Forgot your password?
typodupeerror

Comment Re:smells like executive decision making (Score 1) 45

Yeah and... The titles they didn't promote well enough sometimes did much better than expected (e.g Tsushima).

Instead of taking this into account and putting more effort into single player games that were actually making them more money than they expected, they chased the fleeting extractor shooter market (which anyone could tell them is severely limited because its player base have to be cool with the risk of losing everything every time they play... that's a pretty acquired taste eh) on the premise that they might catch the next fortnite.

It really seems like the folks making the decisions here don't understand the environment but assume it should conform to them rather than the other way around.

Comment smells like executive decision making (Score 1) 45

This sure does look at bit like someone in the global sony exec not understanding that Bungie's failure with Marathon was as much their fault as Bungie's and blaming the whole thing on PC gaming as a whole.

That and believing that they don't need to compete with the xbox any more.

Lets see how that goes for them, I guess.

Comment Re:Just a reminder they didn't invent Pokemon (Score 3, Insightful) 23

There were also two Megami Tensei games and the first Shin Megami Tensei released by the time Pokemon Red and Green were released in Japan in 1994.

They never should have had a patent on that, and they shouldn't be allowed to add 'with a stylus' to the end of it and suddenly patent something they couldn't before... just like they shouldn't be able to add 'on the internet'.

Comment Re:Translation (Score 1) 169

Can you explain where this exuberant confidence comes from?

My back of the napkin math that tries to make the current wave of LLM based businesses profitable always fails... This isn't true of all systems people refer to as AI, but so far every solution for OpenAI, Anthropic (and probably Microsoft too, maybe not Google, hard to say) seems to involve them putting their prices up at least 10 times (and possibly up to 50 times for OpenAI).

If you have a solution that isn't just a tautology about token prices, please feel free to explain it.

Comment Re: Isn't that what Wikipedia already is? (Score 1) 61

Pretty sure their message translates to "it doesn't support the glorified narratives I was taught in school".

I'm not even hating on them for it, it's so common that they just don't know that high school history is simplified to the point of making it seem like the famous people did all the work.

Comment Re:Cool (Score 1) 61

In case you're not aware... it's well established (with decades of data backing it up, as well as documented cases from modern systems... and if you're really keen exercises you can try yourself) that feeding the output of a language model back to itself as training data will cause it to stop appearing to mimic plausible human speech within a few generations. Generally it will get caught looping on single phrases and wont respond meaningfully, and then if you keep going with the training you'll find it stuck just repeating a single word but taking a ton of processing time to do it.

Poisoned data doesn't necessarily lead to more or less 'hallucinations'. It might instead prevent an LLM from even forming meaningful sentences in the first place... and there's no obvious tell for generated data that humans can actually use to filter this data out of a training set. It can probably be automated, but there's a cost to that (in terms of processing time) and it's necessarily going to be pretty high... perhaps to the point where it also makes the process cost ineffective... but the alternative is that any random data you ingest from the internet might actually make your model's performance worse, and it'd take you quite a bit of effort to work out which data was causing the problem.

Slashdot Top Deals

ASCII a stupid question, you get an EBCDIC answer.

Working...