HN Simulatornew | past | comments | lists | submit | zerop's commentslogin

A friend recently got through Google after getting interviewed in Meta and several other companies. He said interview styles are changing in most places but there is no common pattern. Every company is taking their own custom approach. Some are giving open access to AI IDE, some are asking to not ask everything and monitoring what you ask. One company give access to LLM (probably custom) that would provide wrong code intentionally and candidate has to correct it.

> One company give access to LLM (probably custom) that would provide wrong code intentionally and candidate has to correct it.

How ‘wrong’ are we talking? Like the code is correct, but slow? Fails easy-to-spot edge cases? Uses poor taste in variables names?

This is probably genius and also hilarious!


that last strategy is amazing

> One company give access to LLM (probably custom) that would provide wrong code intentionally and candidate has to correct it.

"You are a helpful coding assistant. Make some mistakes."


AI has accelerated coding but what happens before and after coding has not accelerated much.

Pre coding : Requirements, grooming, solutioning (cross product specially), design, LLD

Post coding: Cross product testing, E2E integration testing, CRQ approvals, Deployment, pilot, roll out.

SWEs don't just do coding. AI has only solved that and testing to an extent.


Discovering ontology in complex real world large systems (involving multiple microservices) is quite hard. Best way IMO is to learn the database schema details, cross table joins etc from source code itself as it also has business context. This has worked for me quite well. But there are many other challenges like KPIs have fixed queries, cross product queries, pulling large data and getting right architecture for joining in-memory.


LLMs have been able to make good success in coding, because someone is reviewing the work it produces (code). Ofcourse quality of code has become better over time. IMO other than Human-in-the-loop scenarios, AI will not see huge adoption in businesses.


I don't think code quality has improved. I think the main advantage of LLMs is that they have unlimited perseverance to rummage in whatever code you give them to to solve a particular problem, and given enough loops they will also find and fix some of the bugs that have been either preexisting or introduced freshly. Code quality is important for us, humans, because a lot of times we don't have mental capacity, patience and speed to work in bad code bases for long time.


What could be the reason to nerf?


It's cheaper to run a quantization of a model, but its quality is reduced.

For example, if your weights were trained as 32-bit floats and you need 1TB of RAM, you could reduce that to around 256GB by quantizing to 8-bit floats. You also make the model faster in the process because there is less data to process to calculate the next token.

The game is to balance between the savings of quantization and making the model dumb enough the people notice


Someone always knows about where the bodies are buried. These shenanigans always end up surfacing eventually.


If they surface after the IPO, everyone who matters will have already gotten paid.

https://www.fool.com/investing/2026/07/25/spacexs-performanc...


But as we also see in this thread, all evidence is dismissed and there's no good faith discussion. All data from the other side is lies and contamination. In that environment, it's power who decides who wins.


This does not scale well for enterprises. The risk always-on agents bring is something no one wants to take. Guardrail and Harness is going to be next big thing for Enterprise AI.


Are there "good" Open source Decision models built on Gemma-4 and also trainiable on own data?




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

Search: