Australia still needs expert humans

Artificial intelligence leaders think the technology might get smart enough to improve itself without human help within a year. But they’re not sure.

They’re confident AI will be able to match virtually any human skill, but they can’t agree whether that will happen by 2028 or 2035. The tempo of technical breakthroughs is downright disorientating. And the world’s leading economists can’t agree whether AI is a jobs killer or a generator of unprecedented new demand – or both.

It’s lucky elected policymakers aren’t allowed to throw up their hands and declare it all too hard. As Andrew Charlton, the assistant minister with responsibility for AI, said this week with champion understatement, ‘You are creating policy in an environment of great uncertainty.’

What does Australia do amid this sprint into the unknown? Charlton this week in a speech to the Australian National University’s Crawford School and a two-hour appearance on the Joe Walker Podcast sketched out the most detailed roadmap to date – the glaring conclusion from which is that the safe bet is on people and skills. Whatever else happens – and nearly anything could – raising world-leading Australian talent will be a no-regrets investment and should be a top priority.

AI has three key elements: talent in the form of researchers and engineers who think up innovations in AI architecture and code, data on which models are trained, and computing power to crunch it all. Australia absolutely needs to build computing capacity in the form of data centres; Charlton was clear on that. We also have high-quality data ranging from healthcare to resources.

Where Charlton gave us something new was in arguing that data centres are not enough. Australia needs computing power as a foundation from which to climb up the value chain to where the real returns are made – the intellectual property, such as AI models and applications. That means skills and talent.

We’re all familiar with blockbuster frontier models such as GPT, Claude, Gemini and, increasingly, Chinese competitors such as DeepSeek, Moonshot’s Kimi and Alibaba’s Qwen. But, as Charlton argued, there’s a burgeoning industry of smaller models that are optimised for specific tasks and don’t need the massive grunt of the frontier models.

Australian AI watchers have debated for months whether we should get into the model-building business, but it’s often an all-or-nothing proposition: we go big or we stay home. Charlton argued that breakthroughs in fine-tuning, post-training and distillation, each of which allows users to take an existing model and tweak it for their own uses fairly easily and cheaply, boost the case for getting into the model business.

Now there is a catch: Charlton’s prescription rests heavily on using what are called ‘open-weight’ models, which can be downloaded and adapted by users. They’re much cheaper to use than proprietary models. But the best open-weight models are Chinese, raising security and strategic risks. Australian banks, telcos and power company can’t rely on Chinese open-weight models. Limits on the use of Chinese models is one of the big unanswered questions in this formulation.

It doesn’t necessarily knock the wider argument off course. There’s a strong open-weight push in the US as well. And there are reasons – both economic and security-related – to think China’s government will start to crack down on the open-weight proclivities of its companies. And, as Charlton also pointed out, there is global competition. He cited France’s Mistral, which is the best model builder outside the two superpowers. The industry has a lot of evolving to do yet.

This brings us back to skills and talent. Model-building skills are going to be vital, whether it’s building from scratch, tinkering and improving, or evaluating strengths, weaknesses, risks and dangers of new offerings on the market. These skills will be transferred and deployed as needed, as the specific uses become clear.

Best of all, attracting investment from AI labs and big tech companies in data centres that can be used for training new models, as opposed simply to running them, will generate research clusters that grow the skilled workers and give them reasons to stay. The government should strongly encourage this and extend it further by funding centres of excellence and elite training colleges at universities, in cooperation with industry.

Rightly, the government plans to attach conditions to foreign investment in computing infrastructure, including a demand that the investors reserve some of the capacity for Australian companies and researchers.

‘Our goal,’ Charlton said in his Crawford speech, ‘is to convert physical investment into national AI capability.’

AI will play a huge part in determining the global pecking order in a world in which economic and strategic power is up for grabs. Models lie at the heart of the stack – the layers of value from the chips through to the applications. The people who know how to make them constitute a global elite that is probably the most sought after talent on the planet, as the eye-watering Silicon Valley salary packages testify.

Models are already becoming a strategic resource. Caught between a risky China and a capricious US, Australia’s best bet at sovereignty and agency is home-grown skills in the decisive technology of our lifetimes.

This article was published by The Strategist.