Did the grumpy gamer mean perforce instead of perform and it got auto 'corrected'?

VMware owns Spring Boot https://en.m.wikipedia.org/wiki/Spring_Boot

Impressive! Can't wait to try Fray out at work.

Click the ying yang looking icon to get an explanation

> 4-7-8 Breathing The 4-7-8 technique involves inhaling for 4 counts, holding for 7, and exhaling for 8. This pattern is repeated several times. Developed by Dr. Andrew Weil, it helps reduce anxiety, manage stress, and promote better sleep by triggering your body's natural relaxation response and slowing your heart rate.

Some setups can run devstral small locally. Have you tried?

Also have you tried any other open weight models as the architect?

Third, as far as I understood, devstral was made with open hands so it might perform better there than aider out of the box.

Finally, I'm looking for local llm only setups for apple silicon 32gb and was considering devstral with open hands.

A lot of commenters point out that there already are many established static checkers that do this. That is not what Uber attempts here.

Uber is not proposing a static checker. They even use sonar qube in their architecture. They propose using an LLM to resolve the leak detected by sonar qube.

This architecture is promising. Large legacy code bases can have static analysis violations in the 1000s which devs never have the time to address. I've seen sonar tube reports that require man years to resolve everything.

These existing tools (spotbugs, findings, sonarqube, null away, checker framework, pmd, etc) provide detection but not resolution. Resolution thru LLM is what Uber proposes.

Resolution of the kind of resource leaks that SonarQube finds is usually the work of a few seconds. And it should be integrated into the developer workflow in such a way that those bugs are found and fixed before the code even gets checked in. In other words, if that code is even making it into a repo where GenAI can find and fix it then you have deeper engineering process problems. Fix the root cause rather than patching around them with more tools.

And once SonarQube flags a resource leak to a developer, a competent developer would learn from that and never make the same mistake again.

It's not identifying the leaks. In their architecture they use Sonar qube for the detecting. The LLM provides the fix.

Can the inference piece be partitioned over multiple hosts?

Edit: algorithmed or partitioned in a way that overcomes the network bottleneck

> prima.cpp is a distributed implementation of llama.cpp that lets you run 70B-level LLMs on your everyday devices— laptops, desktops, phones, and tablets (GPU or no GPU, it’s all good). With it, you can run QwQ-32B, Qwen 2.5-72B, Llama 3-70B, or DeepSeek R1 70B right from your local home cluster!

https://github.com/Lizonghang/prima.cpp

Pretty sure llama.cpp can already do that

I forgot to clarify dealing with the network bottleneck

Just my two cents from experience, any sufficiently advanced LLM training or inference pipeline eventually figures out that the real bottleneck is the network!