Yevgen Nerush

Principal Engineer @ Mendix, Siemens

[Haider. @haider1]

Creator of C++, Bjarne Stroustrup:

AI-generated code isn't ready — it generates more bugs, more bloat, more security holes, and is nearly impossible to validate

"senior developers are already retiring rather than deal with it"

The problem is that even a small prompt change can shift the entire codebase in unpredictable ways

[tuhin @tuhinSubmitra]

@nerush is talking about how Mendix AI assist works at #MxWorld2019. We know the science behind it, to us AI is not a gimmick!

[Physics In History @PhysInHistory]

Timelapse footage of neurons making new connections. This is what your thoughts probably look like, quite literally.

[Yevgen Nerush @nerush]

Port of Facebook's LLaMA model in C/C++ with bindings to other programming languages

https://t.co/OoGr12zapR

[Johannes Rudolph @virtualvoid]

A Scala port of @karpathy's llama2.c: https://t.co/C6DsHB5GLI Even without optimizations performance is surprisingly acceptable (-6%, single-threaded on GraalVM). Example: https://t.co/D5wABII06M

[Jake Luciani @tjake]

Pushed the first cut of my weekend hack project. Well like 3 weekends.

It’s called LLM-J a pure Java LLM inference engine. I wanted to learn how LLMs work and try out Java 20+ so here we are.

It supports GPT-2 and Llama 2 (well almost).

https://t.co/LnwkWew5NU

If you’re a Java developer and your brain can’t decipher PyTorch like mine this may help you understand better.

Biggest takeaway from me is dot product is literally 95% of the cpu time.

Speaking of CPUs, this project has made me think maybe the future isn’t all GPUs. Sure you need them for training. But for inference and perhaps LoRA a distributed inference engine could be workable and cheaper to run.

#java #llm #transformer #ai #oss

[Yevgen Nerush @nerush]

safety of LLMs as coding copilots is better for languages where higher order structures (arrays, dicts) are not mutable across function calls like #scala #dsl https://t.co/d12t5ttGzc

[Yevgen Nerush @nerush]

Making GPT model bigger does not mean it becomes smarter. It just has more storage for memorizing the answers.

[Yevgen Nerush @nerush]

@niw @saiedg @antimatter15 @ecjwg @ggerganov Is `ggml-alpaca-7b-q4.bin` available for download? https://t.co/JSrdUSkOvi repo does not have any hyperlink to the file

[Yevgen Nerush @nerush]

@BeLIrina Это достоверная информация. Я лично подтверждаю

[Yevgen Nerush @nerush]

https://t.co/FKF4h9xp13

[Yevgen Nerush @nerush]

@ScalaCon @odersky @scala_lang Is there any video recording available?

[Yevgen Nerush @nerush]

After 8 years of work, 28,000 commits, 7,400 pull requests, 4,100 closed issues – Scala 3 is finally out. Thank you #scala 3 team for being open for the contribution! Glad to be part of it
https://t.co/C3r2HN3TAv

[James Ward @JamesWard]

Just blogged about my journey to a globally distributed, scale-on-demand, edge-cached, webapp; bringing together: GraalVM Native Image, Scala, Buildpacks, GitHub Actions, Cloud Run, and a bunch of other stuff.
https://t.co/W1lrmQXsvo

[Ivan Zub @zubivan]

That would be a fine ending to 2020 black mirror episode! #BostonDynamics #BlackMirror

https://t.co/tqdkbvxDo3

[Andrej Karpathy @karpathy]

How to become expert at thing:
1 iteratively take on concrete projects and accomplish them depth wise, learning “on demand” (ie don’t learn bottom up breadth wise)
2 teach/summarize everything you learn in your own words
3 only compare yourself to younger you, never to others

[Xebia @Xebia_Global]

Are you getting ready for #Scala3?

Make sure to review the evolving @scala_lang Migration Guide! The Scala Center is looking for community feedback and help filling out any missing content in preparation for the big release!

https://t.co/28eU8tbmuR

[Yevgen Nerush @nerush]

#scala #adt in the form of sealed trait and case class comes to #java in the form of sealed class and record. With #scala3, #java becomes #scala2 idiomatic and #scala3 introduces brand new programming patterns, unavailable in the other mainstream programming languages

[Yevgen Nerush @nerush]

if you can run a neural network model at an energy cost of below 1 mW, it makes a lot of entirely new applications possible

[Yevgen Nerush @nerush]

@copyconstruct I found at least 6 (28%!!!) of #gof design patterns superseded in my favorite #scala programming language: #singleton, #builder, #adapter, #chainofresponsibility, #observable, #prototype. How many of 21 are superseded in your favorite programming language?