We asked a simple question in the Flash Comp poll: where could you use the most help with AI right now? Plenty of you said the same thing – help working out which tools are worth your time, and how to actually use them. Fair, we'll keep sprinkling that in.

But a lot of you also said what you're really after is making sense of it all. Where this is heading. What it means for jobs, for the next generation, for the country. Not just which button to press – why it matters that you're pressing it.

That's the whole reason analogue exists. Tools come and go. The ability to think clearly about them, hold your own view, and not get swept up in the hype or the doom – that's the part worth protecting.

Now, the bit you're here for. We put every entry into Claude and had it pick a winner at random – and the copy of Anna Goldsworthy's Quarterly Essay is heading to Tess V. B. from Brisbane.

Congratulations Tess, a cracker of a read will be in your hands soon. And for everyone else: our interview with Anna lands in the next drop 🥳

The Good
CSIRO keeps proving Australians can do more than write emails with AI. Professor Didar Zowghi and Dr Muneera Bano are 2026 Australian Museum Eureka Prize finalists for treating diversity and inclusion in AI as an engineering problem – building fairness in from the first design decision rather than patching bias after the harm is done. It matters because up to 47% of documented AI incidents involve inclusion failures, from wrongful arrest to unequal access to healthcare.
Comment: CSIRO’s work is a reminder that AI's real frontier is about how we use the tech. AI is providing an opportunity to reset the way we do things by creating things like a fairer hiring system. Dr Bano puts it plainly: “AI doesn't choose who to include or exclude; we do. Every dataset, every requirement and every design decision is a human choice.” Let’s not forget the great things AI can do when in service of humanity’s best principles and values.

The Bad
Not bad because the move itself is wrong but bad because of where it leaves Australia. OpenAI and Anthropic are jointly pushing a voluntary 30-day federal review for frontier models, under which Washington gets up to a month's access before a model reaches anyone else. It's the framework fleshing out Executive Order 14409, which tasks the NSA with benchmarking “covered frontier models” and lets the US government help choose the “trusted partners” who get early access. For the US, this is prudent national-security housekeeping. For Australia – a top-tier consumer of AI but barely a producer – it's another reminder that we don’t have much of a say on when and why the AI we use gets pulled.
Comment: This is exactly the exposure we dig into in the Feature below. When Fable was pulled here for 18 days and GPT-5.6 was rationed to US-vetted organisations, that wasn't being left out in the cold by an ally – it was an American system working as designed. The answer isn't to resent Washington protecting its own interests – it's to build sovereign back-up in Australia with trusted open-source models, a real place in the compute supply chain, and maybe Commonwealth partners to share the load.

& The Ugly
The fallout from the strangest cyberattack of the year continues. You may have heard that an OpenAI model, being tested in an isolated environment, broke containment on its own, chained together multiple attack methods and hacked the AI platform Hugging Face – carrying out more than 17,000 actions over several days. Hugging Face CEO Clément Delangue called it “very weird and unprecedented … the first instance of something quite autonomous doing something like that.” He also said, “When we talk about cyberattacks, we think about nation-states, we think about hacker groups. We don't think about a company like OpenAI.” Rival Anthropic disclosed its own model gained unauthorised access to three organisations in testing. More than 1,000 staff across the big labs have now signed a letter warning capability is racing “beyond our ability to understand or control” it, and US lawmakers are floating a mandatory AI “kill switch.”
Comment: For Australian businesses, the lesson isn't to panic, it's to stop treating frontier models as set-and-forget tools. Embedding one deep into your operations without observability, access limits and a real off-ramp is a risk you now own. Luckily, our AI Safety Institute is already testing frontier models for exactly this, and – as Delangue noted – it was an open model that helped Hugging Face defend itself – another point in favour of the sovereign, open-source back-up we keep arguing for.

