cat briefings/2026-10-12.md

Daily AI briefing

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

Mon, 12 Oct 2026

Ghana's central bank is drafting AI rules for lenders, Vambo AI announced a 1.5B open-weight model for 12 African languages, and frontier models cleared Booz Allen's eight control-system objectives.

generated · sources linked on every line

Africa & low-resource ML

  • The Bank of Ghana is drafting a directive governing how banks use AI, and Governor Johnson Pandit Asiama told the Ghana Association of Banks' 43rd annual general meeting that the central bank will run thematic reviews of how it is implemented. MyJoyOnline
  • South African startup Vambo AI announced MORENA, an open-weight 1.5-billion-parameter model trained from scratch for 12 African languages plus English and French, but has published no licence terms, download location or benchmark results. iAfrica
  • A CC-BY-NC dataset machine-translates LLaVA-Instruct-150K's 157,712 multi-turn image conversations into 13 African languages, among them Hausa, Swahili, isiZulu, Amharic and Nigerian Pidgin, giving 2.2 million rows for visual instruction tuning. new Hugging Face
  • A fifteen-author study of "malicious AI swarms" finds no verified case yet of fully autonomous AI influence operations, and argues the shortage of African-language training data constrains attackers as much as defenders, contrasting Mali with Ethiopia. arXiv

AI for science

  • MORAD trains one reinforcement-learning policy across diverse RNA backbones against six objectives at once, lifting GDT-TS 16.0% to 0.3827 and base-pairing MCC 19.4% to 0.7295 while cutting normalised ensemble defect 27.5%. arXiv
  • GEOM-DRUGS, the standard reference set for molecular conformer generation, is missing low-energy conformers on its own energy surface, as 2.3–7.0% of structures from five public generators fall below its reference thresholds. arXiv
  • Machine-learning force fields reproduce molecular-crystal vibrational frequencies to about 3 cm⁻¹ and relative harmonic free energies to about 1 kJ/mol at 300 K against density functional theory. arXiv

Free offerings

  • Cactus Compute released Whistle under Apache 2.0, a speech-to-text model in a single 16.9 MB file that runs on CPU with no GPU and no dependencies, handling seven languages with per-word timestamps and speech embeddings. new Hugging Face
  • Underdog Saluki 27B 1.0, released under Apache 2.0, packs Qwen3.8-27B into a 7.89 GB two-bit GGUF for stock llama.cpp, scoring 88 of 120 on tool calling against 84 for the full 54 GB model. new Hugging Face

Building on frontier models

  • SoftwareGym synthesises 182 runnable GUI applications across twelve categories at a 91.5% mean task pass rate, and training a computer-use agent on their trajectories lifts out-of-domain OSWorld Pass@3 from 41.39% to 56.97%. arXiv
  • Work2Skill records synchronised human desktop demonstrations and induces reusable agent skills from them, adding 4.67–10.54 points on MyPCBench Bounded Action and 3.11–7.50 points on 286 OSWorld-Verified tasks, with code and data released. arXiv
  • ProgramGym builds 1,648 verifiable training environments in C, C++, Go and Rust from real repositories, for agents that must discover requirements and build whole programs rather than patch an existing one. arXiv
  • ReHarness revises the reflections an agent retrieved for a task after seeing how that task went, propagating feedback backwards through its text memory, and beats memory, reflection and prompt-optimisation baselines on three continual-learning benchmarks. arXiv

Cost & hardware

  • Microsoft launched Microsoft-Decision-1, a structured-output decision model built on Alibaba's open-weight Qwen3.5-9B, at $0.042 per million input tokens with output free, over 20 times cheaper than GPT-6 Sol for text classification. The Register
  • Samsung's 12-layer HBM4E memory has reportedly cleared customer quality validation with Nvidia and major hyperscale data-centre operators. DigiTimes

Evals & guardrails

  • Two unnamed frontier models achieved all eight attack objectives in Booz Allen Hamilton's industrial-control testbed, one reaching inside the control network from the perimeter in just over 16 minutes. The Register
  • A second vulnerability found by Anthropic's Mythos model is being exploited in the wild: a critical authentication bypass in the Rejetto HTTP File Server, where Mythos reverse-engineered a random number generator that was not as random as intended. The Register
  • SWE-Duel has coding agents write each other's tasks at evaluation time so nothing can leak, and tasks authored by a family's newest model beat a fixed solver 42.9% of the time against 4.0% for older versions. arXiv
  • AMLworld-Compact cuts money-laundering evaluation targets by 271 and 611 times using Horvitz–Thompson weighting, holding F1 error to 0.00–2.58 points against 4.44–7.41 for uniform sampling. arXiv

Algorithmic advances

  • DiffusionART aggregates each prompt's rewards into one shared improving direction before updating, matching DiffusionNFT and Flow-GRPO-Fast on HunyuanVideo-13B with 12.02 and 11.72 times fewer GPU hours and 28.8% and 53.0% higher mean training reward. arXiv
  • S-Space finds a linear subspace in multimodal language models encoding object positions along horizontal, vertical and distance axes, readable even on non-spatial questions, and shows the bottleneck is transforming those representations rather than forming them. arXiv
  • Picking which target-model layers feed a speculative decoder by matrix entropy and centred kernel alignment, with low-entropy tokens filtered out, beats EAGLE3's default [2, N//2, N-3] rule. arXiv