Hello fellow ML geeks, my name is Andrey. Each day, I understand more about how little I know. Time to cover another gap - scaling LLM training. Most of the time I just plug in a library or ask an expert colleague. And... I still want to do the same thing, but finally have a much better grasp on what is going on under the hood. Join in!

How it works:

- Read a chapter from the list.

- Jump on a call once a week on Thursdays at 14:00 (London time)

- Take part in the discussion and learn something.

- Message here if you want to present one of the chapters (pretty pretty pretty please!)

- All slides will be uploaded to Github

First call: October 30 (Thursday), 14:00 (London time)

Important links:

- Schedule https://docs.google.com/spreadsheets/d/157L7VW-2M1KcwJoRtzadagpJ7LOpos338oEwj8kp2XQ/edit?usp=sharing

- Calendar https://calendar.google.com/calendar/u/2?cid=ZTM4ZmMwYTM5MDhmZjkwNjcwY2EyZDRlNmQzYThjODZmM2I2NzZkMzZiNWExNzc4NWY3MTE4NTk3MDJkNWJlYkBncm91cC5jYWxlbmRhci5nb29nbGUuY29t

- Repo with slides https://github.com/fxlrnrpt/little_ml_book_club

- Chat invite https://t.me/+bleQ2wfB4DBjMGZk

Comments URL: https://news.ycombinator.com/item?id=45695629

Points: 8

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=45610603

Points: 1

# Comments: 0

Sadly, internet became fragmented over the last few years. You move from one country to another and to your surprise find that REPLACE_WITH_YOUR_FAVORITE_RESOURCE is all of a sudden is not available anymore. WTF?

It is a minor inconvenience when it is a feed with cat memes, but a major problem when it is, say, Youtube.

About a year ago I left London to go back to school. As you can guess, the school network (network provider actually, but who cares?) had an extensive blacklist. It hurt my studies and studies of other students. I decided to setup a VPN for me and other students. Well, I am a damn computer science student, aren't I?

Below if the blueprint on how to setup an VPN for your entire org (my uni in this case) in broad strokes.

You will need:

- Any VPS with a good network connection;

- marzban (https://github.com/Gozargah/Marzban) - SOTA proxy management tool powered by [Xray-core](https://github.com/XTLS/Xray-core);

- marzban-verify (https://github.com/aigoncharov/marzban-verify) - sidecar for mazrban to create new accounts with email verification;

- Telegram account.

What you will get:

- A Telegram bot where people can create VPN accounts on their own as long as they have access to their org emails. For instance, you can provide VPN access to all email holder at `@wtf.com`.

- Sane defaults with a 3 month expiration date and 50 GB traffic limit for new accounts (can be changed).

The blueprint:

- Find a decent VPS. Do not trust speed filters on aggregators. They lie. I had to manually search Reddit for a shortlist of suitable providers and then test them one by one. Do not go with the largest ones if you want your VPN to last. The largest providers are the first targets for censors.

- Install marzban (https://github.com/Gozargah/Marzban?tab=readme-ov-file#insta...). Set `DOCS` to `True`.

- Get SSL (https://gozargah.github.io/marzban/en/examples/marzban-ssl) and optionally set it up to work on one port (https://gozargah.github.io/marzban/en/examples/all-on-one-po...). I have a simpler Haproxy config that still works well:

  ```
    listen front     
     mode tcp     
     bind \*:443      
     tcp-request inspect-delay 5s     
     tcp-request content accept if { req_ssl_hello_type 1 }

     use_backend marz if { req.ssl_sni -i end  ADDRESS_OF_MY_VPN_SERVER }
     use_backend reality 
     
 backend reality     
  mode tcp     
  server srv1 127.0.0.1:12000 send-proxy-v2 tfo  
  
 backend marz     
  mode tcp     
  server srv1 127.0.0.1:10000
 ```

- Setup marzban-verify (https://github.com/aigoncharov/marzban-verify?tab=readme-ov-...). Currently, it can send confirmation emails by setting up a standalone SMTP server (prone to being identified as spam, also some VPS hosters block port 25) or by using your own Exchange email (popular in enterprise envs).

- Profit!

No wise closing words. Just send your PRs if you need more email providers.

An actor misbehaved - you stop going to its movies. A company got involved in some shady shit - you stop buying from it. Sounds like sound logic aligned with an image of having principles, right? Well, how about not allowing folks into academic conferences? And not because they personally did something you do not like, but because they come from a certain country?

