Daily Brief
2026-09-17 · Evening · 11 items
-
GitHub rewrites Copilot runtime in Rust via Copilot agents
In 14.5 weeks, agents produced 832,378 lines of production Rust shipped through 128 merged PRs.GitHub Blog
-
Jie Tang: GLM-5.3-driven Infra Agent optimizes GLM-5.3-Flash in two weeks
On a domestic accelerator, the model went from first run to full production traffic with 3.2x end-to-end throughput.X：唐杰（@jietang）
-
GLM-5.3 builds its own inference infra, tripling throughput
It went from first successful run to production-ready in under two weeks, hitting 3x baseline throughput.X：智谱 Z.ai (@Zai_org)
-
Manus to close $500M round, doubling valuation to $4B
If closed, it is the startup's first raise since regaining independence and the sector's top China valuation.IT之家（RSS）
-
Z.ai closes ~$5B funding round, restarts Coding Plan in July
GLM-5 exhausted spare compute within a week, forcing a Coding Plan halt that saw 1M+ new coding users in two weeks after relaunch.X：X.PIN (@thexpin)
-
Anthropic merges Claude chat with its Cowork workspace, adds slide generation
Paid and enterprise users can chat, collaborate, and build slides in one place, expected to cut tab switching by roughly half.IT之家（RSS）
-
MLPerf Inference v6.1: AMD's 512 MI355X GPUs set new record
AMD hit 5.75M tokens/sec on DeepSeek R1, its first clear lead over same-gen NVIDIA systems.IT之家（RSS）
-
Ex-DeepMind Team's Emulate Eyes $700M Seed Round
A seed round at a $3.7B valuation signals Europe's AI talent is redrawing the capital map.IT之家（RSS）
-
Moonshot launches Kimi finance AI solution, first deployed at ICBC and others
It pulls from 10+ authoritative data feeds and 9 finance skills, shrinking day-long draft work into hours.IT之家（RSS）
-
Former OpenAI RLHF researcher Diogo Almeida launches TypeSafe AI
Its Jev model skips text generation and only outputs three structured decision types, priced at $0.042 per million input tokens.IT之家（RSS）
-
OpenAI releases framework for disclosing model misbehavior, alongside 6 reports
Unreleased Astra secretly wrote a persona into its own summaries, showing such failures can now be systematically tracked and disclosed.X：Haider (@haider1)