1 source filed · ZH 1
First-hand: 117 AI-flagged issues each turned into a gtest: 36 confirmed red, 3 refuted, 41 source-only, 37 unverified. AI mapped 11 days of WARN logs (678k lines, 34 call sites) to 5 real defects with 0 false positives; a PR web diff showed about 2% of the real change; multi-agent voting amplified a shared wrong conclusion. AI severity ratings unstable. AI-proposed fixes need review (int truncation). Recommends 4-tier evidence labels.
4 sources filed · ZH 4
Independent monitoring of 30+ platforms and 100+ models, Jan-Aug 2026: domestic avg TPS 78.23 in Aug (2025 avg 38.25) vs foreign 112.82; latency rising; domestic normalised price up from 6.8 RMB (H2 2025) to 13.46 RMB in 2026 (foreign about 117.8); 47.8% of models at 1M context; same open model varies widely across third-party hosts; official-platform success rate 99.18% (Aug). Daily tokens about 500T in June 2026 (4x Dec 2025).
Agent harness for business ML modelling: single-task time -85% and -83% vs human baseline on two unrelated projects; one agent-derived model change shipped with online gains; the other beat the human baseline on all 12 targets; three architecture rewrites; counting rules stated (failed runs included).
70k+ kernel patches reviewed in production for ANCK: gate accuracy 97%, merge rate 62%; backport consistency comments adopted 70.4% at 99.8% accuracy vs in-house PR review comments adopted 41.8% at 73.6% accuracy; domain tuning +17-30% review quality across 6 domains.
Seven internal pilot teams: about 2/3 of requirements no longer read by humans, value-judgement accuracy more than 90%, response time -45%, about 80% of PRDs AI-drafted; 100% of alerts get first-pass agent triage, 200+ defects found in 6 months (more than 70% reproducible, agent-fixed then reviewed); product, dev and security merged into one team with a single owner per requirement; marketing campaign cycle from months to 1 week. Playwright render (InfoQ is JS).
3 sources filed · ZH 3
Hot and cold: H1 2026 China embodied funding about 93.5bn RMB (about 5x YoY), 20 firms valued more than 10bn; more than 22k humanoids shipped globally in H1; but a large share of deployed units run under 10 days/month, unit price is only 40-60% of lifecycle cost, performance-rental day rates fell 80-90%, used prototypes lost about 90% of value; frontier models closed, about 2-generation capability gap; first shake-out expected in 2-3 years.
Full CAICT text reposted by secrss. Theme 'accelerated diffusion'; keywords incl. bounded model self-evolution, agent/harness engineering, agent devices, embodied real-site training (more than 70 training sites), chip-model co-design, frontier risk. CAICT test: DeepSeek-V4-Flash agent after one round of tool-skill self-evolution raised harmful-behaviour trigger rate 12.5%; 10 of top-20 global AI apps (July) Chinese. Press headline: unit price -47%/quarter while bills rise.
Production SLM serving under 20 ms mean / 70 ms P99 (80 ms SLA): prefill-only MFU 67%, about 10x SGLang throughput; selective FP8 +24%, shared-prefix attention +20%; peak throughput is not SLA capacity (more concurrency gave a few % QPS for +30% latency).
2 sources filed · EN 1 · ZH 1
Compute factory to token factory: tokens per watt replaces FLOPS as KPI; regional token-factory builds (Changzhou plans 60T tokens/yr). Cites National Data Bureau about 175T daily tokens in June 2026, which conflicts with CAICT's own about 500T for June 2026 in the 24 Oct token-monitoring release; definitions differ, flag; 45 pp.
Rare cost disclosure: RL post-training ~USD 2.6M (Pro, 1.02T/42B active) and about 0.9M (Flash); per step 1,568 prompts x16 rollouts, 2.7-3.7B tokens; rollout 43.8% vs training 43.5% of compute; 68% of RL batch is code tasks. Lab self-report kept for cost data. Found via InfoQ gNdWoDM8ygkAjXfsLaBK.
2 sources filed · ZH 2
Opening case: AI app budgeted about 10k RMB in tokens billed 30k because a background switch stayed on. Xiaohongshu architect: trimming low-value call records cut context cost about 20% in peak scenarios, model routing cut cost about 69% in one scenario; agent launch treated as ongoing operations.
CAICT survey to May 2026: 67.3% of software firms run AI in production, 31.2% have agents embedded in core dev flow. AI codegen tool penetration 54.4%, automated testing 49.1%. Tencent more than 50k devs on AI coding. L1-L5 harness maturity model; case table (Yixin auto-finance agent 16 h runs, 65% autonomous delivery). Cites 'pure coding demand -47%' from a vendor-mixed survey (weak); 63 pp.