Tue, 8 Sept 2026
East Africa named the six universities that will anchor its shared AI curriculum, and diffusion-augmented LLMs delivered 3× lossless inference speedups without a draft model.
generated · sources linked on every line
Africa & low-resource ML
- East Africa's federated AI Centre of Excellence named its six anchor universities on 7 September at the higher-education quality assurance forum in Kigali — University of Rwanda, University of Dodoma, Dedan Kimathi, SIMAD, University of Burundi and Kabale — tasked with writing a common AI foundational module for every student in the region. Web Rwanda
- Ugandan startup SANDI AI took the $50,000 grand prize at the GoGettaz Agripreneur Prize at the Africa Food Systems Forum in Kigali for AI scoring that gets smallholder farmers loans without collateral, and will use the money to expand beyond Uganda and Kenya. Disrupt Africa
- The One Vecta AI Summit opens today at The Palms by Eagles in Accra, running two days across AI in fintech, healthtech, agritech and climatetech with a MEST Africa startup track and speakers from Microsoft, Amazon, Gartner and Fidelity Bank Ghana. MyJoyOnline
AI for science
Nothing notable today.
Free offerings
Nothing notable today.
Building on frontier models
- Iris trains search agents by reverse-engineering web link structures into tasks and then running reinforcement learning against live search engines, reaching 88.6% on BrowseComp and 92.9% on DeepSearchQA, with weights and the full data and training recipe promised. arXiv
Cost & hardware
Nothing notable today.
Evals & guardrails
- Enoki detects hallucinations at both the claim and span level using open information extraction, matching strong claim-level systems on far fewer resources while localising errors down to individual spans and entities, swapping between LLM, encoder and rule-based extractors to trade cost against accuracy, and ships alongside EnokiQA, an openly licensed dataset pairing claim-level verification labels with span-level localisation annotations. new arXiv
Algorithmic advances
- Diffusion-augmented LLMs bolt lightweight diffusion weights onto a normally trained autoregressive model, giving up to 3× speedups at every batch size with no quality loss and no draft model, with the 8B Uno beating 26B DiffusionGemma and the proprietary Mercury 2. arXiv
- FlowBalance lets a reasoning model learn from its own rollouts by calibrating dense token-level guidance against a terminal verifier — keeping it on positive-advantage trajectories, reversing it on negative ones, disabling it when the group is indifferent — beating FlowRL on Qwen3-4B and 8B maths without response-length collapse. arXiv