cat briefings/2026-09-07.md

Daily AI briefing

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

Mon, 7 Sept 2026

OpenAI declared its automated research intern goal met on schedule, while τ^τ-Bench showed frontier models still fail badly at building the agents everyone assumes they can build.

generated · sources linked on every line

Africa & low-resource ML

  • The University of Ghana will make AI training compulsory for every student from the next academic year, Prof. Isaac Wiafe of the Department of Information Technology's Interaction Labs said on 4 September at the launch of the 2026 Achievers in Technology Ghana Fellowship. The Ghana Report
  • KNUST AI coordinator Prof. Eric Appau Asante told Ghana's PRINCOF conference on 2 September that colleges of education should embed AI competencies in core teacher-training curricula rather than optional workshops, and that each institution should write an AI policy aligned to the National Teachers' Standards. Rainbow Radio

AI for science

  • Hessian-based conformation augmentation generates extra training geometries from vibrational modes, improving machine-learned interatomic potentials for vibrational analysis and transition states without changing the model architecture, so it drops into existing MLIP pipelines. arXiv

Free offerings

Nothing notable today.

Building on frontier models

  • OpenAI said on 6 September it has met its target of an automated research intern that carries out well-defined research tasks under human direction, including work that would take a skilled researcher several days, reporting 3.1 agent-workdays of runtime per human workday by mid-August and a median researcher spending over $600 a day on inference at API prices. Engadget
  • Google's MaxKernel puts an agent loop around TPU kernel writing and matches expert hand-tuned baselines across the 50 kernel tasks in its new JaxBench suite. Google

Cost & hardware

  • Rystad Energy puts 2025 global data-centre water use at 222 billion litres and projects 644 billion litres by 2030 without mitigation, while Nvidia says its 45°C closed-loop DSX designs can almost eliminate water draw at some sites — a saving that trades directly against electricity for cooling. Tech Xplore

Evals & guardrails

  • Sierra's τ^τ-Bench asks models to build a working agent end to end rather than act as one, and Claude Opus 5 running under Claude Code passes 23.9% of evaluation simulations against 82.2% for the expert-authored reference agent. Sierra
  • The Seattle Times and Newsday sued OpenAI and Microsoft over training data on 5 September, joining the growing group of publishers in litigation. TechCrunch
  • Authors are contesting attempts by publishers and literary agents to claim a share of the Anthropic copyright settlement, disputing who is entitled to the payout. TechCrunch
  • Hikers had to be rescued after relying on Google Gemini to plan their route. TechCrunch

Algorithmic advances

  • Cerebras ran over 2,400 experiments on layer dropout and found the right sparsity schedule saves up to 25% of training FLOPs and delivers up to 1.5× inference speedup, arguing dropout should be tuned rather than discarded. Cerebras
  • On-policy distillation needs almost no data: a single training query already covers 71.5% of the student's visited states and 16 queries reach 98.9%, matching full-dataset training. arXiv
  • Salesforce's RISE recursively distils a policy into a self-extrapolated improvement of itself and beats both RLVR-only training and on-policy self-distillation across maths, STEM, code generation and agentic tasks. Salesforce