ivo's home

This is the MaxText configuration for reproducing OLMo 3 7B pre-training on Google Cloud TPUs. It covers stage 1 and stage 2, matching AI2's reference run on 5.93 trillion tokens. The setup uses JAX/XLA and a specific data pipeline to replicate the PyTorch training curve.

It is valuable if you want to verify that a PyTorch recipe translates faithfully to TPU hardware. The catch is that it requires access to Google Cloud TPUs and the specific OLMo 3 data mix. It is not a generic GPU tutorial, but a concrete validation of MaxText's numerics against a known baseline.

more: https://developers.googleblog.com/reproducing-olmo-3-7b-pre-training-in-maxtext-case-study-of-large-scale-training-on-tpus/

my other side: ivo.blogspot.com

This guide walks through migrating an SO-101 robot arm simulation from standard CPU-based MuJoCo to MJWarp, which runs on NVIDIA GPUs. It batches 2,048 parallel environments using NVIDIA Warp kernels, allowing large-scale simulation steps without moving data off the device.

It is useful if you are scaling reinforcement learning workflows and need raw simulation throughput rather than single-step precision. The migration requires writing custom CUDA-centric logic and does not cover policy training, so it is strictly an infrastructure setup task for developers comfortable with GPU programming.

more: https://huggingface.co/blog/nvidia/how-to-use-nvidia-warp-and-mjwarp

my other side: ivo.blogspot.com

This npm package installs a skill for Codex or Claude Code that takes your local app live. It handles hosting, database, auth, domain, email, and payments on your own accounts. The current alpha covers Vercel, Netlify, Supabase, Neon, Porkbun, GoDaddy, Resend, and Stripe test mode.

It is worth trying if you use a coding agent that supports skills and want to automate the post-build setup. The catch is that it is an early alpha, so you must verify the results yourself. It is not a hosted service or a replacement for manual configuration of unsupported providers.

more: https://github.com/mikehasa/golive-skill

my other side: ivo.blogspot.com

Foremerge is a pre-1.0 CLI and MCP server that adds an intent layer to Git. It lets multiple agents declare semantic changes before editing, using a local SQLite store to detect conflicts like replacing a class while another extends it. The setup involves installing the tool and wiring it into clients like Claude Code or Cursor.

The deterministic conflict detection is useful for catching logical overlaps that standard Git merges miss. However, the schema is still unstable, there are no published benchmarks, and it does not support coordination across different machines. It is not for teams needing distributed state or hard file locks.

more: https://github.com/naw103/foremerge

my other side: ivo.blogspot.com