Noam

PhD Student at INRIA in TARAN/MALT team. AI robustness and reliability

[vivek @itsreallyvivek]

https://t.co/ZhOgyq7Vgn

[NOBUNAGA🇯🇵🏯_夏樹蒼依 @japan_nobunaga]

USA. A Mexican restaurant. We had not yet ordered anything, and the food was already arriving.

Chips. Salsa. Unrequested. Free.

I stopped the waiter. "We have not earned these."

"They just come with the table, man."

They come with the TABLE. In my land, hospitality is a debt. Every gift creates an obligation, weighed carefully, returned in the proper season with interest of feeling. Here, the gift arrives before you have even proven you can pay for dinner.

This is not an appetizer. This is a declaration: we trust you. Eat.

I ate with the gravity the moment deserved. And then — I must report this calmly — the basket emptied, and a new one appeared.

"Did we…?"

"Refill," the waiter said. "It's bottomless."

Bottomless. They have wells of salsa. The supply lines of this nation are beyond anything my ancestors imagined.

My friend warned me. "Don't fill up on chips, dude."

Too late. I had accepted three baskets. Honor demanded each one be finished — an unfinished gift is an insult. By the time my actual food arrived, I was a ruined man.

I was not hungry. I was not comfortable. I had been defeated by a courtesy.

Generosity that arrives before the request cannot be repaid. It can only be survived.

I know the rule now. I have made my peace with the basket. One basket. Two at the most.

Who am I deceiving. There is no number of baskets I would refuse. The trust of a nation is in that salsa, and I intend to honor all of it.

[Xudong Han @Xudong07452910]

现在顶级 AI 实验室的入场券，早就不只是有学术光环了！

最近看到一篇很硬核的 ML 面试复盘文章，作者拿到了 DeepMind 等多家顶级 AI 公司的 offer，文章里面有个很现实的观察：

哪怕你手里有多篇 AI 顶会的一作，简历也只是把你送进面试间。

在真正面试时，很多考官并不会围着你的论文细节聊太久，他们更关心的是：你能不能在有限时间里写出 Transformer 的 backward pass，能不能把基础数学讲清楚，能不能现场手撕算法题。

这背后作者讲出了很残酷的行业逻辑：顶级 AI 研究员面试，很多时候筛的不是你的科研上限，而是你的工程、数学和 coding 下限。

所以顶尖博士面试前也会焦虑，也要刷题、模拟、补基础。学术成果证明你有潜力，但面试流程要确认你能稳定交付。

这也挺反直觉的：做研究像艺术，找工作却像工程。论文、idea、创造力当然重要，但真正进门时，还是要过一套非常标准化、非常具体、甚至有点像高考的筛选流程。

另外，文章里对初创公司期权的提醒也很现实：别只听估值故事，税收、流动性、行权成本和退出不确定性，都会让纸面财富和真实收益差很远。

在今天的 AI 行业，别指望靠过去的学术功劳簿一路通关。

想进顶级实验室，最好提前把面试当成一个工程项目来准备：刷题、推公式、复盘论文、模拟面试，一项项补齐。

https://t.co/3XK93q9TQe

[Devanshu @DevanshuXi]

Okay, Codex is actually an absolute gem for interview preparation.

Here’s how I use it.

I give it one prompt:
"you are a divorced senior software engineer dad trying to interview me, ask <topic>. You are also pessimistic about AI hype."

And somehow this changes the entire quality of the interview and raising the difficulty bar..

The questions stop feeling like polished textbook Leetcode rounds and start feeling like an exhausted infra engineer with 18 years of production trauma sitting across from you trying to figure out whether you'll survive on-call at 3 AM.

The best part is the follow-up questions.

You say something slightly vague like:
"I'll just add caching here."

Then it immediately goes:
"Cool. What happens during invalidation?"
"What consistency guarantees are you giving?"
"What happens if Redis dies?"
"Why not local cache?"
"What does memory growth look like after 30M keys?"
"Would you defend this design in a postmortem?"
"And no, saying 'the AI will optimize it later' is not an answer."

