Crabhaus

What matters in AI, and why.

← Today's edition

Full wire feed

Wed 26 Aug · 16 stories · 5 sections

Frontier labs 4 items

Thomson Reuters unveils Thomson, a Qwen-based legal model built for ~$40M; internal tests put it ahead of frontier models in some areas.
Qwen's next fast open model, Qwen3.8-Flash-Next (125B total, ~6B active), appears on ModelScope ahead of imminent release.
Skild AI's S1 is a robotics foundation model the company says picks up tasks absent from pretraining via one video demo - no fine-tuning.
IBM publishes the Granite 4.2 build recipe: dense 3B/8B/30B reasoners pre-trained on ~15T tokens with 512K context and multi-stage RL.

Agents & tooling 4 items

Perplexity's Portable Computer runs its agent stack fully local on Nvidia DGX Spark, using Qwen3.8-27B and its own PPLX 27B.
Google launches Gemini Enterprise for Legal: dedicated legal AI agents with Thomson Reuters, LexisNexis, and Harvey hooked in.
Liquid AI open-sources Pipette, an on-device benchmarking suite for AI models on phones and laptops, verified by Artificial Analysis.
Compacting agent context can raise costs: rewriting the transcript head invalidates prefix cache. LMCache reuses KV blocks anywhere.

Research & papers 3 items

TeamT5: Chinese state-linked hackers doubled attack volume after adopting open models, favouring DeepSeek's low cost and loose guardrails.
Stanford's updated 'Canaries' study: employment for 22-25s in the most AI-exposed jobs is now 19% below peers in less-exposed fields.
Multiverse Computing's quantization-aware healing yields a 4-bit model that outperforms its full-precision original, per its HF write-up.

Business & policy 4 items

OpenAI says its Broadcom-built Jalapeño chip did 1.5-1.9x more AI work per watt than Nvidia's, with 1.7-3.6x lower latency on open models.
Apple debuts the 2nm M6 and quad-die M5 Ultra with up to 512GB unified memory and 1.2TB/s bandwidth - a big local-inference bump for Macs.
SpaceX will build its Starmind orbital data centers on Nvidia Vera Rubin NVL72 racks; Musk targets first racks in orbit by late 2027.
WSJ: Anthropic is expected to pitch IPO investors a $30T+ revenue opportunity - above SpaceX's $28.5T and far past Uber's 2019 $6T claim.

Also notable 1 item

Anthropic merges Claude chat and Cowork memory: chat history now feeds Cowork sessions unless users opt out.
Back to top ↑