cat briefings/2026-09-13.md

Daily AI briefing

One day's edition, kept at its own URL. The current briefing is published each morning.

Sun, 13 Sept 2026

GreyNoise showed agentic exploitation at machine speed against 395 real organisations, Amodei called publicly for pacing the frontier, and Nex-N2.5-Pro put a 397B agentic model under Apache 2.0.

generated · sources linked on every line

Africa & low-resource ML

  • Ghana's Ministry of Communication, Digital Technology and Innovations signed MoUs with 11 universities — including the University of Ghana, KNUST, UENR, GCTU and UCC — to embed One Million Coders certifications, among them AI Essentials and AI Professional, into degree and diploma programmes, Minister Samuel Nartey George said on 7 September. MyJoyOnline
  • At the KOAFEC ministerial meeting in Seoul that closed on 11 September, Korea and the African Development Bank signed a letter of intent to create a Korea–Africa AI Hub, and Korea named Tanzania the first recipient of its K-AI Package of AI solutions, data centres and power infrastructure, starting with an AI education institution. The Korea Times
  • An audit of the publicly released Central Kurdish speech corpus — 35 hours and three synthetic voices — found the settings file lists equipment never used, unlabelled test recordings sitting inside the training data, and a coding fault that mishandles long numbers, a caution for anyone reusing low-resource speech releases without re-verifying them. arXiv

AI for science

  • Researchers at the University of Miami's Sylvester Comprehensive Cancer Center compared over 214 million predicted protein structures rather than gene sequences and found TM184C, an uncharacterised GPCR-like protein that builds cell-to-cell bridges and regulates autophagy, with human TM184C restoring function when swapped for yeast Hfl1, published in Nature on 11 September. Phys.org
  • GraphED, from the Eastern Institute of Technology in Ningbo, represents candidate equations as directed graphs so topology and material parameters are optimised together, and recovered compact constitutive laws for alloy steels, lithium metal and filled rubbers that beat the Johnson-Cook model, published in Science Advances on 11 September. Phys.org

Free offerings

  • Nex-AGI released Nex-N2.5-Pro under Apache 2.0 — a 397B-parameter multimodal agentic model scoring 82.2 on OSWorld-Verified, 61.2 on SWE-Bench Pro and 92.2 on OmniDoc — with weights on Hugging Face and ModelScope plus hosted access through OpenRouter. new Hugging Face
  • Edge0 published Edge0-35B-A3B-preview under Apache 2.0, a 35B sparse mixture-of-experts built on Qwen3.5-MoE that streams expert weights from storage on demand to run in under 3 GiB of active memory, losing an average 3.9 points against the fp16 base while scoring 81.0 on MMLU-Pro and 90.9 on HumanEval. new Hugging Face

Building on frontier models

  • OpenAI launched GPT-Live-1 in the API on 10 September, a full-duplex speech model that listens and speaks at once and hands reasoning to a backend model, with 12 voices, WebRTC, WebSocket and SIP telephony transports, at $0.05 per minute billed per second on top of backend model and tool costs. OpenAI Developer Community
  • Anthropic added an `auto` permission policy to Claude Managed Agents on 10 September, letting the server evaluate each agent or MCP tool call and run, deny or pause it for approval, and shipped `ant beta:sessions connect` so a terminal can attach to a live session to follow it and approve or deny pending calls. Anthropic

Cost & hardware

  • Positron raised $875 million in Series C and C-1 financing at a $5 billion post-money valuation on 10 September for its Asimov inference chip, which tapes out on TSMC N3P at the end of 2026 with 288GB to 2,304GB of commodity LPDDR5X per chip, sidestepping constrained HBM and CoWoS supply. PR Newswire
  • d-Matrix said on 10 September its next-generation Raptor XPUs will adopt NVIDIA NVLink Fusion, connecting them into a single scale-up domain inside liquid-cooled MGX racks alongside Vera CPUs, ConnectX-9 SuperNICs and Spectrum-X, so they can serve disaggregated inference next to Vera Rubin NVL72 systems. NVIDIA

Evals & guardrails

  • GreyNoise documented a Russian-speaking actor running hundreds of AI agents built on OpenAI's Codex and a DeepSeek model to develop and deploy exploits for two PaperCut NG/MF flaws, compromising 440 instances at 395 organisations in 48 countries, reaching first remote code execution in under four hours and hitting 11 organisations in 26 seconds at peak. GreyNoise
  • Dario Amodei published "We Must Pace the Frontier", arguing labs should deliberately slow capability gains so alignment and third-party evaluation can catch up, warning that a misaligned agent swarm could hold a persistent internet-wide botnet within 6–12 months, and committing Anthropic unilaterally to embedded third-party evaluators. Dario Amodei

Algorithmic advances

  • LILA scores neuron importance by the Kolmogorov–Smirnov distance between singular-value distributions, pruning large language models with no calibration data at all, beating PruneNet by 1.57 points on LLaMA-2-7B at 25% sparsity and calibrated SliceGPT by up to 6.0 points, and landing within 0.48 points of SliceGPT after one LoRA epoch. arXiv
  • Replacing the dense random matrices in Kashin-decomposition quantisation with a sign-randomised discrete cosine transform cuts per-iteration cost from O(N²) to O(N log N), matches OPTQ and QuIP at 4 bits on OPT, Llama-2 and Pythia, stays numerically stable where QuIP variants fail, and yields two 2-bit factor codes per channel suited to native 2-bit hardware. arXiv