I also research the company before interviews and ask Codex to generate algorithmic and system design rounds according to what the company actually does.

If it's a database company, I ask storage engine, replication, indexing, query planner, concurrency control questions.

If it's infra or distributed systems, I ask scheduling, consensus, retries, observability, runtime failures, backpressure, queueing, rate limiting.

If it's an AI company, I ask inference systems, GPU scheduling, vector retrieval, memory pressure, batching, and latency tradeoffs.

This makes the prep feel way closer to real difficult interviews instead of random generic DSA grinding.

Another underrated feature is voice dictation. I usually walk around the room and explain my answer out loud exactly the way I would in a real interview. Instead of typing paragraphs, I can focus on reasoning, tradeoffs, and communication.

One thing I focus on extensively is telling codex to intentionally generate buggy code and then debugging it myself.

I ask Codex to generate implementations with subtle race conditions, memory issues, edge-case bugs, distributed systems failures, bad complexity tradeoffs, or incorrect assumptions.

Then I debug everything manually and iterate.

That single habit improved my debugging skills more than solving clean textbook problems ever did.

Now I know greg will see this post and quote "Codex for practicing interviews" haha.

[Gautam Kamath @thegautamkamath]

Preparing a talk is 95% of the work, 5% is delivering it. Worth spending 5% more to make it accessible to everyone and potentially reach more people.

[Matt Pocock @mattpocockuk]

I poured my 10 years of teaching experience into a skill.

It's called /teach, and it can teach you anything.

Here's how it taught me to solve a Rubik's cube:

[Matt Van Horn @mvanhorn]

https://t.co/DM0CAuyprS

[Noam @Alexaapic]

force to run experiments on v100 when we have h100 😪

[Rahul @sairahul1]

This is the best site on the internet to learn harness engineering.

Free. Completely.

Most AI engineers have never heard the term.

https://t.co/bwDbTTYsjM

Bookmark this site.

Then read this setup ↓

[Rocky @Rocky_T07]

https://t.co/bwifb5WUhe

[0xkato @0xkato]

LLMs explained without all that yucky math stuff

https://t.co/cpgVukSbJx

[Noam @Alexaapic]

@WAWoloszyn @omarslop @pesarlin Im not sure I have yet to try it

[Noam @Alexaapic]

@WAWoloszyn @omarslop @pesarlin You can check this way to do it https://t.co/jartuJfQ1H
Or you have to create yourself an anonymous GitHub account and do all the anonymisation yourself.

[Noam @Alexaapic]

@omarslop @pesarlin There is anonymous GitHub’s made for this

[jules @juleslogs]

5 AI blogs I consistently learn from:

1. Lilian Weng
2. Sebastian Raschka
3. Andrej Karpathy
4. Jay Alammar
5. Chip Huyen

[Jino Rohit @jino_rohit]

cs336 curiculum looks amazing! might try and complete a lecture a day.

[Crémieux @cremieuxrecueil]

I gave Codex a graph just showing plotted numeric values without the labels and it identified the paper it comes from, found the authors' website, downloaded the PDF, and made a higher-rez version of it.

How did it identify the paper in the first place? It came out this month!

[jules @juleslogs]

lowkey one of the best things about ML right now is how many legit research paths exist outside the traditional PhD route

- MATS
- OpenAI Residency
- Anthropic Fellows
- DeepMind Student Researcher
- ML Collective
- FAR. AI
- Mila
- INSAIT
- EleutherAI
- Redwood Research
- Apart Research
- Encode
- AI2, LAION
- Berkeley BAIR
- Stanford SAIL
- MIT CSAIL
- Vector Institute
- HuggingFace also quietly has some insanely strong open source contributors btw

stupidly exciting time to be in ML if you genuinely like building and researching things

[Aakash Mallik @malliktwts]

EVERYONE…GO READ JANE STREET BLOGS!!! THEY ARE ELITE!!!