Hacker News

points by mkarliner 1 day ago | 0 comments

"AI can copy, it cannot innovate."

https://blog.karliner.net/posts/chatgpt-and-the-velcro-test/

admittedly a bit old now, but...

3.5 years later, LLMs are solving multiple hard maths problems that humans couldn't solve, not even those with a million dollar prize available to whoever did.

Even if someone wants to deny this counts as "innovation", it sure isn't copying.

[–] jauco 12 hours ago

imagine the ai shops training the next generations of ai purely on the output of the preceding generations. No human input involved.

Would the ai improve exponentially or would it decay towards some base line?

I think you can say that they “create” if it’s the former and “reproduce” if it’s the latter.

[–] VCFundedGenYer 1 day ago

Gonna need to see some sources on that claim.

LLMs do not do math well and most of their calculations are quite false.

Copilot can’t even count to 20.

[–] stalfie 1 day ago

It's hard to overstate how out of date this opinion is now. I thought I would add some additional sources besides just the openAI article just to give you some idea of the scale of the impact. There are thousands of articles, blogs and other sources discussing the topic.

https://www.nytimes.com/2026/10/08/science/mathematicians-re... https://www.nature.com/articles/d41586-026-03389-1 https://scottaaronson.blog/?p=10169

[–] reverius42 1 day ago

https://openai.com/index/sharing-ai-progress-in-mathematics/

[–] saikia81 1 day ago

This comment reads as if it's a year old.

Also many of my calculations in my life have been quite false as well. I like to remember the problems I did solve. Instead of focusing on the few I couldn't solve.

[–] stalfie 1 day ago

It is funny how quickly opinions on this topic expire. It almost feels like witnessing an XKCD 1/10000 moment.

Apropos failing at calculating, I have a relevant anecdote that I can't help but wanting to share. A few years ago (2020) I got inadvertently handed a math problem I had to solve, and failed spectacularly at. I suck at math, had no one around me that didn't also suck, and the problem wasn't Googlable. The issue was creating a staff rotation for medical interns in a rural hospital, in an exceptional 6 month period where the previous symmetrical split of four interns to medicine and four to surgery was disrupted by an extra intern. Interns were on call for both departments on nightshifts and weekends, and since they all needed the same amount of total "on call" hours during a 6 month period, same amount of regular hours, and there were a number of legal and logistical constraints to how working hours could be distributed, the asymmetry turned out to be a real headache. Previously, getting the responsibility for setting up the next rotation was a coveted power amongst medical interns, as it just involved copy pasting new names into a rotational plan established decades ago by some forgotten elder, where you could abuse your privilege to slot yourself into the juiciest holiday periods. So no one realized they were handing over a logic bomb that would overload the atrophied math circuits of some poor schmuck fresh out of med school (me). Anyways I spent days Googling, setting up excel formulas, calling more mathematically qualified friends (who also failed), before finally giving up and spending 20 hours manually setting up a deeply flawed plan that ended up pissing numerous people off.

The other day, just for fun, I dug up my old emails with the original excel sheets and handed the problem over to Opus 5.5. It ran a quick python script that solved the problem in literally 3 seconds, and then seemingly just for fun ran a python solver that nuked my CPU for 10 minutes to find the proven optimal staff rotation. Then it sent me a winkyface emoji after pointing out the funny coincidence that me and my buddy (who was on the email chain) just happened to get Christmas and New years off.

It's a brave new world.

[–] boplicity 1 day ago

I'm not sure that's true; AI can (and does) sometimes brute force a variety of combinations, some of which are new and interesting.

It does beg the question, though: what is creativity?

[–] atleastoptimal 1 day ago

Obviously far out of date and not accurate. AI can solve math humans can't understand now.

So much of many ostensibly "critical" thinkers were deluded by axiomatic prognostications that AI would never be able to pass some certain threshold of capability, and then move the goalposts on what AI will "never" be able to do once those thresholds are surpassed.

How can the author of this blog so confidently claim that a currently existing invention could never have been invented by AI in a counterfactual universe? Such a claim is fundamentally unfalsifiable, but is the bread and butter of the larger industry of "AI capabilities denialism" which acts as a salve against anxiety over AI's genuine capabilities. However living in a false reality eventually catches up with you.

[–] stalfie 1 day ago

May 2023 is pretty much prehistory in AI time. No conclusions from that era about what AI cannot do are valid.

We're starting to approach four digits of LLM-made mathematical papers, cybersecurity being almost entirely taken over by AI redteaming, initial forays into physics, biotech and pretty much every field... and we're going to use some dude prompting GPT-4 about Velcro to form the basis of an opinion? Even back then, using the word "never" was intellectually questionable at best.

The mantra of AI not innovating is getting pretty tired at this point. It's simply wrong as a matter of fact. If the goalpost has shifted so much that solving millenium prize problems is not "innovating", then the goal is no longer on the same continent.