Fuck AI

Exactly. It's best for problems that both:

1. Have a vast possible space of potential solutions.
2. Can be quickly checked for accuracy.
3. Has more than one right answer.

You want to come up with a new chemical to do a thing? Have an AI run through billions of permutations to find the one compound that can do that thing best. Then test that compound. See if it's more effective than current solutions. Maybe it missed a better compound somewhere in the universe of potential compounds. But if what it found is still better than the current solution, that's still progress and has real value to the world.

For an example, think of an AI built to search for high temperature superconductors. There's an endless variety of possible compounds that might lead to this. But the time to test any individual compound isn't that large. At a basic level you just need to synthesize it, confirm it's a superconductor, and measure its critical temperature. Simple stuff (apart from maybe the synthesis.) Testing one compound for superconductivity isn't hard. Trying to screen through trillions of candidate high temperature superconductors is. So you let the AI churn through the vast universe of possible options. Have it spit out the top ten candidates. Then you go to the lab, synthesize them, and try them out. If it hallucinates and is wrong? Very little cost beyond some lab time. If it misses one? Oh well, maybe we'll find it later. Say the current record for a superconductor is 151K. The AI predicts a compound with a transition temperature of 155K. Maybe it misses one that could do it at 157K. But oh well, a better high temperature superconductor is still a better high temperature superconductor!

This is where AI shines. When incremental progress is acceptable and when the solution is easily verifiable.

I’ve been impressed by the benefit derived from Carbon AI. It’s is a deep learning model trained on massive data sets and can identify 150mil weeds near instantly. When paired with a laser delivery system and pulled over a field, the weeds are id’d and precision killed with the laser, which spares the crop. This has great potential for reducing the use of commercial herbicides, which also means we can get away from the “RoundUp Ready” gmo seeds engineered to survive devastating chemical bombardment.

Currently the delivery system is tractor pulled, but I’d be interested to see if it could be made into a compact, low impact form. It could be a tool for wildland restoration crews to use to cleanly eradicate invasives while another crew seeds natives in their wake. Obviously not something that could be used in every terrain, but there’s plenty where you could. Even a consumer grade model could be a boon, instead of people dousing their property with weed killers that drain into local water supplies and nearby rivers, drive your mower-sized laser blaster across your property and zap your weeds.

It’s one of the few applications of AI where I’ve felt like its ability to recognize, process, and respond quickly is worth the application it’s used for. But also, it’s not generative. It’s trained on data collected and input by humans, it just reduces the effort (and side effects) of mass agriculture. Doesn’t eliminate ag workers, it makes an already difficult job easier and cleaner.

I do get your sentiment, and can see it applicable in these cases.. But I also think we should put a lot more focus on creating models specialised for these important tasks, instead of using these super large and heavy genAI models.

As it currently stands, the research seems to be largely focusing on just pushing the expensive limits of these models, and pushing the olligarch agendas, which is a bit unfortunate, since this field has potential to be super beneficial.

In most of your examples it's revealed after the fact that AI didn't discover anything it just defeated the test.