21 comments

[ 3.9 ms ] story [ 54.3 ms ] thread
Clickbait. It gets me really angry when tomshardware is doing this as well.
tomshardware hasn't published quality content since the early 2000's.
for me it just hijacked the browser history or something and I had to go back twice after reading the article to return to HN, doesn't seem to always happen. Anyone else notice? Such a scammy behaviour.
I think this might be a problem with your browser. When going backwards, browser should skip pages that redirect user to other page in less than 5 seconds.
Which spec is that on? Or you mean "should" in the "it would be convenient if your user agent worked like mine in this case" sense?
That's how it should work in my opinion; sadly as far as I am aware no browser implements this.
really? I think using the browser history API a page can push to the history stack on scroll events for example.
This is why I always middle click article links.
What’s the best source for hardware news now?
Yeah, pretty much clickbait... an "almost" is missing there, and "Generating Windows 95 keys" could be replaced by "simple arithmetic", because generating a valid key needs just simple arithmetic (which ChatGPT couldn't do in this instance).

Although the bit at the end where Chatty denies it did it because it'd be against policy is interesting, it means there were some safeguards implemented to prevent Chatty from giving out pirated keys from its crawls.

I've been a developer for 20 years, like many other here. To me, it's unconceivable that this prompt results in correct output. It's just so magical to me. I think because I think in conventional software development patterns? How would I even start to understand this? Am I old now?

Also, this is having an interesting result with my clientbase. I have one client who is using ChatGPT to 'help' write code for certain parts of the business logic. This has created an interesting workflow where we have created the boundries of the application and they can use ChatGPT to inject process businesslogic themselves and with that the responsibility of it's performance and functioning.

Another interesting effect is that clients now have proof that computer programs can indeed do all the magic things that I've, for ever, been telling them were not really possible (within a reasonable budget)

The data ChatGPT was trained on contained real keys. These models are extremely good at memorizing information and encoding it in the network.
I don't think thats true since you can prompt it a completely novel pattern in the same way the W95 key was described and get correct output.

Except it has some limitation in calculating sums and divisibility, however you can even get around that by telling it the steps more explicitly, like you don't say "must be divisible by 7 with no remainder", instead you tell it how you generate numbers that are step by step so its "thought process" is part of the output. Its like you write a program step by step in natural language.

basically if google can answer something, so can chatgpt… but it will also kinda summarize the results instead of sending you to the page.
I’ve had it a couple of times where Google can’t answer something where ChatGPT answers it.

Not sure if it was correct but it sure tried.

It's a LLM. You can't really expect anything else from it than answers that sound correct. Producing actually correct answers sometimes (which have a natural tendency to sound correct) is just a happy side effect.
> To me, it's unconceivable that this prompt results in correct output.

Then you'll be happy to learn that the article agrees:

While the keys passed a casual inspection, it turns out that only about 1-in-30 keys seem to work as expected.

Although that will probably change if/when ChatGPT gets better at basic math

More like you prompt chat gpt with the formula for generating a key. Then ask it to run said formula to generate a few keys.

Kind of clickbaity

** If you already know the algorithm and if you spend half an hour to steer chatGTP towards following that algorithm then it generates valid code but not with 100% accuracy. Then probably it would have been faster to ask it to generate code for the algorithm and then you execute it, since it's surprisingly better at coding than math.
I think thats spot on! Its similar how it uses plugins, because you can also describe a problem tell it to write code to solve the problem, then use a plugin to execute the code and then tell you the output.
"YouTuber uses ChatGPT to recite formula they provide back to them."

"YouTuber uses ChatGPT to perform basic algebra."

"Youtuber uses ChatGPT to do nothing."