Mathieu

creative mind into AI and working corporate. sharing whatever crosses my mind

[Mathieu @matbitzmat]

@retolipp @NZZ are the pisa results getting better and better in other OECD countries?

what's the baseline difference in PISA results?

* makes a statement across groups *
* makes a statement about one group relative to itself *
... and provokes causality ("nevertheless")

[Mathieu @matbitzmat]

@ipezyGJ yes, sorry i misformulated: i used the same set of rules for 5 judges but each had another task.
1/ detects indirect speech
2/ defines if should be indirect speech
3/ controls no 2/ for false positives
etc.

but all had the same definition of what indirect speech is

[Mathieu @matbitzmat]

LLM judges have the exact same issue (see previous prompt). Give one judge five different tasks and the score gets blurry. The judge itself has bias and run to run variation. So now I am judging one unstable output with another unstable output.

[Mathieu @matbitzmat]

@SamTechien @dev_maims mac tells you how long it's going to take
linux tells you what's happening
windows does neither

[Mathieu @matbitzmat]

@axhiii88547 i'd be careful and say: "let's wait and see"

[Mathieu @matbitzmat]

@neeraj7105 98% usage

[Mathieu @matbitzmat]

@neeraj7105 ter minal

[Mathieu @matbitzmat]

@neeraj7105 pro gramming

[Mathieu @matbitzmat]

@neeraj7105 cyber security

[Mathieu @matbitzmat]

@pauliusztin_ what motivated you to build your own from scratch instead of using cursor? cost?

[Mathieu @matbitzmat]

For my bachelor thesis, AI mainly increased the scope of what I could analyse. More than 15 million data points, 3,000 outputs and 30,000 lines of code. Four years ago I would have needed to write every script myself. With agents, I could just do way more.

[Mathieu @matbitzmat]

A lot of models can classify a bank transaction. The harder part in my finance app was everything around it: Swiss taxes, pensions, commune data, forecasts and the actual budget logic. What domain knowledge in your product took longer to build than the AI?

[Mathieu @matbitzmat]

Fable 5, Opus, GPT 5.6. All insanely capable. But if I give them one huge task and the output is wrong somewhere, what exactly do I fix? That is why I keep the steps small enough that I can still retrace them.

[Mathieu @matbitzmat]

I think people underestimate how hard it is to compare prompts when one prompt does five things at once. One part gets better, another gets worse, and now what? How do you even say which version is better?

[Mathieu @matbitzmat]

We had a scribe tool mixing reported speech and direct observations. I wanted to see if a new prompt improved it. I thought this would be a simple before and after test. Then the judge needed grammar, source context and speaker attribution. How would you build that eval?

[Mathieu @matbitzmat]

Your portfolio does not need to be full of polished products. Mine had my bachelor thesis, small experiments, Streamlit apps and tools with real users. Half finished stuff still shows how you think. Got one sitting around? Put it online and send the link.

[Mathieu @matbitzmat]

At uni you could already see it. Some went fully into ChatGPT. Some only used it for summaries. Some never touched it. Back then, whatever, everyone could still get good grades. Now I feel like that gap is getting pretty huge.

[Mathieu @matbitzmat]

One AI engineering internship I got had around 900 Quick Apply applications on LinkedIn. Your CV is sitting in a massive pile. What helped me was having a portfolio people could open and judge themselves. If yours is online, drop it below. I want to see it.

[Mathieu @matbitzmat]

Finding a first job in Switzerland is pretty brutal right now. One AI engineering internship I got showed around 900 Quick Apply applications on LinkedIn. If you hire juniors here, what do you actually look at after the CV?

[Mathieu @matbitzmat]

Detecting whether a sentence uses Konjunktiv is one thing. Deciding whether it should use Konjunktiv needs the original conversation, the speaker and the source claim. One tiny grammar eval suddenly needed the whole pipeline. How would you evaluate that?