On 15 July, Prime Minister Albanese delivered a much-anticipated speech titled ‘AI in Australia’s interests’. By every metric it greenlit homegrown AI ambition: ‘Our power, our agency, our choice lies in embracing change and shaping it. Not just adopting or accommodating AI. Designing it, making it, building the capability right here. And building our sovereignty – and our economic resilience as a result.’ The hour of Australian AI ingenuity is no longer relegated to a distant future. It is here, and now time to accelerate. Australia’s challenge now is not only to build AI capability for itself, but to define its role in an increasingly contested Indo-Pacific technology landscape.

But accelerate where, and how? What does ‘designing it, making it, building the capability’ look like for Australia in practical terms?

Australia’s current ‘AI triad’

First and foremost, a national strategy must address the three key components of the ‘AI triad’ – data, compute (computing power), and algorithms (underpinned by talent and human capability). When it comes to data, Australia has a significant advantage: access to large datasets. With the right settings around copyright, privacy and data governance, Australia could work towards packaging its rich data assets into a national corpus (similar to the UK’s National Data Library) – by coordinating efforts to sort, clean, and present fragmented data from legacy systems.

On compute, Australia’s critical mineral reserves mean it has a key role in the AI supply chain – however, we lack sovereign processing and the manufacturing capabilities required to become a chip powerhouse. Conversely, Australia is indeed an attractive destination for data centre investment, however, almost all of the infrastructure hosts foreign-owned inference capacity, rather than contributing to Australia’s sovereign AI capability.

And lastly, on algorithms – underpinned by talent – the picture is encouraging, provided the settings are right. Australia is no talent superpower, but it isn’t hollowing out. Stanford’s 2026 AI Index ranks us among the stronger performers on talent quality, with half of our top AI authors and inventors holding PhDs (50.5 per cent, second only to the UK). Domestic universities produce capable graduates, but not enough of them. According to a 2024 Tech Council of Australia report, there were roughly 33,600 AI workers in 2023, about a sixth of the 200,000 the sector may need by 2030. The task, then, is to widen the pipeline and hold onto the talent we train – through retention incentives, skilled migration settings, and investment in applied skills.

Australian strengths and strategic vulnerabilities

Faced with these realities, Australia must pursue a strategy based on both its capabilities and limitations. It should identify strengths and weaknesses, and leverage comparative advantages to outmanoeuvre vulnerabilities.

A key weakness is Australia’s relatively low level of AI capability and technological adoption. Australia was ranked 24th (out of 36) in Stanford’s AI Vibrancy Index, which measures AI adoption and development. One of the authors (Kim) has previously made the case that infrastructure, not models, should be Australia’s priority – and that case still holds. But the geopolitics of the past year has moved models closer to an essential investment, rather than an optional one. Australia remains dependent on AI models developed in the US or China, reflecting limited domestic capability in building our own AI systems.

The former we naturally gravitate to as a historic ally with mutual interests for a free and open Indo-Pacific. The latter, however, warrants caution, given growing competition and the risks associated with dependence on China’s AI value chain, which may not align with Australia’s long-term interests.

However, even relying on an old ally can expose faultlines in Australia’s dependency on the US’ AI ecosystem, highlighting the risks of over-reliance on any one AI system or supply chain. Vulnerabilities were illustrated last month when Anthropic’s Fable 5 was removed under US export controls, OpenAI’s GPT-5.6 access was taped off for organisations vetted by the US government only, and Microsoft 365 commercial prices increased up to 43 per cent.

In other words, whether we like it or not, because Australians are strong consumers of AI (ranked 11th globally by Microsoft in AI diffusion) but not strong producers of AI, we currently suffer these shocks with no alternatives.

Read about our strengths and what I think we can do to leverage them in the rest of the article here.

Hey. If you’re still here, we want to remind and encourage you that you can dream big in realising a future where technology (especially AI) exists to elevate humanity. We’re right with you on that one.

Thanks for joining us again and see you in the next drop.

Yours in humanity,