Siqi Miao

ML PhD student @GeorgiaTech, working on geometric & graph deep learning | Previously @Purdue @UofIllinois @TencentGlobal

[Greg Brockman @gdb]

GPT-4 finished training four years ago today.

[Limichange @Limichange2]

GPT 5.6 Sol Max 写代码非常的防御，属于脑子不是那么灵光，但是劲大。你裤子漏个洞，他通常选择掏出一块钛合金帮你补上，你告诉他应该用块布来补，他会说“对，这样更好”，然后飞快把钛合金拔掉，用布补好。

[Nima Shoghi @nimawrites]

Excited to share STAR-MD: a scalable autoregressive diffusion model that generates stable, high-quality protein MD trajectories at microsecond timescales, where existing methods fail catastrophically.

Accepted at ICLR 2026! Links and details in the thread 🧵👇

1/7

[Dongxi 东锡 NLP @dongxi_nlp]

https://t.co/DR6qCCuE4s

[Siqi Miao @siqim311]

I am attending ICML’24 in Vienna to present our oral paper: Locality-Sensitive Hashing-Based Efficient Point Transformer with Applications in High-Energy Physics, https://t.co/brQxxkvXzB.

DM me if you’d like a chat!

[Mckay Wrigley @mckaywrigley]

From now on I’m now going to be using AI to build software by simply drawing it.

[OpenAI @OpenAI]

ChatGPT can now see, hear, and speak. Rolling out over next two weeks, Plus users will be able to have voice conversations with ChatGPT (iOS & Android) and to include images in conversations (all platforms).
https://t.co/uNZjgbR5Bm

[Ninareh Mehrabi @NinarehMehrabi]

What should be a factor to get a faculty job? (3rd option is when your name is on all publications from your lab)

[Jim Fan @DrJimFan]

GPT-4 has one emergent ability that is extremely useful and stronger than any other models: self-debug.

Even the most expert human programmer cannot always get a program correct at the first try. We look at execution results, reason about what's wrong, apply fixes, rinse and repeat. It is an agent loop: we improve our code iteratively given the environment feedback.

I highly recommend this paper: "Demystifying GPT Self-Repair for Code Generation", which quantifies GPT-4's self-debug capabilities against other LLMs. Some key findings:

▸ The core reason that GPT-4 can self-repair is its powerful feedback ability. It is able to self-reflect effectively about what's wrong with the code. No other models can compete.
▸ Feedback model and code generation model do NOT have to be the same. In fact, feedback model is the bottleneck.
▸ GPT-3.5 can write much better code given GPT-4's feedback.
▸ GPT-4 itself can write much better code given expert human's feedback.

It's very likely that OpenAI is training the next GPT by hiring lots of software engineers as teachers in the loop. They don't need to generate. Critique is all you need.

[Paul Couvert @itsPaulAi]

ChatGPT is used by only 2% of people who have access to the internet.

During this time some people use the right prompts and save a lot of time.

Copy-paste these advanced ChatGPT prompts to get on top of the game:

[Rosanne Liu @savvyRL]

Thoughts?

[Aman Jha @amanjha__]

It's here!

Upload *any paper* to Explainpaper and start instantly getting explanations! Ask follow up questions if you need a more in-depth answer.

Go to https://t.co/C7rp4Vj9RS and go read all the papers you've been saving! 📝📝📝

[Michael Galkin @michael_galkin]

🔥 New blog: Graph ML papers at #ICML2022 in 10 categories: from denoising diffusion generation through transformers, expressive and explainable GNNs to algorithmic reasoning, knowledge graphs, and cool applications! Jointly with @zhu_zhaocheng
https://t.co/VhzNp0nzaL