Forgot your password?
typodupeerror

Comment Re:If true... (Score 1) 103

Actually it does not show any of that. Assuming there is no bug in Lean (which may or may not be true, I tend to think Lean is probably ok), there is a huge step in there that Lean does not deliver: The translation from the real world to Lean. That is done by the LLM in this case and LLMs have messed this up for mathematics time and again and proved something else, put in invalid assumptions as base facts and other problems. With this proof it seems the LLM pulled in a lot of existing mathematics and every piece of this has to be formalized correctly. If it just got one existing thing wrong, the whole proof may fall. Now every one of these formalizations needs to be checked and that can be very tricky. I have some small experience in this area and I know that checking or doing formalization is very hard. Checking an existing proof is easy in comparison. In a sense, Lean does very good syntax checking, but it does not understand the content of the text.

Hence it may be years or even decades before we know whether the LLM proof is good. If enough mathematicians even care. Many seem to think the LLM pushers are stealing their lunch without contribution much value and I agree. Many mathematicians have no motivation to invest into helping them.

Comment Re:Or not at all (Score 1) 23

Indeed. I talked to a researcher in this area around 1990. Behind closed doors, nobody expected to ever get an actually useful computation mechanism even back then. The scaling is just way too bad and it is bad (much worse than linear) in two dimensions. The reason I was given for the research to be done was "the mathematics is interesting". I can accept that, but keeping the public lie alive this long? That is just wrong.

Comment Great, more bullshit (Score 1) 23

For all practical purposes. QCs do not exist. The only thing they can do is simulate themselves. Oh, and factor, what was it, 21 or something in 2012. Nobody has been able to improve on that without faking it and that should tell you nicely how true all these claims made are.

This scam has now run more than long enough. Whether these things will ever become useful is completely unknown. From their abysmally bad scaling (inverse exponential both in number of qbits and length of the calculation seems likely, but we may never get QCs large enough to actually find out), they never will scale to anything real.

Comment Re:If true... (Score 1) 103

Thank you. Good to see some people still can think for themselves.

Judging by historical standards, a return to secrecy is strongly on the cards.

Yes. And that is bad. But no mathematical research anymore is worse. Better have even slower progress than none at all after a brief period of straw-fires. I also wonder whether the RH proof may have leaked and gotten stolen. That we only have a basically unreadable Lean version of the LLM proof is highly suspicious. Time may tell.

As to programming, code security needs will kill slop for anything in production use in the next few years. LLM code is insecure and even frontier models do not find some of the most laughably obvious vulnerabilities. But that does not mean attackers cannot find them, far from it. We are also going into an age of regulation, liability (including personal) and qualification requirements for software. Slop will become a huge financial and personal risk and reviewing slop is really something nobody wants to do. It seems to be conductive to burnout and depression.

Comment Re:It's several things at once (Score 1) 19

And ongoing development gets more difficult as the code base degrades into slop.

Which may well be an existential risk. Because at some point you cannot fix things anymore and you may find out only when they get attacked or start to break all the time. This is why good software engineering is so critically important.

Slashdot Top Deals

Anyone who imagines that all fruits ripen at the same time as the strawberries, knows nothing about grapes. -- Philippus Paracelsus

Working...