My name is Andrey. I am ex-big tech software engineer who decided to switch to academia and went back to school. One of the best tech schools in Russia - Skoltech. The one that is currently being cancelled.

Long story short:

My extremely talented colleagues (the ones I look up to on daily basis) have been rejected from CVPR (a top tier academic conference) purely because of their affiliation with Skoltech.

Specific papers:

- Switti: Designing Scale-Wise Transformers for Text-to-Image Synthesis (Yandex, HSE, MIPT, Skoltech, AIRI, ITMO); - Color Conditional Generation with Sliced Wasserstein Distance (Skoltech);

- MaterialFusion: High-Quality, Zero-Shot, and Controllable Material Transfer with Diffusion Models (HSE, AIRI, Skoltech);

- Good Keypoints for the Two-View Geometry Estimation Problem (Skoltech, Slamcore).

Screenshot - https://pbs.twimg.com/media/Gpm-cEdWQAIU4uU?format=jpg&name=large

Why it matters:

There are only a handful of respected conferences that truly matter in academia. You want to be heard? Publish there or perish.

Why it is wrong:

As humanity, we are where we are because even at darkest times we find a way to talk and collaborate. If you do not like someone personally, surely you do not have to play with them. It is basic free will. But do you want to live in the world where people cancel others based on their origin as a whole group? What if it is not about business and profit, but about making the world a better place together through science?

One could argue that it is only Skoltech - a single uni in Russia. Guess what? It is not. Other top-tier schools (HSE, MIPT) are under sanctions and face the same treatment. And not only schools. A few companies that have the resources to fund the research are in the same boat. You want to do good research in Russia - not too many options.

I get that conferences hosted on the US and EU soil have to comply with the laws. It is inevitable. Yet these laws are there to restrict for-profit business, not academia!

The world is dark and full of terrors. It is hard to navigate living from one crisis to another. Yet people somehow manage to keep their obsession with science. They keep grinding and want to share their findings with the world. Since when this kind of behavior should be punished? Could you remind me when division of millions of people into "good" and "bad" worked well? Ever? WTF?

Final message:

World is complicated and unfair, but we can make it better bit by bit. Not through division, but through staying connected and doing our best to find common ground. Science and non-profit are universally good for all of us. Shall we start there?

Comments URL: https://news.ycombinator.com/item?id=43819002

Points: 5

# Comments: 1

I am a Master's student in Data Science in a great school in Russia (Skoltech). Started it to switch to ML research after 9 years in the industry as a software engineer and so far did not regret it one second. Weighing my options for a future PhD. Geography does not really matter. It might easier for me to do it in Russia or in the UK since I lived in both countries, but I am open to all opportunities. What are the places/programs with decent PhD stipends to live comfortably?

Here is what I know of:

- In the UK, Meta offers collabs with UCL and Oxford (see AI Research Assistant on their website). AFAIK, they offer IC3 level comp for the role - 60-70k (ish) GBP. It is more than 3x of the regular PhD stipend in London (~20k GBP) and above median salary in the city; - Same Meta collab in France (do not know the name of the school); - In Russia it might be possible (but not trivial) to find a scientific advisor from the industrial lab (AIRI, Yandex Research, etc.);

What else?

I would enjoy doing research after graduation, but if we take London, to be honest, I am not ready to survive on 20k there (a typical PhD stipend). When I was there last year, my en-suite room in a good place was ~1450GBP after bills. Which means that the entire stipend would go towards rent. I could take a hit financially and burn my savings or rent in a much more modest place for a year or so, but not so sure about committing to it for 3-4 years.

Comments URL: https://news.ycombinator.com/item?id=43270649

Points: 1

# Comments: 3

> Meanwhile deepseek is hiring college grads and the like, for way better tech.

If only I knew Chinese...

> I highly doubt that they've got a better ai but are too scared to show it off. Is there any independent verification of that?

Nope. Just the rumor

Comments URL: https://news.ycombinator.com/item?id=43026464

Points: 44

# Comments: 29

Comments URL: https://news.ycombinator.com/item?id=41044336

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=40882089

Points: 4

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=36069384

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=35142459

Points: 1

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=32808158

Points: 3

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=32721499

Points: 9

# Comments: 19

Comments URL: https://news.ycombinator.com/item?id=32692290

Points: 3

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=31450668

Points: 2

# Comments: 0