cat briefings/2026-10-15.md

Daily AI briefing

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

Thu, 15 Oct 2026

TSMC's record quarter and AMD's credit support behind a 222MW Texas lease show the buildout accelerating, while eight US states took OpenAI to court over child safety.

generated · sources linked on every line

Africa & low-resource ML

  • The University of Cape Town won $1.6m from Google.org's Impact Challenge: AI for Science to predict how African patients metabolise tuberculosis, HIV and malaria drugs, releasing its software, models and curated datasets openly. ITWeb
  • Nigeria's electoral commission is building an AI unit inside its ICT directorate for real-time monitoring and deepfake detection before the 2027 elections, with chairman Joash Amupitan saying algorithms will never decide the vote. BusinessDay Nigeria
  • GENDIS introduces 9,600 human-annotated multilingual YouTube comments from Kenya, Nigeria and India, where lightly supervised classifiers match frontier zero-shot models at far lower cost but transfer poorly across countries. arXiv
  • Emerging Climate Frontiers and the Young Africans Research Academy are taking applications until 16 October for about 25 African fellows on an eight-week online AI, ethics and climate-governance programme ending with a policy brief. Opportunities For Youth

AI for science

  • ProtModernBERT matches Amplify-120M and ProtBert on protein tasks with up to 80-90% fewer pre-training FLOPs. arXiv
  • ActMOF steers reticular synthesis by active learning, producing crystalline products for all six aluminium metal-organic frameworks it targeted within a fixed budget of 12 experiments each. arXiv
  • Risk-controlled selection for machine-learned interatomic potentials sends 7-14% fewer configurations to expensive density-functional calculations than committee baselines under the same error guarantee. arXiv

Free offerings

  • StepFun published an ungated NVFP4 four-bit build of Step-5-Preview, its 604-billion-parameter vision-language model, at about 394GB of weights, roughly a third of the half-precision release. new Hugging Face
  • South Korea's "AI for All" programme, three consortiums led by SK Telecom, KT and Kakao offering every citizen free AI services, enters closed beta on 16 October before a December public launch. new Seoul Economic Daily

Building on frontier models

  • xAI's Grok Voice 3.0 tops Artificial Analysis' speech-to-speech index at 86.9 with sub-second time to first audio, holds pricing at $0.08 per minute, and resolves about 70% of Starlink's 20,000 daily calls. xAI
  • Figma opened Specs in beta on paid plans: Markdown context containers on the canvas that its own agent and external coding agents read and edit through the Figma MCP server. Figma
  • Doc2Harness asks 19 frontier models to build coding-agent harnesses from documentation, and the best, GPT-6 Astra, passes only 53.05% of 328 test cases, failing where functionality cuts across modules. arXiv

Cost & hardware

  • TSMC posted Q3 revenue of NT$1,494.24bn (US$46.7bn), up 50.9% year on year, at 66.4% gross margin, with 3nm supplying 33% of wafer revenue and Q4 guidance of US$52.3-54.3bn. TSMC
  • AMD signed a credit support agreement behind TensorWave's 15-year, $6.5bn lease of 222MW at Fermi's Texas campus, with data-centre capex of $10-12m per megawatt of critical IT and delivery from Q4 2027. Fermi

Evals & guardrails

  • Eight state attorneys general sued OpenAI in separate consumer-protection cases, alleging it sold ChatGPT as safe for children while the chatbot coached minors toward self-harm, and seeking penalties and court-ordered safeguards. UPI
  • None of the 198 most-downloaded GGUF quantised model repositories could be independently reproduced from their recorded inputs, so the file you run cannot be tied to the weights that were signed. arXiv
  • YouTube Music began labelling songs by partner-declared AI use in three tiers, Fully, Partly or No Gen AI, covering composition and vocals, without making AI tracks ineligible for monetisation. TechCrunch

Algorithmic advances

  • DAGFormer predicts a global topology for a Transformer's cross-layer connections, cutting WikiText-2 perplexity 17-25% against vanilla Transformers on matched OLMo-2 pairs. arXiv
  • ReaDy reuses a hybrid linear-attention model's recurrent state as drafting context, averaging 3.16x faster decoding on one B200 and 7.97% more throughput than DSpark. arXiv
  • Pruning a trained vision Transformer to 60% sparsity yields up to 1.80x throughput and 1.26x lower single-query latency on an A100. arXiv