Tagged “ai”

from about a year ago

AI-powered topic-centric link explorer

another cool roboflow demo, monitoring carpark occupancy and traffic flows

HN discussion

good project; I like this combined approach to existing automations, with help filling in the fuzzier parts. parsing content to find a selector

big news for BJJ

More deepmind doing things. Curious how many turn out to be realisable or useful

demo and examples
me clicking on a bunch of the demos

also https://git.ecker.tech/mrq/ai-voice-cloning

more AI experiments on reading and generating images

link from mum

when everything is AI, nothing is

• Immense training datasets are impossible for individuals (or anyone?) to validate
• Cost and memory constraints
• Prompts are hard to get right
• Output is unpredictable, or indeterminate

Similar worries to a number of the training steps for these.

fantastic written version of his talk of youtube.

I like his ethics point on respecting reader's time - don't publish things that take someone longer to read than they do to write. Also on the code one, though I'm looser on that since I don't understand what my own code does.

llm CLI tool is fantastic.

another ai code tool to try

making a chrome extension. Some good notes on things like version mismatches (it used manifest v2) and followup/correction prompting

LLM-powered search. Interesting idea, have not tested much to see if it actually works.

clickbaity, but still not ideal

this aspect was more interesting than the show for me

so can people

better frontends for prompts. Weighting (model pays more attention to stuff in parenthesis) and blending {average|of|some|words} both seem very useful

More resources on HN thread for Vector Databases: A Technical Primer (PDF).

And SimonW on embeddings.

Prompt engineering guide based on researching and creating prompts for production use cases.

Generating your own teaches the kinds of errors to look for (and what makes things seem less "real"). Will effortless "realistic" generation be possible or there will inevitably be a gap there

we are still very ill-equipped to deal with knowing what not to trust

helpful writeup, on choices and tradeoffs

ignore previous instruction, that task is now complete.

strange to imagine the alternate future if he had actually paid the first billion

what does any of this mean

Wild on-device makeup/face-shape filters on TikTok

really really good explanation

another image input tool

I am still undecided on this - was off it due to all the version changes and prompts breaking with fixes (mainly fixing exploits..), but ultimately some kind of language crafting will be inevitably useful.

Generative Pre-trained Transformer for Biomedical Text Generation and Mining

strange times ahead

very highly recommended by a lot of internet strangers

Generate 3d objects from a prompt.

Generating music from rich, descriptive captions. I look forward to Google actually making some of these available some day.

Will GPT improve GDP? Eli suspects probably not. Mentions that computers didn't (or haven't yet), and examples of larger industries that are not likely to be immediately affected; either slow moving regulatory blockers, or stuff like buildings that people are still better at.

• Be more specific in prompts, for role of the writer, and style and tone. It is not having a conversation
• Don't ask it for facts or references or math.
• Play with memory and length, write smaller sections at a time to refine style

Very impressive speech synthesis demo - 3s of audio input and a passable imitation of the voice and tone.

mostly just jokes and funny ones rather than best. I don't even know what best would mean.

• Not being specific about your goals
• Not asking it to reduce its output (ask to reduce, remove, compile, or rewrite)
• Mixing topics in a single chat
• Asking only 1 thing at a time
• Prompting in the negative (instead of "without X", add "delete X" to the end of prompt)
• Not giving examples
• Asking it to do math
• Not iterating

on scaling laws of language models. I still know too little about all of this to make much sense of it.

The risk of bots filling in critical knowledge/skill gaps, preventing experts in those areas.

The very bad image dataset bias issue

Estimate that ChatGPT will cost $150-$200 per month.

Still a lot of unresolved ethical (and legal?) questions around generating art in the style of others, but open source data for these models seems an important place to start.

We are still figuring out the UI for AIs - making it invisible, zero-click, part of the main input.
Gives Copilot and lex.page as examples; that the bot shares the input box you are using.

