cat briefings/2026-10-13.md

Daily AI briefing

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

Tue, 13 Oct 2026

Microsoft's Maia 300 inference chip, Ghana ordering its universities to build Twi and other local-language models, and a proof that trained models can hide capabilities from audits.

generated · sources linked on every line

Africa & low-resource ML

  • Ghana's Minister for Communications, Digital Technology and Innovations, Sam Nartey George, tasked KNUST with leading a Twi language model, assigning other Ghanaian languages to other universities and pairing academic champions with traditional authorities. MyJoyOnline
  • Six African investors told TechCabal that fundable AI startups must supply what a model cannot: voice products working in languages such as Wolof and isiZulu, deep system integration, hard-won licences and revenue. TechCabal
  • A cross-modal attention network trained on 2,560 mammograms from 1,024 patients at four Ethiopian referral hospitals reached 97.8% accuracy and 0.987 AUC, losing only 3.2% on low-quality images. arXiv
  • RunyaNER, the first public named-entity benchmark for the East African language Runyankore, releases 237,000 manually verified words across 30,000 sentences and tests which auxiliary languages transfer best. arXiv

AI for science

  • Prompteina generates all-atom protein structures from natural-language descriptions, trained on 1.8 million structures with over 50 million captions, and nine designed binders against GM2A were confirmed experimentally, the strongest at 107 nM. arXiv
  • Treating the choice of nanobody scaffold and epitope as an inference-time resource-allocation problem, rather than committing upfront, raised designable candidates per GPU-hour nearly fivefold without losing diversity. arXiv
  • RetroTAC predicts route-derived synthetic accessibility for PROTAC degraders at 0.13 RMSE, and its top-scoring candidates average 0.73 synthesizability against 0.31, with 98.75% versus 51.00% of routes fully resolved. arXiv

Free offerings

  • OpenBMB released MiniCPM-V 4.7 35B-A3B under Apache 2.0, the first open mixture-of-experts multimodal model to run on-device, decoding 13-plus tokens per second in 3GB on an iPhone 17 Pro Max. new Hugging Face
  • Sperid Labs released Iris-3B under Apache 2.0, a 3-billion-parameter image generator that paints pixels directly with no VAE and fine-tunes for depth estimation and upscaling, its four-step Turbo variant rendering 1024×1024 in about 0.7 seconds on one H100. new Hugging Face

Building on frontier models

  • Meta released Muse Wonder through its Model API at $4 and $20 per million input and output tokens, while the larger Muse Wonder Full goes only to Muse Code subscribers and is not served over the API. TechCrunch
  • AgentSWE commissions coding agents to create, edit and optimise other agents across 25 tasks graded on held-out cases, and no tested agent delivers what its requirement actually asked for. arXiv
  • TerminalHorizon synthesises 1,500 long-horizon terminal task environments, and fine-tuning Qwen3.5-27B on the resulting trajectories improved all 12 benchmarks covering coding agents, production agentic work and general reasoning. arXiv
  • AI agents rebuilt Call of Duty: Modern Warfare 2's source code in C++ over three months and more than 500 billion tokens, recovering 99% of functions with 83% matching the original byte for byte. Tom's Hardware

Cost & hardware

  • Microsoft announced Maia 300, its third-generation in-house inference chip, claiming 40% more tokens per megawatt than leading GPU systems and seven times the rack throughput of Maia 200, reaching production workloads in 2027. The Verge
  • Nvidia contributed its Vera Rubin rack and compute tray designs to the Open Compute Project through the MGX architecture, and published open AI-factory design and grid-resilience standards with hyperscalers. NVIDIA
  • AMD debuted Helion, a rack reference design for agentic AI servers scaling to 32,768 Zen 6 "Venice" cores, shipping in CPU-only and all-in-one configurations. Tom's Hardware

Evals & guardrails

  • Backdoors that embed encrypted neural networks inside a model's weights, decrypting only on a secret trigger, stay provably hard to elicit even with full weight access, setting a hard limit on elicitation-based capability audits. arXiv
  • DRYLAB-Bench scored 15 frontier models on 30 hazardous mutation-design tasks against 187,695 wet-lab measurements, finding 67.6% of one-shot designs beat the average measured variant and that biological design tools raise risk further. arXiv
  • Agent skills can poison later sessions through files they leave behind, succeeding 46.7–77.4% of the time across Codex CLI, Gemini CLI and Claude Code, which a provenance-restoring advisory guard cuts to 10.9%. arXiv
  • Microsoft president Brad Smith said the UK, not the US or EU, is best placed to set global AI safety rules, and the UK AI Security Institute confirmed it still receives frontier models before deployment. TechCrunch

Algorithmic advances

  • Multi-Stream Attention Residuals gives depth-wise attention multiple heads, beating standard residuals, Hyper-Connections and AttnRes on all 12 benchmarks at 18B and 28B mixture-of-experts scale with a 1.58x iso-loss compute speedup. arXiv
  • SemiCat, which adds discrete reveals to a two-level interpolant, lifts continuous diffusion language models to 49.94% on TinyGSM and beats discrete diffusion on four of six likelihood evaluations at 1.7B parameters. arXiv
  • ReplaySSM caches a state-space model's recent inputs instead of its state, accelerating decoding 1.20–1.47x over tuned vLLM at batch size 256 and roughly 1.8x speculative-decoding throughput on 120B-class hybrids. arXiv
  • Uncertainty-guided progressive contraction matches or beats full-data supervised fine-tuning on reasoning benchmarks using as little as 0.22% of its optimisation budget. arXiv