Nico Axtmann

Building conversational commerce @certainly_io. Sharing random thoughts and sometimes it is about #data/#ai

[Nico Axtmann @_nicoax]

@kaixhin I strongly believe that people become highly professional not due to degrees but due to their ability for the passion of learning process and tinkering.

[Nico Axtmann @_nicoax]

Coding with llms is so hard when the methods of the training data is deprecated... need to read docs again

[Daniel Bourke @mrdbourke]

Mindful machine learning: switch things up sometimes and let your model train on CPU.

Then go for a 2-3 day hike in the woods.

When you come back it’ll be .7% done.

The perfect time for another hike.

[Nico Axtmann @_nicoax]

Spontaneously run my first 50km (+1600 vertical meters, below 6 hours). Happy.

[Nico Axtmann @_nicoax]

@florinpop17 I can recommend Freeletics!

[Nico Axtmann @_nicoax]

@THETYFRANKEL 👋👋

[Nico Axtmann @_nicoax]

@florinpop17 Yes, in a lot actually! Won some cool gear and prizes! For me, it’s great way to get connected with other developers while having fun coding something up that breaks every 20 minutes

[Nico Axtmann @_nicoax]

@JohnSigmon @HamelHusain @sarahcat21 @ledwards I'm using dvc

[Nico Axtmann @_nicoax]

@HamelHusain Have used it in prod but preferring to export the model to onnx and run it with the onnx runtime.

[Nico Axtmann @_nicoax]

@bhutanisanyam1 Congrats! You should definitely join strava and track your runs 🏃‍♂️

[Nico Axtmann @_nicoax]

Wood season finally started 🪓🪵

[Nico Axtmann @_nicoax]

@am29d Sure dm me with possible time slots and I’ll send you an invite.

[Nico Axtmann @_nicoax]

Anyone from ML twitter who knows if a #serverless GPU service for inference exists?

[Nico Axtmann @_nicoax]

@creeeeeeeeeeg @iamtrask @Stepsome Treadmill walking for reading, calls, videos, learning new things, writing docs/simple code/ reviewing code works great for me.

If I do very deep conceptional or thinking work I still prefer to sit down.

[Nico Axtmann @_nicoax]

@am29d Model size is more than 10gb.

Response time isn't important because it takes 30-45 seconds (on GPU - CPU is more like 7-9 minutes) to process the data. I don't want an instance running 24/7. That's why a serverless GPU service came to my mind.

inf1 instance sounds nice. thx!

[Nico Axtmann @_nicoax]

@BenediktSVogler Looks nice but it doesn't offer GPU usage.

[Nico Axtmann @_nicoax]

@iamtrask I bought one when it got colder last year. Definitely can recommend it and it is an incredible life hack.

[Nico Axtmann @_nicoax]

@am29d Costs - internal mvp that is used infrequently with a very big language model to process some text data.

Lambda with GPU would be great here because it provides a huge speedup and we don't need to care about idle costs and scaling of the service.

[Nico Axtmann @_nicoax]

@ameisen_strasse If you are interested in helping small to midsize companies in Europe we can definitely have a chat 🙂Tons of interesting problems from running thousands of individual models to processing millions of conversations a day and more.

[Nico Axtmann @_nicoax]

Seeing projects like @coqui_ai feels so great and exciting! I hope it will bring some democratization into SR. Still hurt by the pain from setting up #kaldi back in the days.

[Nico Axtmann @_nicoax]

@Casey Hiking in 🇨🇷