Alexey Demyanchuk

pragmatic minimalist

[Alexey Demyanchuk @aDemyan4uk]

@thdxr OpenCode Ride

[Alexey Demyanchuk @aDemyan4uk]

@thdxr thanks, it's so clear and to the point, now I can see where the clarity and conciseness of opencode is coming from

[Alexey Demyanchuk @aDemyan4uk]

Huge thanks to the devs behind https://t.co/vtsnwII4Hv for making Triton work on Windows 🙌

Thanks to their work, we could use torch.compile with a Triton backend — even though official Triton doesn’t support Windows yet.

Open-source wins again! 🔧🔥

[Alexey Demyanchuk @aDemyan4uk]

Been digging into tokenization while following @karpathy Zero-to-Hero path.

Wrote my own BPE tokenizer from scratch — added regex-based grouping, chunked encoding (like GPT-2), and learned a ton along the way.

If you're curious:
🔗 https://t.co/MPLZiLNT1t

[Alexey Demyanchuk @aDemyan4uk]

@karpathy More educational content is coming :)

[Alexey Demyanchuk @aDemyan4uk]

@AggrendWoW I would delay! Let non-try-hard people level, I guess a week will be enough even for the busiest of us.

[Alexey Demyanchuk @aDemyan4uk]

@karpathy Waiting for announcement of Andrej joining @huggingface :)

[Alexey Grigorev @Al_Grigor]

I have 3 tickets to the MLOps World conference

Either in person in Toronto or online

Want to win one?

🔸️ Follow me
🔸️ Retweet this tweet

👉 https://t.co/XyOikVMmWR

Winners selected randomly and announced on Wednesday

Good luck 🍀

[Alexey Demyanchuk @aDemyan4uk]

Learn MLOps in a free online course from @DataTalksClub!

- Processes
- Training models
- Serving models
- Monitoring
- Best practices

More info here: https://t.co/Y68jLzbI9N

[Alexey Demyanchuk @aDemyan4uk]

@karpathy Great plan! You're probably already overwhelmed with invitations. But in case you are open to meeting and making new friends, it would be super cool to meet in Germany for beer or tea :)

[Alexey Demyanchuk @aDemyan4uk]

@karpathy Mama

[Alexey Demyanchuk @aDemyan4uk]

@seb_ruder Wish you all the best with the new job and protect :)

[Alexey Demyanchuk @aDemyan4uk]

@borisdayma Hard negative mining? Mix-up augmentation? Misalign x and y?

[Alexey Demyanchuk @aDemyan4uk]

@Bosch_AI Hi, great opportunities :) Just to let you know, the second link is broken and leads to 404

[Alexey Demyanchuk @aDemyan4uk]

@yoavgo One more addition. Here https://t.co/wUrlaiiade is their network bandwidth explanations. I think, one should be able to achieve a better performance than 20 minutes for 200 GB. One more point for trying support :)

[Alexey Demyanchuk @aDemyan4uk]

@yoavgo Even though they have quite extensive docs, it is not that obvious from the first glance to me what would be the bandwidth between gcp and vm. You may try to reach they support, they might have some better options to share.

[Alexey Demyanchuk @aDemyan4uk]

@yoavgo Sorry for posting the same message twice, thought I didn't send the first one :)

[Alexey Demyanchuk @aDemyan4uk]

@yoavgo I think, if the data is somehow transformed on the regular basis, it would be better to save it on persistent (mountable to VM) disc and then do all the transformations.

[Alexey Demyanchuk @aDemyan4uk]

@yoavgo This site https://t.co/mz8gm33LTk
has about the same time estimation