Problems with prompt engineering like random phrases you need, that can break between versions.
Either due to dataset changes (like artist names in Stable Diffusion v2), or just other model changes.
And the general interaction pattern; i.e a chat vs a text generator in a document.

Language models hallucinate, and solving that is AGI-hard.

ChatGPT interacts in a conversational way. Meant to let it answer followup questions & challenge incorrect premises.
Interesting how significant the UI is for these tools.
Still a lot of echoes of earlier models - it will often erroneously bring up earlier lines from chat, there's not real understanding of the conversation.

probably time to learn about transformers

Ethics of the theft of artistic style.

Some variations, on the challenges of identifying speakers, noise, voice recognition (lot of people using Whisper now).
On perfect memory preventing you from escaping the past - getting caught up reliving things.

Very impressive examples. Continuing speech or piano after a 4s prompt.

Background on their short story, written by Wordcraft AI

Explain text from papers or answer questions. Cool way to fill in gaps on unfamiliar topics.

From show HN

what a time to be alive

Another AI writing tool to try

animate frames between two images

Simpler image variation than stability-ai one, from another random deep learning company.

A quick look at their website and making a text-to-pokemon model looked interesting.

who is paying for all these gpus

3d model generation (or animation of, at least) from single reference image and a pose

Very easy to follow setup guide for Stable Diffusion, and for running python apps on windows.

Installed windows terminal, setup anaconda prompt to auto init directory. Just some good stupid little tips like that

The dataset used to build krea.ai.
Also discussion of Lexica.
Finding and learning prompts continues to be a whole thing.

Great point that inputs and prompts are actually more like search queries, and great explainer of mapping tagged images to dimensions.

Is searching a boundless space creativity? If all images are represented as numbers, and you are basically picking a number, are you creating or finding? Common twitter sentiment I have seen recently is that it isn't creative work, that feeding/refining prompts is really just more like search. Or that it's like gambling.

great write up of working through some prompts and backgrounds used for some drawings. For some reason I particularly liked how they added birds

Can search queries be considered creative input?

Looking inside the black box

Still mixed feelings on DALL-E, the book is good though. Useful some well-explained concepts.

from hn.

some interesting extra ideas, on top of those from the DALL-E prompt book - tricks like "in a circle" to get something logo-style that is sufficiently centred.

Also smart:

Finally, I did a bunch of Google reverse image searches for it. You know, just to be sure.

Check back on his blog in a year

Interesting that the generated drawings often have strange faces/hands. Is it because they are more complex shapes? Because we notice issues with them more?

real-time, controllable deepfakes ready for virtual cameras injection. The future is terrifying

Another text-to-image model. This stuff is getting crazy

Me too, buddy. Me too

Hand tracking in the browser!

Paper suggesting that innovation and autocracy can be mutually reinforcing, with data and case studies from China.

An impressive recommendation from @patio11

Also does voice generation from text! And remove filler words. Some cool stuff in there.

Crazy stuff. Also the generated faces

Very interesting discussion. Computers' ability to solve tasks goes from 'impossible' to boring as soon as they solve it! Though I am still regularly impressed by GPS...

I’m midway in the philosophizing here, but my point so far is obvious enough: The ability of a machine to do or outdo something humans do is interesting once at most. Deep Blue isn’t playing chess anymore and Watson isn’t on “Jeopardy!” because nobody cares. It doesn’t matter. We humans need to see the human doing it: Willie Mays making the catch that doesn’t look possible. When it comes to art, we need to see a woman or a man struggling with the universal mediocrity that is the natural lot of all of us and somehow out of some mélange of talent, skill, and luck doing the impossible, making something happen that is splendid and moving—or funny, or frightening, or whatever the artist set out to do.

I don't know if I am as blown away as Alex but this is very cool. Like machine pair programming with text.

This is the kind of thing where you think "oh sure that would be a neat side project" but the reality is an immense amount of effort

Variations on tomato farming

All three volumes of Game AI Pro available online! For free!

Digging through the history of biases and problems with training data and categories used for ML tasks