Wednesday 5 August 2026

The Brief – 06/08/2026

The top things worth knowing about in AI today.

  1. Jeff Dean leaves Google, Hassabis steps back

    Jeff Dean is leaving after 27 years to co-found Discovery Loop, a public-benefit company automating scientific and engineering experiments, taking Oriol Vinyals, Quoc Le and Sanjay Ghemawat with him. Demis Hassabis moves off day-to-day leadership of Google DeepMind to become Chief Scientist of Alphabet, with CTO Koray Kavukcuoglu promoted to oversee Gemini development. Alphabet shares fell 5.4 per cent on the news, and Alphabet is backing the new startup as a founding investor. If you build on Gemini, the people setting its direction changed this week.

    Read more
  2. OpenAI passes one billion active users

    OpenAI says its models now reach more than one billion active users and two million businesses, up from 900 million weekly ChatGPT users. It also cut the price of GPT-5.6 Luna by 80 per cent to $0.20 per million input tokens and trimmed Terra by 20 per cent, crediting efficiency work including model-written GPU kernels that reduced serving costs 20 per cent. Scale has not produced profit: the company lost $21 billion on $13.07 billion of 2025 revenue. For anyone budgeting API spend, the mid-tier models just got substantially cheaper.

    Read more
  3. EU begins enforcing AI transparency rules

    From 2 August the EU AI Office can investigate general-purpose model providers, demand access to models for evaluation, and fine companies up to 15 million euros or 3 per cent of worldwide turnover. Chatbots must tell users they are automated, deepfakes must be labelled, and AI-generated content needs machine-readable marking. The obligations apply to systems already on the market, not only new ones. If your product serves EU users, the disclosure work is due now.

    Read more
  4. AI could lift Australian productivity 2.4 per cent

    A report EY released today estimates AI could raise Australian productivity by up to 2.4 per cent, worth $116 billion in real GDP and roughly 44,000 jobs, against a decade in which labour productivity grew 0.3 per cent a year. Australia is the largest destination for data centre investment outside the United States, but approval times and grid upgrades are the constraint. Energy Minister Chris Bowen said data centres will have to underwrite new renewable generation to offset what they consume. More than 80 per cent of Australians want stronger rules on how organisations use AI.

    Read more
  5. Claude outage takes four models offline

    Anthropic services went down for roughly 7.5 hours on 5 August, taking Mythos 5, Fable 5, Opus 5 and Sonnet 5 offline across the web app, API and Claude Code. StatusGator counts 164 documented Claude incidents since January. Teams putting agents into production workflows should assume multi-hour single-provider outages and plan fallbacks accordingly.

    Read more
  6. Sandisk and SK hynix standardise AI memory tier

    The two companies published the first Open Compute Project technical specification for High Bandwidth Flash, a NAND-based memory layer that sits between HBM and SSDs for AI inference. The spec covers the host interface, electrical guidelines, die-stack packaging and read and write software, six months after the OCP workstream formed. Google and Tenstorrent joined the consortium during standardisation. Inference cost is increasingly a memory-capacity problem, and this is the industry agreeing on how to address it.

    Read more
  7. Microsoft merges Copilot into one app

    Microsoft is consolidating Copilot chat, GitHub Copilot, Copilot Cowork and its Autopilot agent tier into a single app this month under an internal programme called Copilot Fusion. The context is weak conversion: fewer than 4.5 per cent of 450 million Microsoft 365 commercial seats have moved to paid Copilot, and only 20 to 30 per cent of paying users open it weekly. Copilot Podcasts and Copilot Labs have been cut for low engagement. Autopilot will be priced separately because background agents make 10 to 20 model calls per task rather than one.

    Read more
  8. DeepSeek V4 Flash posts large coding gains

    DeepSeek moved V4-Flash-0731 into public API beta and published the weights on Hugging Face, holding the price at $0.14 per million input tokens. Terminal Bench 2.1 rose to 82.7 from 61.8, and DeepSWE jumped from 7.3 to 54.4. The model is a sparse mixture-of-experts design with 13B active parameters out of 284B and a one-million-token context window. Agentic coding at this price puts real pressure on mid-tier pricing across the closed labs.

    Read more
The Weekly AI Brief

Practical AI, distilled.

A short read every week — the few things worth your time, and nothing that isn’t.