dchest

15 ms·

181.
by dchest 3y ago
https://news.ycombinator.com/item?id=38858012
182.
by dchest 3y ago
IANAL, but you misunderstood the law. It only applies if the boycotting company terminates contract with a party in Iran in order to comply with US sanctions. If the company doesn't want to have business with Iran voluntary for other r
183.
by dchest 3y ago
From the linked article: "Last December, we brought Gemini Pro into Bard in English..."
184.
Rust's Problem with Slow Compiles(thenewstack.io)
2 points by dchest 3y ago|0 comments
185.
Kinopio's Design Principles(pketh.org)
1 points by dchest 3y ago|0 comments
186.
by dchest 3y ago
Their tokenizer is open source: https://github.com/openai/tiktoken Data files that contain vocabulary are listed here: https://github.com/openai/tiktoken/blob/9e79899bc248d5313c7d...
187.
by dchest 3y ago
it was a joke. i'm fascinated by his approach. shakti.com has a suspiciously familiar looking snippet of javascript was it him?
188.
by dchest 3y ago
Pity that he writes in C; if he wrote in JavaScript, he wouldn't need a minifier.
189.
by dchest 3y ago
"I think the fastest I got the FreeBSD kernel booting in Firecracker was 21 ms. NetBSD is now at 18 ms... I need to go back and address some more of the issues I noticed but didn't get around to fixing. Anyone know what the curre
190.
by dchest 3y ago
The book and software is under the MIT-like license https://people.inf.ethz.ch/wirth/ProjectOberon/license.txt
191.
by dchest 3y ago
Not only Americans, xerox is a word for copier in other languages. In Mongolia, it's canon instead. https://en.wikipedia.org/wiki/List_of_generic_and_genericize...
192.
by dchest 3y ago
For calendars and contacts you can try Radicale if you don't need web access: https://radicale.org/v3.html
193.
by dchest 3y ago
My theory is that these designers use very large displays with very high resolutions and put them meters away from their eyes. Although, I'm not sure how they can read text on UI elements at the same time.
194.
by dchest 3y ago
Yes.
195.
by dchest 3y ago
You can't buy YouTube Premium everywhere.
196.
by dchest 3y ago
It works great, but doesn't work against YouTube's anti-adblock yet.
197.
by dchest 3y ago
There's no "in theory" or "breaks the 100% hardware guarantee". Passkeys were specifically designed to be able to sync. There are also device-bound passkeys that can't be synced, if you want that. That is all.
198.
by dchest 3y ago
"Organizations of all sizes and across all industries are increasingly looking to accelerate digital transformation and power AI-driven innovation." How did Google become like this?
199.
by dchest 3y ago
Lit is made by Google.
200.
by dchest 3y ago
Lit is not a polyfill, it's an easier way to write Web Components without writing DOM manipulations manually. Building on top of the Web Components standards, Lit adds just what you need to be happy and productive: reactivity, declara
201.
by dchest 3y ago
Again? This happened some time ago with different search terms, and they fixed it server-side. I believe it was related to their autocompletion thing ("Siri suggested website").
202.
by dchest 3y ago
The lossless format (where this bug's in) achieves amazing compression rates, compared to PNG.
203.
by dchest 3y ago
To clarify: I meant that in general they are commonly applied, but in this case they weren't as the author confirmed. The repetition, of course, doesn't happen all the time.
204.
by dchest 3y ago
Yes, they are applied. Here's OpenAI doc which describes how to set various sampling parameters for GPT-3: https://platform.openai.com/docs/api-reference/completions/c... See presence_penalty and frequen
205.
by dchest 3y ago
There's no such thing as "base models have only the temperature setting". Models do not have sampling settings (temperature, repetition penalty, etc), the sampling code does, which obviously you can use on any model. For exam
206.
by dchest 3y ago
Try running llama.cpp with 0 temperature and without repetition penalty and you'll sometimes get caught in a loop. Or, if you're okay with a smaller LLM, go here, set temperature to zero and enjoy repetition: https://tr
207.
by dchest 3y ago
Even large language models with billions of parameters get caught in the cycle. You don't usually see it exposed to users because there are sampling tricks applied, such as repetition/frequency penalty.
208.
by dchest 3y ago
I loved it in OpenOffice Writer in early 2000s: type a couple of letters and press Enter to complete. Same with autocomplete when programming.
209.
by dchest 3y ago
I'd guess they already trained it on high quality (and highly censored) data. It's GPT-2-like architecture, but not GPT-2 weights. phi-1.5 is 1.3 BILLION, this one is a lot smaller at 34 million.
210.
by dchest 3y ago
It's powered by the same thing. They even mention "Even more accurate autocorrect" in iOS 17 on the product page.

More ›