Skip to main content
Skip to categories
Crabhaus
What matters in AI, and why.
← Today's edition
Full wire feed
Tue 25 Aug · 14 stories · 6 sections
Frontier labs
1 item
Z.ai's GLM-5.3 reuses GLM-5.2's 743B base, gaining via scaled post-training alone; DeepSWE jumps 46.2 to 66.9. Weights due in ~2 weeks.
↗
Agents & tooling
3 items
Docs: Meta plans to launch Hatch, its OpenClaw-style agent platform, in coming weeks; new model Watermelon lands in October.
↗
ByteDance is folding coding platform Trae and agent-builder Coze into Doubao, plus a Doubao Work tier to rival Tencent's WorkBuddy.
↗
Gradio adds gr.Workflow, turning multi-step AI pipelines into an inspectable visual graph you can run and deploy like any Gradio app.
↗
Research & papers
3 items
Google's Recirculation loops deep-layer representations back to earlier layers at inference, weights frozen; perplexity -23%, GSM8K +21%.
↗
METR finds AI's acceleration of science is lumpy: dramatic in cyber vulnerability discovery, modest in math, not yet measurable in AI R&D.
↗
SPADE has an LLM alternate between writing executable training environments and solving them; Qwen3-30B gains +8.1 over base on games.
↗
Business & policy
4 items
Nvidia's inference chip Groq 3 LPX enters full production claiming 3,431 tok/s on Gemma 4 31B at 100K context; Nebius is first customer.
↗
Hugging Face, host of 2M+ models, is exploring a sale at $13B or more - nearly 3x its 2023 valuation, Business Insider reports.
↗
NYT: Russia used Nvidia Jetson modules in fully autonomous AI drones; Ukraine says one chose its target in a July strike that killed three.
↗
The Information: Nvidia is in talks to invest in Perplexity at a $30B+ valuation, after weighing a tech-licensing deal and hiring staff.
↗
Tutorials
1 item
Daily Dose of DS's three-part hands-on guide to preloading a corpus into KV cache instead of retrieving per query, with break-even math.
↗
Also notable
2 items
Ox Alpha update: full DeepSWE runs put it at 63%, near Fable 5 at fewer tokens; its OpenAI cl100k tokenizer shifts guesses to Microsoft.
↗
OpenAI previews Private Safety Processing: cross-interaction misuse detection that stays compatible with Zero Data Retention deployments.
↗
Back to top ↑