Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Sat 29 Aug · 15 stories · 6 sections

Frontier labs 4 items

Z.ai releases GLM-5.3's weights; a new license makes companies with $10B+ annual revenue pass Z.ai's security review to host the model.
Tencent previews Hy4: a 770B-param open MoE (49B active) with 1M context it says narrowly beats GLM-5.3 and Kimi K3 in internal tests.
DeepSeek-V4-Pro-0813 exits preview with big agentic-coding gains plus an MIT-licensed, model-agnostic agent harness; API prices rise.
OpenAI and Cerebras preview Ultrafast: GPT-5.6 Sol served from on-chip SRAM at up to 750 tokens/sec, ~11x its standard API throughput.

Agents & tooling 1 item

Salesforce and Anthropic launch Claudeforce, a 37-skill sales plugin for Claude; Claude also becomes the default model in Slack.

Research & papers 3 items

Anil Madhavapeddy saw exploit probes minutes after opening a fix PR: agents can weaponize the rumour of a bug, breaking embargo models.
Self-GC (Xiaohongshu) has an LLM planner fold, mask, or prune agent context: ~44% of tokens cut while keeping needed detail 85% of the time.
Jordy Zomer turns agent memory into program analysis: facts stored with dependencies, so a falsified assumption retracts its conclusions.

Business & policy 4 items

OpenAI will cut Cursor's access to its models on Nov 12, ending the contract after SpaceX's acquisition; future models are excluded.
The Information: the US is drafting a rule to close the loophole letting Chinese firms use restricted AI chips via third-country DCs.
a16z creates a $1.1B Machine Age fund focused on hardware, aiming to accelerate the physical buildout of AI.
South Korea will give citizens free access to premium AI tools via KT, SK Telecom, and Kakao, treating AI like a public utility.

Tutorials 1 item

Daily Dose of DS implements a Siamese network with contrastive loss in PyTorch end to end: paired data, training, similarity scoring.

Also notable 2 items

UCLH surgeons complete the first AI-assisted brain surgery: a live model watched the camera feed and flagged arteries and optic nerves.
Ten days with Qwen3.8 27B on a Mac Studio: timed 3.8-vs-3.6 runs, RAM needs per quant, and a measured 1-bit quantization quality drop.
Back to top ↑