Two things about artificial intelligence kept getting bigger today: the amount of money going into it, and the amount of permission being handed to it. A third thing did not get bigger — the quality of the evidence that any of it works as advertised. In fact, the clearest new evidence pointed the other way. That gap, between how much we are betting and how well we can actually measure the payoff, was the shape of the day.

The money and the mandates

Start with the scale, because the numbers are hard to believe. In the first half of 2026, two companies — OpenAI and Anthropic — raised 43% of all startup funding on Earth, across every sector. Global venture funding hit a record $510 billion in six months, more than all of last year — and in the second quarter alone, more than 70% of it flowed to AI, up from just under half a year earlier. This is no longer a hot sector inside venture capital; it is venture capital, with everything else sharing the remainder.

Permission scaled up too. Nevada cleared Tesla, Waymo and Uber to put up to 7,000 robotaxis on the roads around Las Vegas — one of the most permissive driverless markets in the country. Though, tellingly, the permits are ceilings, and Tesla's own engineer admitted the company won't come close to filling them. Even the permission is running ahead of the reality.

The capability, and the asterisk

The tools kept getting cheaper and more open, which is the real good news of the day and worth saying plainly. Alibaba released a 27-billion-parameter model that runs on a single consumer GPU and claims coding scores near the paid frontier — downloadable, free, no datacentre required. The catch is in the word "claims": every one of those scores is the company's own, and that is the norm at launch, not the exception.

Which is exactly why the skill of the moment is reading a benchmark chart with a cold eye. We published a plain guide to doing that — who ran the test, on which exact version, could the model have seen it in training, and does the benchmark measure the thing you actually care about. A launch-day number is the start of a question, not the answer to one.

The evidence that cut the other way

And then the day's hardest fact. A 30-month study of 26,811 students found that AI homework help raised their homework scores 18% — and then cut their exam scores by 20%, with the damage to high-stakes exams building slowly over two years. It is an unusually large and long look at what happens when the tool does the work instead of the learner, and it found that finishing and learning quietly came apart.

That study is a useful lens on the whole day. The homework scores went up immediately; the real capability went down slowly, and only careful measurement over years caught the difference. Swap "homework" for "benchmark" and "exam" for "the job you actually needed done," and you have the general problem the money and the mandates are both racing ahead of.

The honest read

None of this is an argument against the technology, and the cheaper-and-more-open direction is real progress. It is an argument for the unglamorous discipline that the scale keeps outrunning: watch the process, not just the score; test the claim before you quote it; and treat a number that flatters the seller as a starting point, not a verdict. Record money and record permission are easy to measure. Whether the thing works is harder — and today, that was the number worth watching.