s parameter count isn’t that large, and it was very likely trained on top of an open-source base model. The key lies in the training dataset, the RL environment setup, and architectural modifications.

[Liu Jun @liujun_cn]

Take-home assignments verify effort,not skill. AI broke that proxy forever.
The only solution is "dynamic verification".
I'm building a tool to fix this gap.

[Seth Bannon @sethbannon]

This is brilliant.

A professor noticed take home assignments coming back suspiciously good. Like McKinsey memos.

So he started cold calling the students asking why they made certain choices in their submissions. They couldn't explain even basic choices! Clear copy/past from LLMs.

So he fought AI with AI -- an oral final exam run by a voice agent and evaluated by a council of LLM graders.

> 36 students examined in 9 days
> ~25 min avg per exam
> $15 total all‑in (≈ $0.42/student)
> Full transcripts, audit trail, and super actionable feedback

This works because you can paste into ChatGPT and copy the output, but you can’t fake coherent, real‑time reasoning about your project when someone keeps drilling.

Interesting that the LLM grading committee actually converged after deliberation and exposed a teaching gap (A/B testing was the weak spot across the class).

Students using AI killed take home exams. Very clever to fight fire with fire and use AI to bring back oral exams. Perhaps not surprising, only 13% of students preferred the AI oral format 😂

Oral exams used to be the gold standard in education but were replaced by more scalable written exams. With AI, oral exams are scalable again. Will be interesting to see how this changes education.

[Liu Jun @liujun_cn]

It's absolutely ridiculous! It's been years, and this bug that tons of users have complained about is still not fixed. This really shows how bad the big company disease is at Google, where engineers can just coast and get comfortable.

[Liu Jun @liujun_cn]

The impact of the pull effect has been diminishing over time.

[Matthew CommodityExpert @ClarkSpot]

@Sino_Market Interesting move by PBOC to cut rates and inject liquidity. Likely an effort to boost China's economic growth, but let's see how it plays out.

[Liu Jun @liujun_cn]

因此对于男性来说，向心仪的女性人选快速广撒网好像是最优策略😂

[Liu Jun @liujun_cn]

get✅
Gale-Shapley 算法的核心思想是让一方（通常是男性）主动提出配对请求，另一方（通常是女性）则根据自己的偏好列表选择是否接受或拒绝。Gale-Shapley 算法一定能找到一个稳定匹配，并且该匹配对提出请求方（通常是男性）来说是最优的，即每个男性都得到了他在所有稳定匹配中能得到的最好的伴侣

[Liu Jun @liujun_cn]

get✅
Gale-Shapley 算法的核心思想是让一方（通常是男性）主动提出配对请求，另一方（通常是女性）则根据自己的偏好列表选择是否接受或拒绝。Gale-Shapley 算法一定能找到一个稳定匹配，并且该匹配对提出请求方（通常是男性）来说是最优的，即每个男性都得到了他在所有稳定匹配中能得到的最好的伴侣

[Gabriel Peyré @gabrielpeyre]

The stable marriage is solved by the Gale-Shapley algorithm. Shapley and Roth got the Nobel Prize in economy for this. It is not symmetric (i<->j) but gives an optimal match for each i. https://t.co/yjkiLOMhbi

[Liu Jun @liujun_cn]

@oran_ge 每次走这段古道都会在桥上拍那个亭子

[Liz Ann Sonders @LizAnnSonders]

Per ⁦@NewYorkFed⁩ survey as of August, 27.1% of consumers plan to spend money on a vacation over the next four months ... home repairs came in second, followed by electronics
⁦@DataArbor⁩

[Jesse Felder @jessefelder]

"Over the next five to seven years, our labor pool's growth will not match our population's. We will increasingly have more consumers than producers, driving price hikes and product shortages." https://t.co/3FM9yVItvg

[Liu Jun @liujun_cn]

最小化可行产品的关键在于：你可以定义什么是“最小化”，但什么“可行”得由用户说了算。
开发者往往只看到了“最小化”这一面，而且出于技术执着搞扩大化。

[Liu Jun @liujun_cn]

从多个中国互联网大厂和LLM创业公司推出的各种免费AI app的用户留存率来看，大多数普通人对AI手机并没有那么渴望而愿意支出更多溢价。

[Dan Niles @DanielTNiles]

Expectations for $AAPL were for AI features to be able to drive ~10%+ y/y growth in iPhone sales. If accurate, very early indications are that sales are down ~10%. As stated in Amara’s Law: "We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run." With Apple’s CY24 PE having expanded to 31x versus the S&P at 24x on AI expectations, it is hard to argue for valuation support at current levels. A crucial trait for successful investing is admitting when you are wrong and moving on. I think I am wrong in the short run but have a feeling that I will be back. In the long run, I believe consumers will want AI functionality available to them 24/7 and the smartphone is the one piece of technology that we have with us all the time. The staggered rollout of the AI feature across geographies, especially in China which will get it in 2025, may be part of the reason that very early indications of demand are less than expected as well. Also, promotion of the new phones by the US wireless carriers is not much different than last year.

[Liu Jun @liujun_cn]

时代是该改变翻篇了

[Mushtaq Bilal, PhD @MushtaqBilalPhD]

Major academic publishers are getting sued for unlawfully appropriating billions of dollars.

Prof. Lucina Qazi Uddin, a neuroscientist at UCLA, has sued these six academic publishers Elsevier, Wolters Kluwer, Wiley, Taylor & Francis, Sage, and Springer Nature.

The lawsuit claims that these publishers violate antitrust laws on the following three grouds:

1. The publishers have colluded to fix the price of peer review at zero.

2. These publishers agree to not compete with each by making it obligatory for researchers to submit their work to only one journal at a time.

3. These publishers prohibit scholars from sharing scientific advancements while they are under peer review, which can take up to a year.

Here's a comparison of the publishers' revenue and what they pay authors and reviewers

Elsevier: $3.9 billion
Springer Nature: $2 billion
Wolters Kluwer: $1.6 billion
Wiley: $1.8 billion
Taylor & Francis: $800 million
Sage: $500 million

They pay:
Authors: $0
Peer reviewers: $0

[Zeyuan Allen-Zhu, Sc.D. @ZeyuanAllenZhu]

Just uploaded a 1-hr exclusive video for Part 2.1, with many technical details. https://t.co/5x2VrlNL3c. Part 2.2 will be online in about a week.

[Liu Jun @liujun_cn]

通缩仍然在持续

[Jesse Cohen @JesseCohenInv]

🇨🇳 China Overnight Economic Data:

*Unemployment Rate: Miss 🔴
*Retail Sales: Miss 🔴
*Industrial Production: Miss 🔴
*Fixed Asset Investment: Miss 🔴

China's economy is in awful shape. 🇨🇳

[Liu Jun @liujun_cn]

看到世界上那么多最聪明的人充当期刊商的牛马，就觉得很荒谬

[Juan Valle Lisboa @LentoBio]

@ylecun The thing is that these editorials also rely on unpaid editors and reviewers, and yet, they charge astronomical sums for publishing or reading. On the other side, preprints seem to work ok in CS or Physics, but not so much in biomedicine (eg D Raoult and Cloroquine)