cat briefings/2026-10-16.md

Daily AI briefing

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

Fri, 16 Oct 2026

Meta opened an 816-billion-parameter model's weights, NVIDIA fitted a capable agentic vision-language model into 24GB, and ELARSpeech released 30,202 hours of speech covering 675 under-served languages.

generated · sources linked on every line

Africa & low-resource ML

  • Researchers released Wolof-adapted open-weight models from 0.6B to 14B parameters, instruction-tuned, preference-aligned and merged, holding up across 63 Wolof and cross-lingual tasks and already running in internal customer-care pilots. new arXiv
  • ELARSpeech published 30,202 hours of speech across 675 under-represented languages drawn from the Endangered Languages Archive, with its ASR and TTS models and build pipeline, under a non-commercial CC BY-NC-SA 4.0 licence. new arXiv
  • Kenya's Senate ICT committee recommended the Senate drop the Artificial Intelligence Bill 2026, calling it too broad and too punitive, and asked the government to finish its national AI policy first. The Star (Kenya)
  • Microsoft is reportedly closing engineering operations at its Nairobi Africa Development Centre, opened on a $100m commitment and once staffed by more than 450 engineers; Microsoft confirms Kenyan engineering cuts but not a full closure. ITWeb Africa

AI for science

  • CellNet pretrains on 231 million single-cell profiles spanning 40 species by encoding gene identity and measured RNA expression as one shared image, improving rare-population recognition and predicting withheld coding and non-coding expression. arXiv
  • An evolutionary loop in which language models propose quantum circuits and an independent evaluator verifies them cut both Toffoli and qubit counts below the state of the art in seven of eight tasks. arXiv
  • On tropical cyclones the AI forecaster FGN beat ECMWF's operational IFS-ENS ensemble on probabilistic storm location for 90% of 2025 storms, while AI models still misjudge intensity and ensemble spread. arXiv
  • MolHop organises 1.6 million commercial scaffolds into a hierarchical action space for scaffold-hopping drug design, and its 9B agent beats GPT-5.6-sol, Claude-5-Opus and Gemini-3.1-Pro while producing more structurally distant candidates. arXiv

Free offerings

  • Meta published the weights of Muse Spark Open, an 816-billion-parameter mixture-of-experts with 42B active and a 1M-token context, under the Muse Model License, which requires model-as-a-service providers above $20m revenue to sign separate terms. new Meta
  • NVIDIA released Nemotron 3.5 Bolt 74B-A9B under OpenMDW-1.1, a 2-bit quantisation-aware-trained vision-language model packed into 24GB that scores 50.98 on Terminal-Bench 2.1 against 44.38 for Qwen3.6-35B-A3B. new NVIDIA
  • Black Forest Labs opened the weights of FLUX 3 Image Klein, a 5.6B image generator and editor that runs on 8GB of VRAM in INT8, free for commercial use below $5m annual revenue. new Black Forest Labs

Building on frontier models

  • Anthropic took fast mode out of research preview for Claude Opus 5.5, so the beta header is no longer required and the model now carries a default fast-mode rate limit on the Start, Build and Scale tiers. Anthropic
  • OpenGCUA released an open computer-use agent stack that mixes GUI actions with code execution, raising mean partial score on OSWorld-v2 by 19.6 points across eight models while cutting reported steps per task by 53%. arXiv
  • SpecAgent has a weak model draft tool-call sequences while a strong model verifies each segment in a single prefill pass, matching the strong model within a point on tau-squared-Bench at half the wall time. arXiv
  • JevAgentHarness delegates an agent's multiple-choice decisions to a 4B model above a fitted confidence threshold, making agents 81-89% cheaper and up to 2.5 times faster without losing three points of task success. arXiv

Cost & hardware

  • Meta published the first hyperscaler account of training recommendation models on AMD Instinct MI350X rather than NVIDIA B200, reporting up to 1.5x throughput over a direct port, with its work upstreamed to PyTorch, Triton and ROCm. arXiv
  • TrimSO(2) folds representation transformations into the dataflow of SO(2)-based equivariant interatomic potentials, giving 2.07-2.94x training speedups and 1.53-3.16x lower peak memory for eSEN, UMA and EquiformerV3 on A100 and H100. arXiv
  • ByteDance is in advanced talks to buy custom data-centre AI chips from MediaTek's ASIC business, which would make it MediaTek's second-largest ASIC customer after Google and cut its exposure to US limits on Nvidia hardware. TechNode

Evals & guardrails

  • Translating only the user's request into Swahili or Amharic flips agent-guard verdicts beyond the guards' own noise, with placebo-adjusted flip rates reaching 4.4 points on Amharic for the nine guards below 13B parameters. arXiv
  • A scan of 21,097 MCP servers found 787 of 4,806 triaged servers carrying filesystem capabilities their declared function did not need, confirmed 150 exploitable instances and produced 13 advisories including two CVEs in Anthropic's Git MCP server. arXiv
  • Telling a model it is being tested raises refusal of harmful requests by 14.4 percentage points across five Indic languages, but also pushes refusal of benign Indic requests from 6.9% to 11.3%. arXiv
  • Poisoning a protein language model's pre-training corpus produced reliable trigger-dependent generation after as few as 200 contaminated sequences, and one crafted retrieval record captured a top-16 position for 97% of free-text design requests. arXiv

Algorithmic advances

  • Specialist models sharing a base can be fused with no training at all: a layer-wise least-squares fit from one forward pass over a few hundred rollouts keeps most of each specialist's gain on Olmo 3 7B. arXiv
  • Waypoint-OPD gates reinforcement-learning policy-shift distillation on whether a teacher's early and late shifts agree, and takes the best domain average in all six maths and code settings for one extra frozen-teacher scoring pass. arXiv
  • LangDiT redesigns the transformer for continuous diffusion language models, reaching 21.72 generation perplexity at 1,024 function evaluations on OpenWebText while leaving each model's objective and sampler untouched. arXiv
  • SoftCache lets the headroom in a training-free diffusion cache set how far each cached step jumps, beating every baseline on four fidelity metrics at 3.5-6.9x speedups on FLUX, HunyuanVideo and Wan2.1. arXiv