kolinko

37 ms·

751.
by kolinko 3y ago
Feedback much appreciated!
752.
Show HN: a tabloid auto-generated from Hacker News
5 points by kolinko 3y ago|2 comments
753.
by kolinko 3y ago
If I'm not mistaken, it recognises voices already? At least once I had it tell me it didn't recognize my voice and couldn't do a thing I requested.
754.
by kolinko 3y ago
Stack Overflow is a capitalist venture. So is HN by the way.
755.
by kolinko 3y ago
We would all be using Internet Explorer now, if not for the US Government. So using browsers as an argument against regulations is a bad analogy.
756.
by kolinko 3y ago
It's not how AI works though - there is no rearrangement, and I think even most artists learned that it works differently. At the initial days of GPT people thought it actually copies and pastes content from the internet, but by now mo
757.
by kolinko 3y ago
That's impossible. There are multiple accounts and tests that show you should be getting >10h with just video meetings.
758.
by kolinko 3y ago
How? That's unusual unless you play games or do heavy computations - but then a regular laptop would last you 1-2h max.
759.
by kolinko 3y ago
"Radio" would be a proper replacement for "wireless".
760.
by kolinko 4y ago
Did you try building a SAAS service? I built a successful one, and sure - it gave me a decent income for 10 years after the initial 1 year of working on it. But altogether it was around €0.5M over it's lifetime, and it took me 5-10 yea
761.
by kolinko 4y ago
Variable length markov chains would merge some states, sure, but it will still be a similar order of magnitudes. Anything longer than 4 tokens/words of context and you bump into 30k ^ 4-10 -> cross a billion/trillion state boun
762.
by kolinko 4y ago
Unless you can steal an actual university's website, fabricating university website won't give you much - a good agent can have a list of sources of truth and proceed from there. Sure, some sources of truth can be compromised, but
763.
by kolinko 4y ago
Not op, but I can share my approach - I went line by line by Recmo's Cria: https://github.com/recmo/cria - which is an implementation of Llama in Numpy - so very low level. Took me I think 3-4 days x 10 hours + 1-
764.
by kolinko 4y ago
You're using billions and trillions loosely here. The hidden state in HMM would be (num_tokens ^ context size), so something like 60000^2000.
765.
by kolinko 4y ago
Kind of, but not exactly. In the end it's content that matters, not the form that it's send to - I don't care if it's my grandma sending me offers for viagra pills, or a spammer, and I don't care about the language
766.
by kolinko 4y ago
If you can use it for research/purpose purposes, you just do, and if you manage to build something good enough, you can get funding to train your own model then. Also, with business there are few "can do / can't do"
767.
by kolinko 4y ago
but what for?
768.
by kolinko 4y ago
How much spam do you get a day? I have an account since the beginning of gmail, and I get around 120 spam messages a day (roughly one every 10 minutes) Also, with self-hosted spam filters, I had issue with false positives - that is my filte
769.
by kolinko 4y ago
This. Google has the best spam filters so far. I wonder though, with the advent of new LLM models, it should now be trivially possible to build a zero-shot spam-filtering bot that is self hosted.
770.
by kolinko 4y ago
I never use Booking. Once got a terrible (and overpriced) location that was highly rated according to them, but above all - I more than once arrived only to realise my booking wasn't there. On top of that, I hate their UX. My first go-
771.
by kolinko 4y ago
When I was 15 years old and my sister was 4 or 5 years old, I wrote a pascal program that showed a big letter on a screen when my sister pressed such letter on a keyboard. She had fun pressing letters on a keyboard and then seeing them on t
772.
by kolinko 4y ago
It really depends on a car. Camaro has possibly the best knobs for changing stations and AC I've ever seen. The AC ones are integrated with two air outlets in the center - they are big, distinguishable and I could find them blindfolded
773.
by kolinko 4y ago
there are no developers sitting on a beach drinking pina coladas
774.
by kolinko 4y ago
I may be wrong, but I think EU already has a soft ban on private entities doing facial recognition in public. Even if it hadn’t, if a government entity is banned from doing something, it doesn’t matter if it’s hiring employees or contractor
775.
by kolinko 4y ago
A system with a bad architecture is way worse than a system with "no" architecture.
776.
by kolinko 4y ago
I am wondering about this as well - wondering how difficult it would be to build an analog circuit for a small LLM (7B?). And wondering if anyone's working on that yet. Seems like an obvious avenue to huge efficiency gains.
777.
by kolinko 4y ago
That's not true. You can buy Raspberry PI, which is 10x cheaper and 10x more powerful than the computers at the beginning of 2000s. Ditto with mobile phones. iPhone may be more expensive than when it launched, but you can buy dirt-chea
778.
by kolinko 4y ago
So far, with technologies, it's been that new tech is both cheaper and better than the previous one. To not look far - gpt3.5 turbo.
779.
by kolinko 4y ago
This tech can more or less improve productivity in every sector of the economy, and that includes energy/climate change. Personally, I've used it with success with R&D and due dil of projects related to climate change. LLMs (a
780.
by kolinko 4y ago
Taxes in the bay area can be insane - ~40% if I remember correctly. On top of that you have crazy-expensive healthcare, and crazy expensive housing costs. ~100k€ in (western) Europe may be comparable to ~200k€ in Bay Area.

More ›