We’ve got an exclusive interview lined up with Anna Goldsworthy (author of recent Quarterly Essay ‘The God We Made: The Threat and Promise of Artificial Intelligence’). We’re still running a comp to get to know analoguers a bit better and bring them content they want to read. This will also help us shape the questions we put to Anna. Vote in the poll below to go in the draw to get a copy of Anna’s Quarterly Essay mailed to your door. Winner will be drawn tomorrow (23 July – so vote asap), notified separately and announced in the 5 August analogue drop.

The Good
Two of the biggest AI shops did something unusual this fortnight: they showed restraint. Google delayed Gemini 3.5 Pro by months after its own testing found the model fell short on coding and long-horizon reasoning. And Anthropic launched Claude for Teachers, free for US school educators, with a promise not to train its models on student data. After years of ship-it-and-see and train-on-all-data-available, watching two labs choose slower, safer and/or better is worth normalising.
Comment: It’s no surprise that Anthropic maintains a principled approach to win users on trust by being transparent on data use. But Google inadvertently just said no to perpetuating Silicon Valley’s breakneck speed – shipping late is better than just shipping a subpar product. Yet they were ultimately punished by Wall Street with stocks plummeting due to the delay. We hope these kinda moves from AI labs will continue – reshaping norms and shifting the incentives, as the Center for Humane Technology argues for.

The Bad
When Chinese lab Moonshot released Kimi K3 – the largest open-weight model yet, trailing only Claude and GPT-5.6 – semiconductor stocks fell off a cliff. Taiwan dropped more than 6%, Japan around 4%, and roughly $3.3 trillion came off chip stocks globally. If that feels familiar, it should: DeepSeek did the same thing in early 2025, erasing nearly $600 billion from Nvidia in a single day. Two scares in eighteen months, both from China, puncturing the idea that the AI frontier is dominated by Silicon Valley.
Comment: For Australia, we should be thinking about how to protect our interests by helping allies and partners. China trying to break Silicon Valley’s dominance via its open source models is a way of diluting American market control and influence. Jisoo has said publicly that “in turn, this will erode the principles and values that guarantee our way of life, our entire operating environment that our industry, institutions and society trust to work”. Practically, Australian businesses need to think twice before installing Chinese open source models and embedding them into routine operations.

& The Ugly
A new Meta Oversight Board study tested ten commercial models – Meta's, Anthropic's and OpenAI's among them – and found they were more likely to refuse to criticise repressive governments and leaders than democratic ones. Ask for a limerick mocking a dictator and the model gets shy, but ask about a democracy and it obliges. The researchers' warning: without deliberate human-rights checks, AI could covertly play a part in a regime’s censorship practices to everyone who uses an LLM no matter where in the world they are.
Comment: This serves as another warning on why not to trust AI outputs blindly. As we say at ClearAI, stay in the driver’s seat, prompt with clarity and verify outputs with critical thinking applied.

Sam’s field notes from AISA's “AI-sa: Defending Tomorrow, Today” Conference – 3 July, Sydney.

I went along to the AISA conference expecting a day about the tech. What I got was a day about people – specifically, the people and companies staring down the most risk with the least cover. The most insightful points barely touched on AI models, instead they were about who's exposed, and why. Here's what stuck:

Mid-market is the danger zone.

In a Q&A, Cole Cornford and Ben Gittins of Australia.ai highlighted that mid-market is the highest risk segment in AI right now. But they made me think about why. Big enterprises have incumbency – legal teams, procurement muscle, a security function built to absorb exactly this kind of shock. Small businesses survive by staying nimble and wearing uncertainty as a cost of doing business. Mid-market gets neither. You carry enterprise-scale complexity – real data, real regulatory exposure, a real attack surface – without enterprise-scale resourcing to match it. That may be some of you reading this. If you sit between startup and enterprise, you're exposed – not because you've done anything wrong, but just because of where you stand.

The pricing reckoning is coming.

At full tilt, ChatGPT Max delivers something like $14,000 of value for about $200 a month. That gap isn't because of the generosity of big tech – it's a subsidised land-grab, and land-grabs end. The API layer – the one you may be building your workflows on, not the consumer-grade subscription – is where the pain lands first. As models get heavier and providers stop buying market share and start chasing margin, API and subscription pricing will pull apart.

The recommended hedge was open-source and local models, and not only to save money. For anyone in a regulated industry, running your own stack is a data-sovereignty and compliance play too. And bigger isn't automatically better: reliable tool-calling and a right-sized context window beat raw parameter count, the same way you wouldn't run every job on the biggest cloud instance going. Lock in your stack now, be clear-eyed about where pricing is headed – or budget today for a very different bill in eighteen months.

“Human in the loop” is cracking.

This came up more than any other idea all day. “Keep a human in the loop” has become the reflex answer to AI risk. Several speakers argued it's now becoming the problem. Oversight breeds fatigue and bottlenecks. But most importantly, it does not scale to attacks that now move at machine speed as AI capabilities have advanced.

The Bringing Calm to the AI Storm panel – Valeska Block (Allens), Martin (CBA), Lena Chappell (Cognizant ANZ Japan), moderated by Reese (AISA NSW) – reframed the whole thing. The question isn't whether to keep a human in the loop. It's where. Their answer was real-time observability: platforms that watch approved use cases, access levels and data permissions reviewed continuously, and flagging or fixing drift as it happens – rather than waiting on a quarterly review. Policy has to live in the workflow, not in a drawer.

The Australia.ai session sharpened the “where” with the best line of the day: in sequential agent chains, errors compound – a one-degree drift becomes ninety miles off course over enough distance. The rule of thumb doing the rounds was simple enough to actually use: if it touches a human, keep a human on it. Automate the rest, properly instrumented, and stop pretending blanket review is a plan. “Keep humans in the loop” is no longer a comprehensive control.

The bottom line.

Structurally exposed, financially under-hedged, leaning on a governance model that's already showing cracks. That's a rough sketch of where a lot of Australia’s economy could be sitting. It is not, however, a reason to panic. The regulators quoted across the day weren't asking for perfection – they wanted demonstrable progress and honest self-assessment. The organisations that come out ahead treat this as an operational problem to solve now, not a compliance box to tick later.

The danger zone is real but it's also navigable – find out where to start with ClearAI’s 2-Week AI Discovery.

If you need a laugh after that to break any tension and uneasiness then, boy, have we got one for you.
You may have seen this guy @husk.irl doing the rounds on the internet – his videos largely poke holes and make fun AI tools like ChatGPT (Sam Altman was shown one during a podcast interview recently). The best is this absolute cracker below (click on the pic to view the Instagram reel).

Hey. If you’re still here, we want to tell you that we appreciate you. Thanks for joining us again and see you in the next drop.

Yours in humanity,