Satya Nadella of Microsoft agrees that we need a brake.
1 Like
The thing about AI is it is basically a prediction model, and nowhere near as intelligent as people like to think. Actually not really intelligent. One aspect that nobody mentions is that in statistics we try to predict, and one problem is out of data prediction. We obviously don’t have all possible data points, so we hope that everything follows our model, for example a linear relationship. AI are horribly nonlinear and they try to produce results for a far greater range of inputs than they are trained with. End result is they can and do produce strange results. I’ve seen it do that and someone had an example of asking an AI to plan her holiday in New Zealand and it created a bus route of 20 miles across a mountain range to replace 300 or 400 miles around. She didn’t realise until she tried to book the bus. There are supposed to be limitations on military equipment. There is an Israeli technology that is used on US tanks that intercepts incoming shells. It tracks them from the source but is more accurate close to the tank. It determines the firing location and sets up the firing parameters but requires approval from the tank commander. I expect if the attacker is close the commander just pushes the button.
1 Like
That’s a simplistic and slightly outdated view of how the big, well-known AI systems operate. Yes, the older models are almost strictly LLMs, which are pattern-matching prediction engines. But many of the newer models incorporate additional methods that go beyond mere statistical prediction and can include genuine reasoning models among other things.
Yes, they still get things wrong. They still get things wrong a lot. Which is why the brakes are needed. But one can no longer say that, say, ChatGPT is “just an LLM” and be accurate. There’s more under the hood now than just linguistic manipulation.
1 Like
They call them reasoning, but it is still based around prediction using neural networks. Reasoning is just a fancy way of saying we will make the result a series of steps and train the AI to reproduce those. Does it do better? Supposedly, and I guess that is because it is constrained at each step. Most of an AI is not an LLM. Correct. But the rest is mostly a neural net. The main problem with AI is that it doesn’t really understand anything. I love the following from Google. The AI doesn’t understand that Mayo Clinic is an organisation so when it misreads the sentence it produces a ridiculous output. It was fixed a few hours later.
2 Likes
I’ve said it before, but remember that Google’s AI Overviews is using the least capable models available because it has to scale to such incredible volume. It’s the one of the worst examples of AI in common use. Ask the same question of Claude or ChatGPT or Gemini (which uses different Google models), or even Siri AI, and you won’t see this sort of silly mistake.
There are plenty of other problems, certainly, but it’s very important not to underestimate what the AIs can do at this point in time and how fast they’re improving.
1 Like