Hacker Newsnew | past | comments | ask | show | jobs | submit | chs20's commentslogin

Location: Tübingen, Germany

Remote: Yes

Willing to relocate: Yes (Europe/UK; open to exceptional opportunities elsewhere)

Technologies: Python, PyTorch, NumPy, transformers, multimodal ML, LLMs/VLMs, adversarial robustness, AI security, red-teaming, research

Résumé/CV: https://chs20.github.io

Email: schlarmann.christian@gmail.com

PhD researcher in ML at the University of Tübingen seeking Research Scientist / Research Engineer roles. I work on robust and secure machine learning, with a focus on multimodal models, adversarial attacks and defenses, and AI safety. Publications include ICML and NeurIPS; I also have hands-on experience red-teaming frontier AI systems. Open to AI security, multimodal/generative ML, and strong research-oriented ML roles more broadly.


Interesting that someone with your CV is looking to get hired on HN?

There should be lot of options in Germany currently, what I see when screening some boards?


Yes, surely. Just using this as another channel, often the most interesting opportunities are rather hidden.


I wish you all the best, esp. since the jobmarket is worsening right now!


Transformer / attention block is missing


On my todo list to add this as another level.


In my experience, papers published in these "special issues" don't really get any attention.


Simulating cross-wind landings doesn't sound that hard, why do you think current sims don't do it well?


Because I practice them a lot and the dynamics of a real x-wind landing is not reproducible in a home sim. You can practice where to put the controls. But in a trainer like a 172 or DA20 you can’t practice the forces you feel pushing back at you and the other seat-of-pants aspects.

One of the drills I do is to descend to 50ft over the runway and then hold the plane on centerline in a eg 20kt crosswind. The sim just can’t even come close to how dynamic and effective that exercise is. Also the feeling of putting down one wheel first, feeling the bump, then the other. The slight skid and side g forces of you mess it up slightly. Those kinds of things.

There are many things that sims can’t reproduce that don’t matter that much when it comes to productive learning. But with cross wind landings, these things matter a lot.

I’m specifically referring to home simulators and MSFS and X-plane specifically. I have no experience in full motion commercial jet or military sims.


> the forces you feel pushing back at you

A few other posts mentioned that even the most sophisticated sim machines cannot simulate (de/)acceleration. Do you think "the forces you feel pushing back at you" here could mean deceleration?


Seems like Microsoft is getting the rest of OpenAI for free now.


Why does he need to subsample? Seems like it should be feasible to find the minimum of a 6M array


Will be interesting to see if they have taken any precaution in terms of adversarial robustness in particular to vision input.


From which database are the papers that it searches?


According to the FAQ [1] the source is the "Semantic Scholar Academic Graph dataset".

[1] https://elicit.org/faq#appendix-how-does-elicit-work


What is the motivation for lambda calculus?


It was an attempt to find the simplest possible mathematical system that was universal, i.e. could compute any function. It turns out that the answer is that function application by itself is universal (if functions are first-class entities). This was a surprise at the time, and it still generally surprises people today when they first learn of it.

You might find this interesting:

https://flownet.com/ron/lambda-calculus.html


A little bit of historical background.

There was a concerted effort in the late 19th, early 20th century (perhaps earlier too) to mechanise computation i.e., reducing it to a pure symbolic manipulation. There were obvious benefits, a famous one being Enigma Machine that was used to (successfully I think) break the German code during WW-2.

On the philosophical side a parallel and overlapping effort was going on to figure out if mathematics could represent all possible truths, again as symbolic manipulation system. Bertrand Russel's magnum opus Principia Mathematica [1] was one such famous work towards it. Kurt Gödel then made a breakthrough when he proved that such a system is impossible; i.e., a system can either capture all the truths or it can be consistent but not both[2]. Put differently, any mathematical system capable of representing all the truths will necessarily contain contradictions within it.

Now coming to your question.

Lambda calculus emerged in this milieu. Alonzo Church[3] invented one such system to mechanise computation, named ƛ-calculus. Using this system one can mechanically compute any function purely by symbolic manipulation. Later on Turing, Church's student I think, invented a totally different system named Turing Machine with the same purpose. Later on it was proved (by Church and Turing I think, but I'm not 100% sure) that ƛ-Calculus and Turing Machine are equivalent, Church-Turing thesis[4].

All these work, and more, laid the theoretical foundation for the modern computers. If we can today safely assume that computers are provably correct it's because of them.

Phil Wadler has an absolutely delightful talk where he takes us through a whirlwind tour of the history of the mathematical foundation of computers [5], highly recommended.

[1] https://en.wikipedia.org/wiki/Principia_Mathematica

[2]https://en.wikipedia.org/wiki/Gödel%27s_incompleteness_theor...

[3] https://en.wikipedia.org/wiki/Alonzo_Church

[4] https://en.wikipedia.org/wiki/Church–Turing_thesis

[5] https://www.youtube.com/watch?v=IOiZatlZtGU


ooh! phil wadler lectured me at glasgow.. he was always listenable but he's got funnier...


But only regional trains. To me the biggest advantage is not having to worry about what tickets to buy in foreign cities.


To give people an idea what regional trains means:

On regional trains you can travel Germany from north to south in less than 14 hours with 5 changes as opposed 8 hours with 2 changes for 125 EUR one-way on long-distance trains.

With the 49 EUR/month ticket you can hop off the train anytime you are tired of the journey. Can't do that with the 125 EUR one-time, one-way long-distance train ticket.

I totally see young and elderly people do this on regional trains.


Consider applying for YC's Summer 2026 batch! Applications are open till May 4

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: