今日技术情报 · 2026-11-09Daily Tech Intel · 2026-11-09本日の技術インテリジェンス · 2026-11-09

📌 Today’s Top Stories

1. Goldman Sachs pulls Claude Code back from traders and bankers, HN 669👍
This is the first case of a major financial institution publicly withdrawing a coding agent, and the reason isn’t model capability but compliance and audit boundaries—code changes in regulated roles require a traceable responsible party, and diffs generated by agents cannot satisfy this. For teams pushing agents in regulated industries like finance and healthcare, this shifts the question from “can an agent write production code” to “who signs off on code written by an agent.” What to do this week is fold agent output into your existing change approval flow rather than opening a separate channel. · The State of AI

2. Alibaba Co-Design Bench: GPT-6 Astra fully solves only 26.4% of real user requests, versus 99%+ on old benchmarks for the same model
The benchmark uses “real requests that actually stumped agents,” not synthetic tasks. The gap between 26.4% and 99% shows that existing SWE-bench-style evaluations have lost their discriminative power, and any selection decision based on old benchmark scores needs to be re-run. For teams evaluating options like Ripple-65B and mini-swe-agent, this is the only number worth looking at this week—first confirm how many requests in your eval set are “requests that actually failed before.” · The Decoder

3. Karpathy uses Claude Opus 5.5 over a weekend to filter 43 million HN comments down to 6,454 videos, says results beat YouTube recommendations
A single long-context batch process plus structured filtering let one person accomplish in two days what used to take a recommendation systems team months. The reproducible part is the “using LLMs for offline content curation” pipeline: the 43M→6.4K compression ratio means roughly 0.15 outputs per thousand comments, and the cost is calculable. For teams working on content retrieval or knowledge bases, what’s worth copying is the structure of its filtering prompt, not the results. · The Decoder

4. llama-modchip: 284B DeepSeek runs on 64GB of RAM with no GPU, HN 388👍
Squeezing a 284B model into 64GB of consumer-grade memory relies on weight chunking plus on-demand paging from disk, at the cost of throughput. The real value here isn’t “it runs” but that it provides a reproducible offload implementation, lowering the VRAM barrier for local deployment from “multiple GPUs” to “a single machine with lots of RAM.” For teams evaluating private deployment, test its tokens/s first before deciding whether it’s worth replacing API calls. · GitHub

5. Nvidia board investigates Jensen Huang’s relationship with suspected Chinese spy Alice Huang; Huang admits to extramarital affair
A security incident at the CEO level, with a Megaspeed executive alleged to have been recruited by China’s Ministry of State Security to conduct intelligence activities against Nvidia. For teams procuring Nvidia compute or with supply-chain ties to the company, the short-term impact is that export controls and compliance reviews may tighten; the long-term view is a political risk premium on the AI chip supply chain. · TechCrunch / The Verge

⚡ Trend Signals

The evaluation standard for coding agents is shifting from “can it do it” to “who’s responsible once it’s done”: Goldman pulling back Claude Code and Co-Design Bench’s 26.4% are two sides of the same thing—capability is already sufficient; the bottleneck is accountability and real-task success rates. When selecting tools, putting “audit trail” and “pass rate on real failure sets” into your evaluation table is more useful than looking at SWE-bench scores.

Continuing from 11-05 OpenAI’s acquisition of Pierre (a permission model for agents pushing directly to main) and 11-08 apiiro’s credential leak scanning—all three point to the governance gap once agents enter production. Not a new signal, but today brings the first withdrawal case in the financial industry.

The memory barrier for local inference is dropping; the VRAM barrier hasn’t changed: llama-modchip’s 284B/64GB and ISTA-DASLab’s DeepSeek-V4.1-Flash Coder GGUF appearing at the same time shows the community is pushing toward “big-memory single machine + quantization.” The cost model for private deployment should be recalculated based on “memory bandwidth” rather than “VRAM capacity.”

Continuing from 11-07 Ripple-65B’s Apache 2.0 full-weight release—the supply of privatizable coding models is increasing, and the bottleneck is shifting from “are there weights” to “can you actually run them.”

🛠️ This Week’s Actions

• Run your current coding agent against Co-Design Bench’s public task set, measuring “full solve rate” rather than “partial pass rate,” to verify how large the gap is between old benchmark scores and real-task success rates
• Grep through agent-generated config files in CI (continuing from 11-08’s credential leak), confirming no plaintext tokens have entered committable paths

📂 Source Roundup

📰 Tech Media

• Epic Games Launcher is coming to SteamOS The Verge: Fortnite is compatible with Linux handhelds, and SteamOS’s supported-distribution list now formally includes Epic—channel expansion for teams working on Linux game compatibility layers or Proton-related work.
• Sutton reiterates that raw compute is insufficient to produce intelligence; former OpenAI researcher says blind scaling “reward hacked a generation of scientists’ brains” The Decoder: For teams betting their entire training budget on scale, this is another argument for “invest in new learning methods,” but it offers no actionable alternative.
• Elvis, Eminem rights holders sue Suno The Verge: Copyright risk in AI music generation is extending from “training data” to “outputs imitating specific artists”; teams doing audio generation should check output-side style filtering.
• Apple accelerates making Siri more agentic The Verge: Ternus pressures Federighi to match OpenAI Dots and Meta Muse; iOS 27.3 beta already shows a “Conversational Apps” flag, and the entry point for third-party chatbots into HomeView is opening up.

🔥 GitHub Trending

• niubigeo ⭐+228: Open-source AI brand visibility and competitor reporting tool in the GEO (generative engine optimization) space, with the paid portion relying on human testing; the core analysis logic is closed-source, and what’s open is just a shell.
• Agentero ⭐+145: A research workbench for human-agent collaboration, with cited Q&A over PDFs + Vault + writing, filling Zotero’s human-machine context gap; no benchmark provided for citation accuracy.
• google-colab-cli ⭐+104: Official CLI for operating Colab runtimes, moving notebook execution from the browser to the terminal, suitable for scripted experiments; depends on Colab quotas.
• mini-swe-agent ⭐+86: A 100-line GitHub issue-solving agent with SWE-bench verified >74% and no giant config; compared to the full SWE-agent it lacks the tool orchestration layer, so its ceiling on complex tasks is lower.
• mujoco ⭐+41: DeepMind’s physics simulator keeps appearing on the list, a standard component for robotics/RL training; no new version information.

🤗 HuggingFace Trending Models

• Scantist/GLM-5.3-SLIM-GGUF text-generation: ❤️1292, the SLIM-quantized version of GLM-5.3; downloads are only 4k/month but likes are high, indicating the community is waiting for a locally deployable version at this size; suitable for replacing a local general-purpose chat model.
• ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF image-text-to-text: ❤️206, 258k/month downloads, the coder-quantized version of DeepSeek-V4.1-Flash, runnable locally with llama-modchip-style offload; no data provided on code capability degradation after quantization.
• PrunaAI/P-H3-Image-Comfy text-to-image: ❤️254, 625k/month downloads, Pruna-optimized ComfyUI version of the H3 image model, focused on inference acceleration; suitable for replacing the original H3 node in ComfyUI.

🧠 AI/ML Papers

• Exploring Weight-Space Programming: Capabilities, Generalization, and Limitations: Systematically reproduces the WSP reported on 11-03, with 16 editing trajectories, 7 benchmarks, and two tiers (Qwen3.5-0.8B/9B); the conclusion is that gains come from highly localized interventions—providing the first citable boundary condition for “should we downgrade the LoRA pipeline.”
• AgentBaton: Efficient Model Collaboration in Agentic Tasks: Puts strong and cheap models into collaboration within the same long-horizon agent trajectory, rather than query-level routing; for cost-sensitive agent deployments, this is a finer-grained cost-saving path than “use strong models for everything.”
• Confucius4-R2T2: A Stable-Prefix Streaming ASR Model: Post-trained on Qwen3-ASR-1.7B with a frozen audio encoder, ensuring already-emitted tokens are never modified—for voice agents and real-time translation, this eliminates the problem of “tokens being revised causing downstream action errors.”

💬 Hacker News

• Goodbye Claude, hello Pi 👍716: The author’s hands-on record of migrating from Claude to Pi; 444 comments debate “the lock-in cost of coding agents,” with the core conclusion that migration cost lies mainly in the prompt and tool adaptation layer rather than the model itself.
• Declarative Dragging 👍419: Makes drag interactions declarative; 53 comments discuss the trade-offs versus existing imperative drag libraries—a concrete pattern for the frontend interaction layer.
• KetchUp – an open-source SketchUp alternative in Rust + WebGPU 👍307: 3D modeling running in the browser with a WebGPU backend, targeting SketchUp’s local-install model; no performance data for complex models.
• Outlook.com begins silently rejecting email from domains without DMARC 👍304: The hard threshold for sending infrastructure rises another notch; any team running its own email sending should check today whether there’s a DMARC record in DNS.

🚀 Product Hunt

• Sage2 by Levanto Labs: An AI research/writing assistant whose differentiation is long-document context management; specific technical details undisclosed.
• Zigzag Browser: A browser product; differences from existing options not explained, insufficient information.