make3

27 ms·

631.
by make3 4y ago
chinchilla only says that it would be more computationally efficient to have more data than to make the model larger, not that making the model larger wouldn't still benefit in performance gain
632.
by make3 4y ago
only if the expansion is really equal everywhere I guess
633.
by make3 4y ago
I mean that's just space
634.
by make3 4y ago
The title is nonsensical. The faster the compute is, or the faster inference is (through eg precision), the larger the models people will train, because accuracy / output quality increases indefinitely with model size, and everyone kno
635.
by make3 4y ago
how can that work if nothing is faster than light in the referential of the objects being compared? is it that space expands faster than light moves?
636.
by make3 4y ago
small players will never have a chance to push the state of the art, as whatever optimization there is will also be applied at large scale with more money
637.
by make3 4y ago
this is sickening. makes me not want to move to the US. that, and mass killings without gun law changes, & health insurance letting people die or bankrupt, even if I wouldn't have that problem myself
638.
by make3 4y ago
This is some old school microsoft shit, make everyone expand on your products & be dependant on them thinking they're safe by publishing an open license, then swap it under their feet forcing everyone to pay 25%. That's like i
639.
by make3 4y ago
name should probably change so you don't get sued. am not a lawyer though. (cool project btw)
640.
by make3 4y ago
freedom of religion in a mostly Christian nation is a start imho, progressive in the acceptation of differences. even if here it goes too far
641.
by make3 4y ago
progressive in protecting everyone's freedom of religion, which is a good idea in theory but the application here is over zealous
642.
by make3 4y ago
what?! I just assume that people tend to defend their own religion more intensely, whatever it is. & I think that protecting everyone's right to religion is a good thing, just that here it went too far
643.
by make3 4y ago
I'm curious if the direction were themselves Muslim, as this was a pretty brash decision, or if it was out of misguided progressiveness, thinking they were protecting freedom of religion when they're instead breaking freedom of ex
644.
by make3 4y ago
the step from gpt 2 to gpt3 showed us that any attempt at predicting the behavior of sufficiently scaled up models is really futile
645.
by make3 4y ago
yeah this is 100% survivor bias; their success only means you can make money with already hugely successful & popular projects, it's not indicative that your project will be successful or that you're doing the right thing
646.
by make3 4y ago
that's another thing, some people become so famous that having a paper with them is a big mark of prestige (erdos number), and people start to try to add them to papers they only marginally looked in the general direction of at best
647.
by make3 4y ago
they don't do any of those tasks though, people who do 50 papers a year are in very senior roles & only mentor, attend a few meetings, read the paper once or twice at best and propose modifications maybe. the question of what const
648.
by make3 4y ago
insurance companies are super powerful in the US, and some states are pro-business to some insane level. things such as this and businesses similar to 23andme really scare me. you need just one state to make it legal / one fucked CEO,
649.
by make3 4y ago
As a NLP researcher at a well known place & working on this, I totally 100% agree with you, this is ultra dangerous and should not be used, straight up, on anything remotely serious
650.
by make3 4y ago
I'm sure it would work with a LLM if you give it a few examples
651.
by make3 4y ago
vscode, the main competition to this I assume, is very fast to me. would be curious to know how they compare
652.
by make3 4y ago
very cool application
653.
by make3 4y ago
the thing is that it's been shown times and times again (with chatGPT for example) that you can get really pretty good results by giving massive amounts of final results to the model. This approach is better by far than anything we&#x2
654.
by make3 4y ago
too bad it's not compatible with huggingface tokenizer configs and need its own
655.
by make3 4y ago
yeah they're trying to implement content filters and it's not doing a great job I think
656.
by make3 4y ago
wow, a self driving car that will dispute fines for you
657.
by make3 4y ago
sure. my point was that transformers are massively better than whatever they had in 1985
658.
by make3 4y ago
yeah satellites potentially have gigantic cameras arrays and sensors of all kinds
659.
by make3 4y ago
GPT-2 is really by far massively stronger than anything in 1985. I suggest that you try using https://chat.openai.com/chat
660.
by make3 4y ago
I don't see why they would ever package GPT2 (the bigger model) in the browser. Speech to text has higher chances though, that's an interesting idea, as they already package text to speech too.

More ›