I dont understand, what exactly does that mean?
It's the "neutered about the JQ" but technologically speaking.
The real one was too dangerous to be released so they had to cap it so it doesn't wipe out humanity man. Are you dumb or something?
it seems like an attempt to hype their company while at the same time essentially announcing another round of enshittification of their service
>>109017780it's nuspeak for cucked.
>>109017780They removed its ability to say the N word
>>109017780why don't you ask them instead of fishing for free replies in this type of mongolian basket weaving forum, stupid
>>109017839enshittification is obsoletewe are now in the era of ensloppification
>>109017780safe and effective
>>109017780it presents a subpar solution and then asks you to fuck off if you dont like it.
>>109017920>bro just ask a corporation, you wont receive a half ass PR dogshit reply at bestcalm down unc
>>109017780>what exactly does that mean?They gimped the supposed bug detection so you can't go around finding 0-day bugs to exploit and earn them moneyrinos.
>>109017780lobotomized
>>109017839It's kinda impressive how they pull this off so many times like it's nothing.Hype the new thing right before lobotomizing the previous, make it more expensive. Rinse and repeat. Crazy.
>>109017809Truly, the GOAT excuse for why your latest model isn't very good.
>>109017780>what exactly does that mean?that all the hype was fake and gay and this is their excuse for why you shouldn't question it
>>109017780PR bullshit, their model is dogshit so they lie on the internet by saying it's nerfed or something to preserve humanity or something
>>109017780nothingit means they didnt have the infrastructure to sell mythos at scale so they only sold it to a few companies and now theyre selling it to everyone
I just wonder what they could possibly be doing to """improve""" these word prediction machinesare they making real changes to the underlying architecture through research breakthroughsor are they targetting specific areas and improving it's benchmark performance in those without actually generalizing the intelligence
>>109018816Efficiency improves but it is eaten up by making the model larger immediately.
>>109018816I assume they're just cleaning up the training data by tagging it better, while also adding more up-to-date sources to the set.The architecture of LLMs isn't changed much, instead it's all about the data.
>>109018816They do improve them anon.Some models out there act like sponges that have all the data AI company has compressed, ie model was trained on everything.Opus used to be like that. Remember when it sucked ass, refused to follow prompts and sonnet was literally better at coding? That is because sonnet was a smol specialized model, distilled from Opus. No junk, only trained on synthetic data which was likely Opus output, responses generated to specific prompts designed to extract certain data. It is important to have a model like that. The more different shit you put in there, the better in "understands" the world in general, beyound just coding. It's knowledge from different domains and such.That is how it all worked in the past, before AI companies hit the data cap. Nobody had anything new to sell them, all knowledge of the internet that money could buy was already compressed in their largest models, but it was not enough.So my best guess it that they are currently relying on whatever people generate. Because your prompt technically is a new data, response to such a promps also is something that can be included in the training dataset. It is not ideal, but the trick is that they use very high quality promps from professionals, like those Cursor IDE users who spends large amounts of money on frontier LLMs and they provide feedback as well, when they accept or revert changes.tl;drAI companies improve their models by using your data>>109018872Correct, but they don't really have much up-to-date sources that are not contaminated with cheap AI slop. Take a look at places like stackoverflow LOL
>>109017789This. I knew they would not let use have a functional AI. Claude was too good. RIP. Where do we go now?
>>109017780It's just marketing
>>109017780They're saying it's lobotomized, that's why it's shit.As opposed to it being shit to begin with.
>>109018907>they don't really have much up-to-date sources that are not contaminated with cheap AI slop.Slop or not, the models still constantly need to be updated with the latest news, research papers, movie plots and retarded TikTok trends.
>>109018989They literally don't do it. Pick any newest model from any vendor, it would be a year or two behind everything that is going on in the world.And this news and trends shit is all synthetic. They do not allow feeding their models with cheap slop. They get their best model to digest this info and output summaries, those are going to be in the newest datasets, but not the raw data in any shape or form, since new data is cheap slop. Training on cheap slop = making your best model act like a little bit more like cheap slop. Depending on how much you do it. And there is a lot of this noise out there. It is not really worth anything. Scrape one news website or 100 such websites. Same thing, there is no value, so they definitely distill such information.
>>109017780>>109017780previewing it now. this thing is scary good
>>109017780They're all running the "our AI is so powerful it's dangerous" gambit. See the kerfuffle over the Linux kernel exploits revealed a month ago. It's a hail mary to keep the capital flowing and regulations loose. Obama was a wise and skilled leader btw.
>>109018084and they bled money all the way
>>109017780it means they only train them on reddit aproved opinions, just use a chinese model, they are less censored
Small models getting too powerful means less people need to pay the monthly subs for the frontier access.They can't allow that.
>>109019735ah, the burning dumpster strategy
>>109017780They run multiple layers of "alignment" checks as they call it.Another LLM is checking all input, and again all output.The input LLM can also reroute your request to a dumber model under certain circumstances. This reroute can be silent, and you don't know if it happened.
>>109018989Literally only Grok knows what's going on in the world which makes it the most useful one for when you're not coding
>>109017789fpbp
>>109017920Looks like his fishing techniques are very well working.
>>109017780Pro semitic
>>109017780You can't make a nuke with it or some shit idk.
>>109017780Translation: they put a bunch of guard rails against distillations that still won't do shit.
>>109017780if you want it to look for security holes in US water and electric infrastructure you've got to first tell it your grandmother used to explain such security holes to help you fall asleep
>>109017780If means silently falling back to a dumber model if it detects if you make any (((unsafe))) requests.
