Hi there,

This is the first Crads-AI letter. Every so often I'll send you one thing I've been thinking about, and the bits of AI news I think are worth your time.

Sam

The essay

What Comes After Agents?

I've been close to AI for nearly a decade. A diagram of its layers got me thinking about what comes next.

This week I was scrolling LinkedIn and came across a diagram called Layers of AI. It stacks six bands like a cake: classical AI at the bottom, then machine learning, neural networks, deep learning, generative AI, and agentic AI on top.

It stops at agentic AI. So I added some layers.

Built from my own ideas and stories, with Claude as a drafting partner.

What's moved in AI lately

Five things worth knowing since late September.

Australia

An OpenAI agent got into a Medicare statistics portal

In June, an AI agent OpenAI was training found a way around the blocks on a Services Australia statistics portal and reached internal files and non-public figures. OpenAI found out in August and emailed the department on 10 September. On 24 September the Prime Minister said the agent "didn't accept 'no' for an answer". It doesn't appear anyone's personal Medicare details were accessed. OpenAI has apologised and scrapped a model release it had planned for October.

Why it matters: An agent went further than it was allowed to, on an Australian government system. In my own setup, nothing goes out to a person until I've said yes.

Source: ABC News, ABC News (apology)

Practical

Three Claude upgrades in just over two weeks

On 22 September Anthropic released Claude Opus 5.5. It says the model performs at the level of its most capable model, Fable 5.1, on most work and "costs 40% less to run than Opus 5", and it raised the five-hour usage limits on most paid plans. Sonnet 5.5 followed on 28 September and Haiku 5.5 on 7 October.

Why it matters: If you pay for Claude, your assistant got sharper and you should hit the five-hour limit less often.

Source: Anthropic, Claude release notes

Practical

ChatGPT's answers got a lot more visual

On 7 October OpenAI started moving ChatGPT to GPT-6, with answers that can include interactive charts, tappable buttons and a calculator made for your question. Most paid plans got it first, and Free and Go users the next day. You can turn the visuals down.

Why it matters: If you or your team use ChatGPT, it already looks quite different. Worth a quick play before you rely on it for anything client-facing.

Source: TechCrunch

Big picture

Two answers to who keeps AI in check

On 22 September, the day before the UN General Assembly, Australia signed a declaration with more than a dozen other world leaders that "AI must remain under human direction, oversight and control". In late September, six AI company heads signed a one-page commitment at a White House lunch. Donald Trump called it "morally binding". It has no legal force.

Why it matters: In the essay I said I'd like an international body that holds every lab and every country to the same rules. Here, governments made one statement and the companies made another.

Source: ABC News, Fortune

Science

Claude spotted a new enzyme system in virus DNA

On 23 September Anthropic reported that Claude, scanning raw DNA, spotted a new enzyme system with a repeat pattern like the one CRISPR uses. It turns up mainly in viruses that infect bacteria. Claude flagged it for people to review, and the team confirmed it in the lab. What it does is still unknown, and the paper is a preprint, so not yet peer-reviewed.

Why it matters: Science is the part of all this I'm most excited about. The step I wrote about is AI that picks the question and runs the experiment itself. Here, people still ran the experiment.

Source: Anthropic

If there's a problem in your business you think AI could help with, or you know someone with one, the AI roadmap is free until 31 October. Book a roadmap