blixt

14 ms·

151.
by blixt 2y ago
They're fair questions, and yeah it will have to disambiguate in such cases. I tried it now and it picked: - Springfield in Illinois - Sandwich in Massachusetts (the exact coords: https://www.google.com/maps/place&
152.
by blixt 2y ago
I do use a service for getting the weather (backend: https://github.com/blixt/sol-mate ) Yeah the second part I consider a bonus. Especially because I can use street names or slang for locations that no rigid API suppor
153.
by blixt 2y ago
Thank you for sharing – I love that the second one also has a microscope closeup of the display while it's changing. Happy hacking, hope to see your projects on here soon!
154.
by blixt 2y ago
If using the Raspberry Pi 5 I wouldn't go for a battery-powered solution because it's pretty power hungry. But you can run this screen just fine off of a Raspberry Pi Zero, or ideally a microcontroller if you want basically no pow
155.
by blixt 2y ago
My bad, I understand now! I think it's at least as good as a Kindle. I understand why someone who is comparing to bleached paper or OLED/modern LCD screens would say it looks purple, but yeah, I have no complaints. Edited to add:
156.
by blixt 2y ago
I appreciate that it's just a nice picture that happens to update with the weather, but there's no reason to only use this for your current physical location. It accepts any place in the world (and also fictional places), so you c
157.
by blixt 2y ago
I don't understand how I would get a unique but accurate prompt that makes Dall-E produce an image that besides the weather/lighting conditions also has a randomly picked iconic scene for the location, people in the picture that h
158.
by blixt 2y ago
I asked it to generate the weather for "The Purple Forest" (Narnia) and this is what it came up with: https://i.imgur.com/kmKiqbO.jpg You can zoom in on the purples. I would say it looks like a wine purple IRL.
159.
by blixt 2y ago
Nope, it's a completely physical process and state, so you can disconnect it and the image will remain the same. So you could use it as a reusable post card if you wanted, but for $80 maybe it's not such a good idea... I don'
160.
by blixt 2y ago
I love windows! And doors. And pictures that update in response to the weather.
161.
by blixt 2y ago
I'm just parroting Waveshare's own recommendations[1], but yeah probably it's not such a big deal. I'll link their docs in the README to help people make a more informed decision. [1]: https://www.waveshare.co
162.
by blixt 2y ago
I use Floyd-Steinberg dithering, and if you look at the "More images" section you can see it does a reasonable job at representing colors. Also worth nothing that 2 of those 7 colors are black & white, so questionable naming p
163.
by blixt 2y ago
Nice, super clever use of it! Yeah it's super fun and makes me want to put these screens all over my house.
164.
by blixt 2y ago
It uses AI for some things you can't really use free APIs for: - Turn a colloquial name for a neighborhood into the exact latitude and longitude (I'm sure there's some API that can kind of do this but I don't know a free
165.
by blixt 2y ago
The recommendation from Waveshare is "refresh at least once every 24 hours"[1]. It's probably not such a big deal in reality, but maybe there's some risk of the pixels getting stuck if they are active for too long? [1]:
166.
by blixt 2y ago
Since GPTs only have the one prompt I had to specialize it for a multi-step process with tool use (since it has to call an endpoint and draw an image in one go). Here's the prompt that I use in the GPT: https://gist.github.c
167.
by blixt 2y ago
I'm well aware of that, but there are plenty of ways to induce verbatim quoting from "hidden" information, and mostly verify it (through sampling a large number of times in separate runs). Models are improving in truly hiding
168.
by blixt 2y ago
Ah, you're right! My bad!
169.
by blixt 2y ago
I think that would have pretty serious implications for the transformer architecture though. If they're not embedded like text tokens, how would attention, etc work? And a conversation with multiple images back and forth? Not to mentio
170.
by blixt 2y ago
I would assume it has a "mode" token where it switches between text/image (and now audio), or you'd have to try to maximize the number of reserved tokens between multiple modes. GPT-4o did go from 100K to 200K vocabulary
171.
by blixt 2y ago
I did something similar when GPT-4V came out, partially with the goal to figure out the input format (I did not get anywhere other than "magic vectors"), but also to roughly estimate the amount of data you can get back out of a 51
172.
by blixt 2y ago
I went through a similar journey back when GPT-4V came out. Here's an additional puzzle for you: GPT-4V knows the exact pixel dimensions of the image (post-resize since there is a max size for images in the pipeline, besides 512x512)
173.
Show HN: E-Paper 7-color display showing the current weather(github.com)
229 points by blixt 2y ago|93 comments
174.
by blixt 2y ago
I guess a language model like Llama 3 could model surprise on a token-by-token basis and detect the areas that are most surprising, i.e. highest entropy. Because as one example mentioned, the entire alphabet may have high entropy in some re
175.
by blixt 2y ago
I guess you could take all lines in the file except the one you're testing and measure the filesize, then add the line and measure again. The delta should then be more fair. You could even do this by concatenating all code files and th
176.
by blixt 2y ago
I find these polarizing opinions confusing. There are a very large number of applications of generative AI. People believe a lot of them will produce a lot of value. So yeah, a lot of grifters abound. That part is not dissimilar to the cryp
177.
by blixt 2y ago
Really fun! I realized each delve carries the context of the previous ones. So I got to StarCraft II from the initial example of "Faster language models", but it mostly talked about how SC2 can be used for reinforcement learning.
178.
by blixt 2y ago
Reminds me of that time 10 years ago when I tried to make Go more like PHP instead of the other way around: https://github.com/blixt/go-pher More seriously, happy to see worker mode in this server. PHP does come with s
179.
by blixt 2y ago
It is possible but I wanted to maintain the full feature set of ChatGPT, plus it feels less shady to do it this way. What I did consider is just leaving the tab in the background so you can use it as a dumping ground for “side quests” while
180.
by blixt 2y ago
I use it for very random things while I browse the web. One example was ordering train tickets in France, I asked it to summarize all the options I was presented with.

More ›