Revolut moved ~2,000 engineers onto a self-hosted GLM for coding and placed Claude Code and Codex on an exception list. The deployment runs on 56 Nvidia B200 GPUs with a blended cost of 30 cents per million tokens, about one-tenth of Anthropic’s price.
Key points
Engineers: ~2,000 engineers migrated to self-hosted GLM for coding.
Infrastructure: 56 Nvidia B200 GPUs, blended 30 cents per million tokens, roughly a tenth of Anthropic cost.
Sources
libghostty 1.4 speeds up IO up to 16x in benchmarks
libghostty 1.4 speeds up IO up to 16x in benchmarks, with ASCII 1.5x and cursor/erase 16x gains. The fast-path now handles mixed width characters and hyperlinks. SIMD detects pure-ASCII and processes directly from the read buffer without a copy.
Key points
Fast-path now supports mixed width characters and hyperlinks.
SIMD fast path detects pure-ASCII and avoids the UTF-8 decode copy.
Sources
P-H3-Image open-source release 5× faster 8-step image gen
P-H3-Image is an open-weights image generation and editing model built from MiniMax-H3, generating 4 MP images 5× faster in 8 steps. The model is 7× smaller than H3 and runs at 8 MP using about 21GB VRAM.
Key points
Size: 20GB download versus 143GB for H3 (7× smaller).
Speed: 4 MP image in 2.6 s vs 12.4 s for H3 (5× faster, 8 steps instead of 50).
Sources
Elon Musk AI chips vs power wall by end of 2026
Elon Musk says the AI boom will hit a wall by the end of 2026 because electricity supply will lag chip production. He claims more AI chips will be made than can be powered, making raw electricity the bottleneck.
Key points
AI chip output is exploding while electricity growth is nearly flat outside China.
By the end of the year, more AI chips will be produced than can be turned on.
Sources
Claude Code systematic debugging skill added
Claude Code systematic debugging skill forces investigation of root cause, testing hypotheses, and verification of fixes. It is installed with a single npm command.
Key points
Prevents Claude from jumping to conclusions when debugging.
Install: npx claude-code-templates@latest --skill development/systematic-debugging
Sources
MiniMax-H3 converted to world-class image model
MiniMax-H3 video model was turned into a world-class image model, with all work open and detailed in a technical blog post.
Key points
Process: model brain surgery was performed on the MiniMax-H3 video model.
Result: the output is a world-class image model and everything is open, with a technical blog post describing it.
Sources
- Original post
- pruna.ai - Linked in the post
Sable verifiable inference system for Dolly models
Sable is a verifiable inference system that will be deployed for Dolly models, providing cryptographic proofs of committed weights, quantization, sampled randomness and more. The announcement includes a link to code and a paper.
Key points
Sable is a verifiable inference system.
Deployment will include cryptographic proofs of committed weights, quantization, sampled randomness for Dolly models.
Sources
open-source Palantir for Cybersecurity tracks breaches and CVEs
open-source Palantir for Cybersecurity tracks live data breaches, exploited CVEs (including PHP/Linux) and malicious packages, and indicates if your stack is affected. It offers a one-click coding agent to fix identified issues.
Key points
Tracks live data breaches.
Tracks exploited CVEs, including PHP/Linux.
