scarmig

32 ms·

451.
by scarmig 3y ago
Another angle to look at would be the S4 models, which admit both a continuous time and recurrent discrete representation.
452.
by scarmig 3y ago
Proving that two real numbers are the same is not decidable either. Despite that, there are some real numbers that we can prove are equal to each other (rarely).
453.
by scarmig 3y ago
Hosting sites will simply advertise AI-generated porn as real people. Or the videos will go through a chain of transfers that give the sites plausible deniability for the false advertising. AI-generated stuff will outcompete human-generated
454.
by scarmig 3y ago
I had the exact same experience, and I was also using it for a web application powered by the Mamba web framework.
455.
by scarmig 3y ago
> Generate a three sentence comment saying it's easy to use LLMs to write a comment in broken English. Use a style and diction suggesting the author is an English second language learner from subcontinental Asia. Using LLMs for maki
456.
by scarmig 3y ago
Fair enough; that was more directed at LLaMA and derivatives, which have commercial restrictions.
457.
by scarmig 3y ago
Some of us sleep and aren't to be arguing against responses left at 1AM their time :)
458.
by scarmig 3y ago
It would obviously be better if everything were done in the open. But you can't simply say "corporations will corrupt AI with copyright regime X" and compare it to a hypothetical universe "copyright regime Y will result
459.
by scarmig 3y ago
Who cares if it's a PR stunt to improve developer good will? It's still a good thing, and it's now the most open model out there.
460.
by scarmig 3y ago
Preventing or slowing the development of AI is just a startling contempt for human life. Every day millions of people die unnecessarily. Willingness to let that happen because you think your family (or, most likely, a giant corporation) des
461.
by scarmig 3y ago
The logic looks something like "The metrics are there to measure SWE productivity and quality, if we let them change it, they will game the metrics."
462.
by scarmig 3y ago
Very few are meaningful on a resume. The best I can think of for a SWE is RHCE. Regardless of if you get the cert/keep it updated, it's stuff you should know. Another respectable cert track is CCNA/CCIE. That is, of course, e
463.
by scarmig 3y ago
One interesting tidbit from the technical report: >HumanEval is an industry standard open-source evaluation benchmark (Chen et al., 2021), but we found controlling for accidental leakage on webpages and open-source code repositories to b
464.
by scarmig 3y ago
Just gotta make sure the captcha requires a >1M token context length to solve...
465.
by scarmig 3y ago
The technical report: https://storage.googleapis.com/deepmind-media/gemini/gemini_...
466.
by scarmig 3y ago
"The user is wrong" is not a way to justify bad statistics. Saying a violent assault doesn't count as a violent assault unless the victim is willing to spend god knows how many hours going through bureaucracy purely for the s
467.
by scarmig 3y ago
A single block in SF has more charm than the whole of Sunnyvale. Despite some of the toxic people who live here, there's a reason many people want to.
468.
by scarmig 3y ago
People burned up a Muni bus after the World Series, and no one stopped them. And it was certainly a collective thing, not a lone miscreant. Vehicles that are owned by large apersonal organizations seem like a common target in this type of t
469.
by scarmig 3y ago
Data point: after being physically attacked in the street by a (presumably) homeless man, the SF cops refused to take my report unless I went down to the station. I'd regard these statistics as, very generously, aspirational.
470.
by scarmig 3y ago
It's pretty simple: use both GPAs and standardized scores. To prevent rampant grade inflation, normalize each school's GPA (and for ones where the top-end indicates significantly less rigorous coursework than other schools',
471.
by scarmig 3y ago
Although ethically I agree, I suspect future AI will consider us more like ants than moral beings, regardless of what we do. And although some humans do give moral consideration to ants (I do!), it's far from a guarantee.
472.
by scarmig 3y ago
That makes me think of a good use case for AI: delete all the flowery spiel and giant photos and ads, and return the recipe I actually am looking for.
473.
by scarmig 3y ago
ML and Kalman filtering try to solve similar problems. Some theories even suggest that Kalman filtering (or a similar algorithm) provides a basis for neurobiological learning. See predictive coding (e.g. https://arxiv.org/pd
474.
by scarmig 3y ago
The goal isn't to suffocate them, but to draw them into a trap--they're attracted to CO2 because we exhale it. Over time, the hope is you trap and kill the entire population. I think it'd work for mild infestations, but for l
475.
by scarmig 3y ago
> The correct answer should have been: (f (h (g (h z)))) AKA ((1 + 5) * 2) + 5. Isn't that (h(g(h z)))? And, FWIW, at least in my test, Gemini gets that in its final answer, though it failed in the two other drafts: https:/&#x
476.
by scarmig 3y ago
Google Brain and Deepmind.
477.
by scarmig 3y ago
Section 6 (post training) and section 7 (responsible deployment).
478.
by scarmig 3y ago
Google also updated the Gemini technical report : https://storage.googleapis.com/deepmind-media/gemini/gemini_...
479.
by scarmig 3y ago
I went through a progression, where I was first accepted to Google (genuinely one of the most exciting moments of my life), and I gradually fell into a deeper and deeper cynicism as I saw all my fantasies about how the world should work eva
480.
by scarmig 3y ago
It all depends on the team, of course, but hell really isn't a good way to describe Google employment. Of the half dozen companies I've worked for, Google consistently offered better life quality than the others, even ignoring the

More ›