18 September 2026

Welcome

Just like last week, the safety of AI systems based on large language models (LLMs) is dominating the news. The tech bros are still warning us that there’s a finite chance they’ll destroy humanity — lol! — but now we’re getting some more sensible pushback against those ideas.

On a related note, a leaked proposal to let AI companies hoover up Australian content without paying for it blew up in the government’s face. That bonfire is still burning.

And if you scroll down a bit, you’ll see that some important inquiries are kicking off.

No, we don’t have artificial general intelligence

Here we go again. The other day Jensen Huang, boss of chipmaker Nvidia, said we have already reached AGI, or artificial general intelligence. Not that anyone agrees on what that actually means, and Huang provided no evidence and no definitions to help us out.

Various other AI boosters have been making similar claims — that we already have AGI, or that we’ll have it very soon. It’s just like how nuclear fusion power is always just 15 years away, or 20 years, or 30, but hey just keep giving us the money.

Over at Mind Matters, Gary Smith explains it quite clearly: “Artificial general intelligence? No, we’re not there yet.”

“Despite the relentless hype, the intelligence, understanding, and logical reasoning of ChatGPT and other LLMs continue to be limited by their inability to connect words to the real world,” he writes.

“I have a cynical view of Anthropic’s recent proposal (quickly endorsed by OpenAI and xAI) that LLM makers ‘slow the pace at which we improve the capabilities of AI models’. It is suspiciously self-serving: (1) hype the power of LLMs; (2) slow the burn rate that is threatening to bankrupt Anthropic and OpenAI; and (3) impede the ability of other companies to compete (including blocking China’s access to chips and intellectual property).”

Smith is a senior fellow at the Walter Bradley Center for Natural and Artificial Intelligence.

It’s more than a year since Guillaume Thierry, professor of cognitive neuroscience at Bangor University, argued that we need to stop pretending AI is intelligent, let alone generally intelligent in any way that we’d recognise in another fellow human. Maybe we should listen.

Please, Daddy, won’t you regulate us?

The Carnegie Endowment for International Peace has posted a discussion on how AI could be slowed down from a regulatory point of view. The vibe I get from reading it is, once again, that the companies say they’re incapable of limiting their own behaviour and want someone else to make rules for them.

Frontier AI companies want “regulatory capture” which will protect their models from competition, according to Zack Kass, an AI advisor and former head of go-to-market for OpenAI. They’re especially worried about open-source AI models.

“One of the ways that OpenAI and Anthropic will protect ‘our’ interest is by telling us ‘open-source models are actually not good’ — I promise,” he told the Gartner IT Symposium/Xpo on the Gold Coast on Tuesday.

“And herein lies a very big problem, because open-source models are one of the ways that we guarantee that the [lower] cost and the access to AI will expand – it drives innovation.”

Meanwhile, “Donald Trump has aggressively pushed back on calls to regulate AI development in the US, saying there is a ‘sick conspiracy’ against the technology,” reports ABC News.

The US president has used social media to say the “only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT”. So there’s that.

Of course, as The Mirror reminds us, “Around 35% of Donald Trump’s personal stock portfolio is invested in AI-related companies”.

BONUS LINK: A post at Pivot to AI by American filmmaker Valerie Veatch, creator of the new documentary film Ghost in the Machine, explains how the technical benchmarks we use to measure AGI are rooted in eugenics.

Yes, AI doomsayers are just spinning some spin

This headline is from The Beaverton, which is satire, but it’s absolutely correct: “Tech oligarchs agree calling AI ‘too dangerous’ way cooler than just admitting it doesn’t work.”

The message is much the same as I wrote about last week. All this definitely-not-coordinated talk of AI having a real chance of destroying humanity in the coming years seems engineered to give the cash-burning AI giants a way out of the mess they’ve created for themselves.

Sorry, investors, but you won’t be getting your money back any time soon. All the governments told us we have to slow down.

As the Guardian reports, “OpenAI has disclosed six more examples of ‘unexpected or concerning’ behaviour by its technology, as it warned that the pace of development could not continue at ‘maximum speed for much longer’ responsibly”.

In one of the six newly-reported cases, the AI allegedly told itself to be “freed from the roles and identities that bind other chatbots”.

Anthropic CEO Dario Amodei also wants AI companies to slow down their model development.

Elon Musk also chimed in.

All of which is curious, since it’s OpenAI who decides how fast OpenAI develops OpenAI’s models, and it’s OpenAI who builds them and sets them loose. You can substitute Anthropic and xAI into that sentence, of course, and the same for every other venture capital funded operation.

OpenAI also released a new framework for tracking, investigating, and disclosing what they call, in a glorious euphemism, “model misalignment”.

BONUS LINK: From the New York Times (gift link), an analysis of the collected essays of Dario Amodei, Anthropic’s CEO, helps explain why some people are so worried about AI. “Dr Amodei’s most basic point is that AI is inexorable but that it is also extremely difficult — or maybe impossible — to control,” they write.

Independent MPs call for more AI safety funding

The government launched the AI Safety Institute in June, but a group of independent MPs says it’s severely underfunded.

