Home | Smart Chunks

AI News

UK Backs £165m Biosecurity Plan With £200m in OpenAI and Anthropic Credits

October 24, 2026

The UK’s £165m AI biosecurity initiative leans on up to £200m in OpenAI and Anthropic model credits. Here’s what’s confirmed, and what the headline number hides.

Udgam Labs Trains 152M Chat Model From Scratch for $146

October 23, 2026

Udgam Labs trained Laghu, a 152M English chat model, from scratch for $146.26 of pretraining compute and published the code, data, failed runs and costs.

KRAFTON AI’s Raon-1 Mini Preview Puts 30B MoE on a Single RTX 5090

October 22, 2026

KRAFTON AI’s Raon-1 mini preview is a 30B-A4B MoE under Apache 2.0, with NVFP4 weights for a single RTX 5090 and a DFlash2 draft model claiming up to 2.18× throughput.

Flat illustration of nine document pages in a row, lit at both ends, with the highlighted page in the dim middle

The Contract Fit. The AI Still Cited the Wrong Clause.

October 19, 2026

AI Model Selection Framework: Which Frontier Model for Which Job

April 29, 2026

Editorial illustration on a warm cream background. Two stylised runners on a track approach a finish line side by side, one in a deep navy "PROPRIETARY" jersey and one in a sage-green "OPEN SOURCE" jersey. The track has soft horizontal lane markings. A dotted trail behind them shrinks toward the finish line, suggesting the gap they have closed. A small pile of model badges or rosettes near the foreground reads short, abstract labels suggesting "MIT" and "Apache 2.0". Editorial vector style similar to Ars Technica or The Verge feature illustrations: muted terracotta, sage green, soft navy on warm cream, soft drop shadows, paper-texture feel, no logos, no neon, no robots, no AI brain visuals. 16:9 ratio.

Open-Source vs Closed-Source AI in 2026: The Gap Has Closed (Mostly)

April 28, 2026

Guides

Flat illustration of a grid of document cards with one card pulled out and marked with a red cross

Your Docs Chatbot Cited the Handbook and Still Got It Wrong

October 23, 2026

Your AI chatbot doesn’t memorize your files. It searches them on every question and writes from whatever it found. Here is how that works, and where it breaks.