Beer van der Drift

Software at @getcontextdev (YC S26)

[Beer van der Drift @BeerDrift]

You know you're enjoying your new work at @getcontextdev if you're devving late at night while watching Portugal - Spain ⚽️

[Bela Wiertz @blwiertz]

The best time to start building is now!

If you want to lock in for 3 months and just ship & sell, join us in Berlin at The Sandbox.

3 months in a room together with other builders.
No strings attached.
Just a support system for upcoming builders.

Everything from pure software to robotics, but no need for boundaries in your thinking.

You can join alone or with your co-founder.

Supported by Anthropic, OpenAI & Runpod (means you'll get some credits)

⬇️

[Beer van der Drift @BeerDrift]

@mynameisyahia @boardyai Well, you heard the Board Yahia!

[Beer van der Drift @BeerDrift]

@dominikkoch @getcontextdev Happy to be of service 🫡

[Beer van der Drift @BeerDrift]

We're live! https://t.co/nLei3eeLs6

[Beer van der Drift @BeerDrift]

@leodev @mynameisyahia @Stock_Alarm 100%

[Beer van der Drift @BeerDrift]

@elymadebynormal @mynameisyahia @Stock_Alarm Thanks thanks <3

[Beer van der Drift @BeerDrift]

@mynameisyahia @Stock_Alarm Really excited to go on this adventure with you and the team! Awesome stuff on the roadmap! 💃

[Beer van der Drift @BeerDrift]

@nielsgeo @cursor_ai Today is my first day using it! Man it's a hose at max pressure haha.

[Beer van der Drift @BeerDrift]

@futurenomics I've never seen the difference put into words so well!

[Beer van der Drift @BeerDrift]

@dwarkesh_sp That's exactly what world models like JEPA are for, to massively reduce the amount of training required.

[Beer van der Drift @BeerDrift]

@mynameisyahia Models are great at building using chopsticks and duct tape. Right now it still requires human eyes for solidifying it to the level of reinforced concrete so you can build on top.

[Beer van der Drift @BeerDrift]

@mynameisyahia Don't behave

[Beer van der Drift @BeerDrift]

@mynameisyahia Ouch. Hits home. Just crossed this exact same threshold for the first time just now.

[Beer van der Drift @BeerDrift]

Well, this time it should work right?

[Beer van der Drift @BeerDrift]

It's been much harder training an addition model than I thought.

First ~88% accuracy, then ~95% accuracy and now ~97.5% accuracy.

But how does this 4.7M model vs established open source models compare in a+b addition? Is it feasible to get 100% perfection?

I rented some RTX Pro 6000 GPU's to test!

Turns out; Addy is actually already quite advanced!

She does better than Qwen3 8B (~96%), but worse than Gemma 4 26B (~99.6%). Those are full language models of course, but are also about 2000x and 6000x bigger.

How much further can I push it? Does she just require even more training time? Or do I need to do something else?

[Beer van der Drift @BeerDrift]

So fun to see Addy training

[Max Schmitt 🍰 @Mackseraner]

How a Potential Client Almost Cost Me My Google Account:

Yesterday I received a project request through my website.

(thread below)

[Beer van der Drift @BeerDrift]

I always chuckle when coding agents blame every change I do on the 'linter'. "The linter refactored the code to be much simpler"

[Beer van der Drift @BeerDrift]

@typesfast KLM flights do have full free internet in all classes.