Courses, newsletter, and the AI stories I am tracking.
Open courses
Stories
A US government agency has now run more than 40 evaluations of frontier AI models, including ones the public has never seen.
The push into legal and finance is working out pretty well for Anthropic it seems.
Anthropic announced Project Glasswing and its Mythos model on 7 April. It took about Five weeks (11 May) for OpenAI to launch Daybreak, its own AI security stack.
The NZ government wants to make $2.4 billion in savings by cutting 8,700 public service roles by mid-2029. Because AI.
Anthropic shipped Claude Opus 4.8 this week, I reckon it was a rushed response to how crap 4.7 release was. But they are also fixing its most annoying feature.
New Zealand consented its first "AI factory" and given the track record of US data centres I was initially quite sceptical.
Google just deprecated its open-source Gemini CLI and is forcing all users over to its closed-source clone, Antigravity CLI. Oh boy, are people pissed.
The cheap gemini model billed nearly twice what the expensive one did for the same task.
Google just told us its AI Mode passed 1 billion monthly users, one year after launch.
Anthropic measured its own traffic and found that roughly 1 in 1,300 Claude conversations leaves users with a less accurate picture of reality.
Andon Labs gave four frontier AI models their own radio station, a bank account, $20 each, and one instruction: build a personality and turn a profit. As far as you know, you broadcast forever.
An OpenAI model just disproved a conjecture Paul Erdos posed in 1946, and a Fields Medalist verified the proof.
Most supply chain attacks need a vulnerability. This one just needed a credit card.
Oracle asked workers to document their workflows so it could train the company's AI. Then it started laying them off.
US utilities plan to spend $1.4 trillion through 2030 to keep up with AI data center demand. Households are projected to cover nearly half of it.
The UK's National Cyber Security Centre just told organisations to prepare for a "vulnerability patch wave."
Yet another 'company records staff computer use, trains an AI and lays off a bunch of people' story.
ANZ NZ, ASB, BNZ and Kiwibank all joined Visa's APAC Agentic Ready programme on 30 April.
China spends roughly $12.4 billion a year on AI. The US spends about $285 billion.
A fleet of Figure 03 robots just ran 200 hours straight, fully autonomous, zero hardware or system failures, and sorted 249,560 packages along the way. About one every 2.6 to 2.83 seconds, human speed, all onboard. No teleoperation, no cloud, no human in the loop.
On 13 May, ASB launched a free 12-week AI Bootcamp for New Zealand businesses. 4,000 SME seats, built with Xero and academyEX.
Bias and error abound in human systems, but when they get baked into AI it's easier to point the finger.
BNZ wrote NZ$352 million off its capitalised software balance. Their argument is that AI is making enterprise software obsolete faster than the old depreciation schedules assumed.
Starbucks spent two years pulling humans out of its stores and betting machines could cover the gap. It just admitted that bet failed.
I think most people are sleeping on how vulnerable LLMs are to prompt injection attacks.