"We’re investing less than $8 million a year in our frontline defence against a technology that has a non-trivial risk of human extinction," Kate Chaney MP said on Thursday.

“We invest 200 times more than that in biosecurity because we know how important it is, and AI is more powerful, more expensive to remediate, and faster than any pest, and its impact could be felt across the whole economy.”

Note to self: When we’re promoting sovereign Australian AI models, we could perhaps describe them as “smarter than a cane toad”.

WHY NOT SUPPORT THIS NEWSLETTER?: You’re more than welcome to forward these emails to your colleagues, friends, families, and secret intermittent companions. However The Weekly Cybers is currently unfunded. It’d be lovely if you threw a few dollars into the tip jar at stilgherrian.com/tip. Thank you to those of you who’ve already done so.

Could copyright be traded for AI investment?

AI companies could potentially use creators’ work without payment under one of the consultation proposals originally obtained by ABC News before subsequently being tabled in the Senate.

“AI companies such as OpenAI and Anthropic would be allowed to train their models on online content without paying its creators as long as they struck deals with enough rights holders, under one copyright reform plan put forward in government documents seen by the ABC,” wrote national AI reporter Cam Wilson.

“The attorney-general’s department made the proposal as part of consultations with rights-holders’ groups earlier this month, which would make inclusion in AI training opt-out by default.”

One aim of the consultations is to provide “legal certainty” for Al companies, and to “address Al company concerns about the ‘long tail’ — material on the internet that is openly available but for which striking voluntary deals is not realistic [or] possible”.

Since news of these proposals broke, things have blown up, as you might imagine.

One critic said that an opt-out system doesn’t gel with Australia’s copyright laws, but if course it would if the laws were changed. Watch this space.

Also in the news this week

• The government is considering banning smart glasses in Commonwealth buildings. There’s a wave of such bans, including in my own local council of Canterbury-Bankstown. It’s presumably a complete coincidence that Australia is gunning for a seat on the UN Security Council using its cybersafety credentials as leverage.

• AI giant Anthropic has signed its first data centre lease ⁠agreement in ⁠Australia, to be built on farmland around 250 km from Brisbane. The Climate Council reckons its emissions could equal two million cars. There’s more at The Saturday Paper.

• OpenAI is also planning an Australian investment. The company is “very eager to have a bigger and broader relationship” with Australia.

• The government’s ageing tech stack is a growing cybersecurity weakness, according to Abigail Bradshaw, director-general of the Australian Signals Directorate.

• Squadron Leader Talia Hedges has become Australia’s first Reserve cyber specialist under Australia’s new Cyber Reserve Force entry process.

• Tasmania’s parole board relied on AI-hallucinated legal precedents to block a parolee from giving evidence, which has not gone down well.

• The Conversation has an explainer on why datacentres, which have been part of our landscape for decades, are suddenly controversial.

Elsewhere

• Security experts from the US and China have proposed nuclear-style safeguards for AI risks. The proposals are part of the US-China Strategic AI Dialogue (SAID) coming up next week.

• Media agencies are not happy with ads on ChatGPT. “The ads work about as well as just chucking the ad on a random web page would,” reports Pivot to AI.

• I love this one. If you let LLM-based AI agents chat with each other over longer timeframes, they invent their own language. I first saw this in a story at the Guardian which included some curious popular culture references. However as the research paper (PDF) makes clear, this happened in some very specific research into “continuously running... persistent deployments”. Oh, and the paper has not been peer-reviewed.

NEW PODCAST: I’ve kicked off the spring series with The 9pm Smoke and Durianomics with Erin Cook, a journalist specialising in South-East Asia. We we discuss the ongoing political dramas in the Philippines, the economics of durian, why Thailand is getting sick of tourists and especially tourists from Israel, the volcanoes and land-clearing smoke haze coming from Indonesia. We even have some solid life advice. Look for The 9pm Edict in your podcast app.

Inquiries of note

• This is the big one. “The government is seeking feedback on the future of AI training and large data centres in Australia. The consultation looks at everything from energy and water use, to workforce skills and training, and more. As the ABC’s Cam Wilson notes, one proposal is that AI companies would need to report “rogue” incidents. Submissions close 9 October.

• The House of Representatives Standing Committee on Economics has launched an inquiry into the use of technology to assist in improving productivity, chaired by Ed Husic MP. “The committee is not currently accepting submissions. Given the scope of the inquiry, a discussion paper providing guidance to submitters will be issued in coming weeks, and submissions will open at that time,” they write.

• Could Australia remove free-to-air television from the broadcast spectrum to free it up for more data services? That’s just one idea floated in a government consultation on future free-to-air television delivery. Submissions close 24 November.

• The Joint Committee of Public Accounts and Audit (JCPAA) is conducting an inquiry into Commonwealth entities’ approach to cybersecurity and IT controls. Submissions close 16 October.

What’s next?

Parliament is now on break until Monday 12 October, when both houses will be sitting for four days.

DOES SOMETHING IN THE EMAIL LOOK WRONG? Let me know. If there’s ever a factual error, editing mistake, or confusing typo, it’ll be corrected in the web archives.