Show HN: SOS – Project state between coding agent sessions
Built after watching Codex repeatedly lose track of project status on long-running work, forcing the author to re-explain completed steps and open decisions each session.
AI Daily Recap
15 articles · 6 categories
The finishable daily brief
What happened in AI — Sep 11, 2026
Friday, Sep 11, 2026
15 articles · 6 categories
read top to bottom · then stop
In 30 seconds
- Anthropic accused Chinese AI labs of illicit distillation and said they routed requests to Claude at least 35 million times this summer, alongside separate US agency accusations against six Chinese AI firms.
- DeepSeek shipped a smaller, faster model and cut pricing further, sending renewed volatility through Samsung and SK Hynix memory-chip shares.
- Moonshot AI and Z.AI both moved to lock in capital — a $2B revenue run-rate target on open-weight Kimi K3, and a $5B Hong Kong share/bond sale — as Chinese open-weight labs race to fund frontier compute.
- Two Show HN projects (Viaduct, SOS) both target the same gap: coding agents losing architectural and project-state context across sessions.
- AWS and InfoQ both published on agent observability today — AgentCore's dual-layer eval/monitoring approach and session traces plus cost controls for diagnosing runaway agent spend.
Anthropic and US agencies escalated scrutiny of Chinese AI labs today, accusing several of illicit distillation and revealing that Chinese labs routed requests to Claude at least 35 million times over the summer. DeepSeek countered with a smaller, faster model whose pricing pressure is already rattling South Korean memory-chip stocks.
On the builder side, two Show HN launches tackled the same pain point from different angles: giving coding agents durable state and change-set structure across sessions, while AWS and InfoQ pushed agent observability toward standardized traces and cost controls.
Coding Agents Get Cross-Session Memory Tooling 2 items
Two independent Show HN launches converge on the same fix: coding agents keep losing track of project state and architectural decisions between sessions, so builders are shipping external memory layers instead of waiting on model context windows to solve it.
Show HN: Viaduct – Architecture change sets for AI coding agents
Packages architectural decisions into discrete change sets so a coding agent can apply, review, and reason about structural edits instead of freeform diffs.
Agent Observability Converges on Traces and Cost Controls 2 items
Two vendors published overlapping playbooks today for the same problem: multi-agent failures don't show up in traditional monitoring, so teams are standardizing on execution traces plus cost/spend guardrails to catch tool-call loops before they burn budget.
Monitoring production agent lifecycle with AWS DevOps Agent and AgentCore Evaluations
Proposes a dual-layer approach — Amazon Bedrock AgentCore Evaluations for continuous quality scoring, paired with AWS DevOps Agent — because multi-agent systems fail in ways traditional monitoring misses.
Session Traces and Cost Controls Help Diagnose AI Agent Failures
Session traces paired with cost controls are emerging as the standard way to spot tool-call loops and runaway spend while keeping enough context for post-incident review.
Inference and Training Infrastructure at Scale 3 items
Today's infra writeups span the two ends of the deployment spectrum: OpenAI's storage layer holding up 1 billion ChatGPT users and 22M requests/second, LinkedIn cutting training cost with multi-teacher distillation, and NVIDIA pushing inference back down to a user's own local hardware.
Rapidly scaling online storage to serve over 1 billion ChatGPT users
OpenAI details how its Habitat storage platform grew from a Python library into a globally distributed system now serving 1 billion ChatGPT users and 22M requests per second.
How LinkedIn Trains AI Job Search 8x Faster with Multi-Teacher Distillation
LinkedIn compresses knowledge from multiple large teacher models into a compact 0.6B-parameter student to cut its AI job-search training time by 8x.
NVIDIA Personal AI Router Distributes AI Tasks Across Local Compute
NVIDIA's beta Personal AI Router (PAIR) pools inference capacity across multiple machines on a local network and auto-routes AI requests among them.
DeepSeek's Efficiency Push Rattles Chip Demand 3 items
DeepSeek shipped a smaller, faster model alongside the V4.1-Flash release that's pushing API pricing lower, and the efficiency gain is specific enough that South Korean memory-chip stocks moved on it — a live signal that model efficiency is now a direct input to hardware demand forecasts.
DeepSeek launches smaller but faster AI model
DeepSeek released a new, smaller model tuned for faster inference, continuing its pattern of trading parameter count for latency and cost.
DeepSeek V4.1-Flash puts pressure on AI pricing
DeepSeek V4.1-Flash's pricing is compressing margins across the AI API market, continuing the lab's pattern of undercutting incumbent providers on cost.
DeepSeek efficiency gain sparks new volatility for South Korean chipmakers Samsung, SK Hynix
The efficiency gain behind DeepSeek's latest release is specific enough that investors moved South Korean memory-chip stocks on it, tying model architecture directly to hardware demand expectations.
Distillation and Provenance Scrutiny Escalates 3 items
Western labs and US agencies moved from suspicion to specific accusation today: Anthropic says Chinese labs both distilled its models illicitly and routed tens of millions of requests to Claude, while separate US agency action names six Chinese AI firms directly.
Anthropic accuses Chinese AI labs of illicit distillation attacks
Anthropic says Chinese AI labs used illicit distillation techniques to copy capabilities out of its models.
China's star AI labs routed user requests to Claude at least 35 million times in the summer: Anthropic
Anthropic says Chinese AI labs routed user requests through to Claude at least 35 million times over the summer, a scale that complicates any clean separation between their models and Anthropic's.
US agencies accuse six Chinese AI firms
US agencies leveled accusations against six named Chinese AI firms, adding regulatory pressure alongside Anthropic's distillation claims.
Chinese Open-Weight Labs Chase Capital to Fund Compute 2 items
Moonshot AI and Z.AI both moved on funding today — a $2B annual revenue run-rate target built on the open-weight Kimi K3, and a $5B Hong Kong share and convertible-bond sale — signaling that sustaining open-weight releases at the frontier now requires public-market-scale capital.
Moonshot AI eyes $2B annualized revenue as Kimi K3 lifts sales
Moonshot AI is targeting a $2 billion annual revenue run rate, with sales lifted by its open-weight Kimi K3 model.
China AI developer Z.AI launches $5 billion Hong Kong share, convertible bond sales, term sheet shows
Z.AI is raising $5 billion through a combined Hong Kong share offering and convertible bond sale.
You are caught up for this edition