Hannu Varjoranta

[Hannu Varjoranta @hvarjoranta]

Jev is a good idea: stop using generative models for decisions that should return probabilities.

I looked at the architecture, calibration and open reproductions.

My take: useful pattern, strong product, narrow moat.

The model assesses. Code decides.

https://t.co/8NRKw8vzSP

[Hannu Varjoranta @hvarjoranta]

@avreacom Ship fast with Avrea!

[Avrea @avreacom]

@namespacelabs Nice unboxing. We call that Tuesday. 😏
Glad to see the laptop strategy is catching on 😂

[Hannu Varjoranta @hvarjoranta]

Full piece, with the numbers and what the other industries actually did:

https://t.co/fFpqH5AWX4

[Hannu Varjoranta @hvarjoranta]

31% more pull requests now merge with no review at all.

That is Faros AI telemetry across ~22,000 developers as #AI coding tools arrived. Throughput genuinely rose. #Review did not keep pace, and it did not degrade gracefully. It got skipped.

[Hannu Varjoranta @hvarjoranta]

The practical version is two questions rather than one.

Is there a fast test that settles this?
Can it be undone?

Automate hard where a test exists. Spend human attention where none does and the door only swings one way.

[Hannu Varjoranta @hvarjoranta]

because benchmarks are public and competitive.

So the routing question is whether the property has a fast oracle. If yes, automate it. If no, that’s where the human stays.

[Hannu Varjoranta @hvarjoranta]

“More review raises the floor, it doesn’t move the ceiling” is the line.

Worth adding: a reviewer from a different lab doesn’t move it either.

Different lineage decorrelates knowledge errors, not objectives. Labs differentiate on data and converge on evaluation, …

[dex @dexhorthy]

gave up on waiting on nikita for the articles fix - part 1 is here part 2 is coming https://t.co/8aS4vN6LK3

[Hannu Varjoranta @hvarjoranta]

We just shipped M5 Max runners for macOS and ARM Linux CI.

GitHub tops out at M2 Pro for macOS. We're running M5 Max for both macOS and native ARM Linux builds. Faster hardware, lower price.

https://t.co/NpIQ225QKU

[Hannu Varjoranta @hvarjoranta]

We came out of stealth with Avrea!

The CI I wish existed when I was paying the GitHub Actions bill. Faster runners, lower cost, one line to migrate.

https://t.co/MFBwJJv2Px

My favorite feature is ability to SSH directly into your runner. 30 sec to setup, one line diff in your Github workflow.

[Hannu Varjoranta @hvarjoranta]

A salary was never just money.

It was proof that someone needed you tomorrow.

What happens when survival no longer depends on employment?

Link in reply.

https://t.co/MQlyzR3iMI

[Hannu Varjoranta @hvarjoranta]

GGUF is the boring choice for running a 70B at home. It's also the right one for almost everyone, and the reasons aren't what the benchmarks suggest.

Quantization stack, May 2026: what works, what's a trap, and the math corrected.
https://t.co/vp31nS9MOH

[Hannu Varjoranta @hvarjoranta]

This is uncomfortable for engineers who built careers on implementation prestige.

It should be.

The engineers who thrive now will not be the ones who can merely build software.

They will be the ones who understand exactly which human pain is expensive enough that removing it changes behavior.

[Hannu Varjoranta @hvarjoranta]

Software is getting cheaper. The market is also done being impressed by it.

For years, technical skill scarcity warped the economics. Building distributed systems, scalable backends, internal tooling. These commanded attention because few people could.

Implementation itself carried scarcity value

[Hannu Varjoranta @hvarjoranta]

The technical layer does not disappear.

It becomes substrate.

Important. Invisible. Worth less per hour to anyone outside the discipline.

Engineering prestige migrates underground while public attention chases the next demo.