Hacker News
points by program_whiz 4 days ago | 0 comments
Arriving at truth values has limited value in science, obviously (with full understanding being more useful). For example, knowing a chemical reaction will produce heat is useful (cavemen discovered that one), but knowing _why_ unlocks an entire world of useful tools (the combustion engine, space heaters, etc).

However, even getting truth answers will soon be beyond AI. It is of course making impacts in Math and in publishing in "data heavy" subjects where you can just formally verify, program, and/or munge data.

But in "real science" where we actually need to build a lab, buy supplies, hire personnel, get approval, form physical test harnesses, install sensors in proper locations, calibrate everything, then finally make one measurement, question the whole thing, then finally make a conclusion, AI is of limited benefit.

Yes, it can help speed up coming up with some formulas, spitballing, or checking existing research. It can even help write the jank code for sensor calibration, measurement taking, etc. That's not nothing, but we are _very far_ from meaningful contributions to real experimental hard sciences.

As for math, theoretical physics, computer science, and other "logic based" studies, the question becomes one of benefit to cost of true understanding versus simply knowing something is true.

I'd argue that it paints a starkly negative picture of most science. Once answers are easy to obtain, it seems most people are concluding that actually understanding why its true isn't worth the effort, which means that the question maybe wasn't worth asking.

Its the science equivalent to the recent realization in tech: most ideas suck, your product didn't need more bells and whistles. Actually the engineers pushing back on every shower thought from product was a good thing. The friction that stops you from putting every random-ass feature into your already bloated useless SaaS product was good, and now the sales guy is pushing features to prod for a customer demo. We thought "too many ideas, not enough engineers", we now know "not enough meaningful quality ideas".

In the same way we thought "too many questions, not enough scientists", we now know "not enough real study on meaningful topics".

[–] podocarp 4 days ago
I'm gonna be annoying, but space heaters don't have chemicals
[–] program_whiz 4 days ago
I was referring to those big shop heaters that use forced-air propane, butane, or kerosene to generate massive heat. But yes, there are electric heaters too. It wasn't too annoying, just opinionated and misinformed.
[–] rowanG077 4 days ago
This misses that AI is already designing and carrying out experiments autonomously. This will only explode with the growth of capabilities in robotics.
[–] program_whiz 4 days ago
No I accounted for that. Sure, if you have everything setup for full automation, then yeah, you can write a computer program to execute many experiments in parallel, and that is no doubt useful. But at the point your question doesn't fit into a repeated process, the physical provisioning, sourcing, hiring, contracting, politics, space acquisition, calibration, layout, etc. all applies.

For many types of study, especially new ones, having such a setup is just as hard as doing the thing that may or may not require massive automation. For example, testing a new type of high acceleration engine. How would you even setup a lab to have hundreds of test beds for this? Would you need to manufacture hundreds of prototypes of the engine to even do the test? Where would you source that if the whole point is proving whether the engine works and produces more thrust. You probably just have one prototype sitting on a table. Sure when you get to the point of refining a tested and well-studied design, automation helps.

I'm sure there are some areas like organic chemistry where you can just toss compounds together and vary the heat, etc. But what are you getting? Don't you need to do specific tests and measurements to figure out if its useful? All of that requires setup with all the caveats above. What if you're doing something else, like metallics. Well then you need a whole lab with temperature and atmosphere controlled furnaces, various cooling surfaces, additives, dopants, etc. Knowing when / where to add, how to cool, rotate, etc. could all play a factor. Automating that right off the bat isn't feasible until you've proven it can actually work, so AI help is of limited value.

Now, if we have AGI with full embodiment, that's a different story (since with full embodiment it can handle all these tasks if given authority to do so).

[–] rowanG077 3 days ago
> or many types of study, especially new ones, having such a setup is just as hard as doing the thing that may or may not require massive automation. For example, testing a new type of high acceleration engine. How would you even setup a lab to have hundreds of test beds for this? Would you need to manufacture hundreds of prototypes of the engine to even do the test? Where would you source that if the whole point is proving whether the engine works and produces more thrust. You probably just have one prototype sitting on a table. Sure when you get to the point of refining a tested and well-studied design, automation helps.

This is the point where robotics can re-arrange their own labs.