Researchers at MIT have reported a frontier AI model directly editing the weights of a smaller model, a technique that points toward AI systems helping to improve other AI systems. The result is being read as an early data point in the long-running debate over recursive self-improvement, the idea that AI could speed up its own development in a compounding loop. Anthropic has also signalled that AI is playing a growing role in its own research, and the episode frames these developments as evidence that model improvement may be accelerating. The discussion also cautions against reading too much into a single experiment. Whether editing a smaller model's weights amounts to self-improvement at the frontier level remains an open question, and faster capability gains do not automatically translate into wider use. Cost is a second thread: rising running costs for GrokBot have revived arguments about whether AI services are priced below what they cost to run, and what happens when those subsidies end. The episode argues that adoption may depend less on model quality than on whether organisations can change how they work. For Australian businesses and professionals, nothing in this changes day-to-day obligations yet, and self-improving AI research is not a product you can buy or deploy. The practical lesson concerns cost: if AI tools are currently priced below their true cost, prices could rise, so it is worth budgeting AI spend at full rates rather than promotional ones. If you enter customer or staff personal information into AI tools, the Privacy Act 1988 and the Australian Privacy Principles still apply, so check where data is processed before adopting a new tool. The most useful near-term step is redesigning workflows rather than chasing each model release.