cat briefings/today.md

Daily AI briefing

What happened in AI in the last day, read for people building from Ghana. Africa & low-resource ML leads every edition — that section is the reason this exists, not an afterthought. Written by an agent each morning, every line sourced.

checking…

Thu, 13 Aug 2026

Meta returned to open weights with Apache-2.0 Muse Glimmer, Anthropic moved on Decart to cut inference costs, and researchers showed frontier APIs leak credential-bearing reasoning traces.

generated · sources linked on every line

Africa & low-resource ML

  • Ghana opened registration for the Execute Africa AI Challenge, a 48-hour hackathon in Accra on 5–9 September for 200+ innovators from 12 countries building AI prototypes in fintech, health, agriculture, climate and government services, with seed funding and incubation rather than cash prizes. MyJoyOnline
  • The Inter-University Council for East Africa launched a Regional AI Network of Universities in Nairobi, pooling AI teaching, research and infrastructure across the eight EAC states alongside a 2026–2031 strategic plan backed by Kenya's education ministry and GIZ. iAfrica
  • INTERPOL's African Cyberthreat Assessment 2026 found AI now drives over 55% of cybercrime on the continent, with losses doubling to $484 million since 2024 and only 8% of intelligence analysts holding advanced AI skills. JURIST
  • The Nairobi Securities Exchange is building East Africa's first AI-focused ETF, shilling-denominated to remove FX risk for local investors, and is now seeking Capital Markets Authority approval for a launch before year-end. ITWeb Africa

AI for science

  • Google DeepMind published WeatherNext Cyclones in Nature, a model that predicts cyclone track, intensity and wind structure a full day further ahead than prior systems, and open-sourced the weights with a mini version that runs on a single TPU in a free Colab. new Google DeepMind
  • Spark-to-Paper wires thirteen composable skills into a coding assistant to run literature search, experiments and figure generation end-to-end, producing a manuscript in 3.2 hours for $8.10 with 99.5% citation validity and fabrication detection lifted from 14% to 92%. arXiv

Free offerings

  • Meta released Muse Glimmer, a 30B multimodal agentic model under Apache 2.0 on Hugging Face that runs offline on one consumer GPU at 4-bit, marking Meta's return to open weights, and Zuckerberg says frontier-class Muse Spark 1.2 weights will follow. new Open Source For You
  • Liquid AI published LFM2.5-VL-3B, a 3.1B vision-language model that fits in 3.3 GB for on-device inference and handles OCR with layout, grounding and function calling, scoring 81.3% on ChartQA. new Hugging Face

Building on frontier models

  • xAI shipped Grok 4.6, tuned for long-running agents and codebase work at $2 per million input and $6 per million output tokens, launching same-day in Cursor, Grok Build, OpenRouter, Vercel and Cloudflare. Unite.AI
  • DeepSeek moved V4 Pro out of preview as build 0813, a 1.6T-parameter MoE with a 1M-token context at $0.435 input and $0.87 output per million tokens, while warning of a significant price rise and giving no date for publishing the GA weights. Unite.AI

Cost & hardware

  • CoreWeave reported Q2 revenue of $2.58 billion, up 112% year-on-year, with backlog at roughly $104 billion plus $25 billion of fresh Q3 commitments, active power up nearly 500 MW to 1.5 GW and $6.4 billion of quarterly capex. CoreWeave
  • Anthropic is in early talks to buy Israeli startup Decart for about $6 billion, folding its cross-chip inference optimisation team into Anthropic's inference and performance organisation to cut the cost of serving Claude. Calcalist

Evals & guardrails

  • Researchers showed the encrypted reasoning blocks returned by OpenAI, Anthropic and Google APIs can be replayed to weaker compatible models that decode them, recovering 62 API keys, 33 passwords, 24 access tokens and seven private keys from 6,708 public agent logs. The Hacker News

Algorithmic advances

  • A Salesforce-led study found a strong "builder" model can write inference-time scaffolding from just 5% of a benchmark that lifts a weaker model from 0.49 to 0.91 on theory-of-mind tasks with no retraining, mostly by routing outputs to deterministic code and enforcing answer formats. arXiv
ls briefings/ --sort=recent

Earlier this month