pstorm

The first is Millenial as a whole, the second is Millenial broken down by age. Essentially, the oldest Millenials have a ~65% homeownership rate and Millenials as a whole have ~54%.

That's not really how MoEs work. They never directly interact with eachother. There is one manager type model that takes a prompt, directs token inference to 1 or more models, chooses the best response, and continues. The analogy would be closer to a "swarm of agents". (There are a handful of names for this approach, I think swarm is catching on the most)

Couldn't you do the same with Kindle samples? Seems lower effort.

Took me a while to find the configuration settings you mentioned. For anyone else, you go to a specific TV model's page, and there is a tab called "Settings."

These are tame compared to some of the stuff I have been tasked with at a previous company. I don't want to go into specifics really, but I can tell you that it sort of became normal. Should we test out defaulting the newsletter checkbox? Sure, oh great, it improved newsletter conversions! Should we show a fake "X amount of people viewed this product"? Sure, if it works we can actually implement it! Etc. etc.

Until I was outside of the culture of the company for a bit, I didn't see any issue. Improvement of metrics was all that mattered.

Couldn't you just compare the similarity of the embeddings? I imagine that would work in the vast majority of cases and save a lot of LLM calls.

I tried it out for fun and it was surprisingly insightful. I can't see the prompt you used, but I appreciate the subtly fun personality as well.

Paste it into ChatGPT and try it out. I didn't quite understand it until I did that and it's pretty cool.

OpenAI has their own version: https://platform.openai.com/playground

Eh, for some people the sky is falling. For instance, the original hacker was specifically offering profiles of Ashkenazi Jew, which included full name and location information. I understand how that could be scary in the hands of the wrong people.

You are absolutely right. I have a built a handful of complex interconnected LLM calls, and the thing I am dying for is a simple way to inspect what is happening at each step. For instance, one chain has 5+ steps where data is fetched, transformed, sent to an LLM, and transformed again. I have a homegrown solution to seeing what's going on, but it is missing a lot of what I would want to see.

I've seen some attempts at providing RAG tracing as a service, but for my specific use case, where in some instances I am making 100+ LLM calls per chain, it isn't a fit quite yet.

The idea of the NPCs being autonomous and having all the capabilities of human players is really cool. It's similar to the few simulations that have popped up, but with humans and NPCs. I'll be following along!

Running the curl below, I see it as available from the API:

curl https://api.openai.com/v1/models/gpt-3.5-turbo-instruct -H "Authorization: Bearer <YOUR_TOKEN>"

It's an interesting idea that you might be able to make some money off. Getting anyone to spend any money on your product is difficult - I would not consider the idea validated, but you know at least a few people have that need. Now you need to figure out if more people have that need and if you can profitably reach them.

Check out indiehackers.com for more info on running these micro businesses. In my personal opinion though, I think this business might be tough to make decent money on. I think there isn't much defensible, and if you get too successful, you will get a hundred competitors over night. But there are many examples to the contrary like the guy who has been running the store locator saas: https://www.indiehackers.com/product/storemapper

Either way, best of luck!

Ah I see! So, the large/sliding window chunks act as a pre-filter for the medium chunks. That makes a lot of sense. I appreciate the response

I'm impressed by your chunking and retrieval strategies. I think this aspect is often overly simplistic.

One aspect I don't quite understand is why you filter by the sliding window chunks vs just using the medium chunks? If I understand it correctly, you find the large chunks that contain the matched small chunks from the first retrieval. Then in the third retrieval, you are getting the medium chunks that comprise the large chunks? What extra value does that provide?

The higher on the screen you play, the more velocity!

Here is a direct link to the github.io page: https://jonfranco224.github.io/not-my-cows/

I have only read about Worldcoin from comments on Hackernews so I was surprised by such a balanced take on it from Vitalik. I'm not necessarily in favor of it now, but it doesn't seem quite as evil as I had previously believed. I'm not ready to scan my iris or anything, but I'm pretty interested to see how this will go.

Found it with google: https://github.com/ruairidhflint/days-left

I love this. Is it open source? I would love to create my own with my milestones.

Zero shot implies that it was given no direct examples [0] So, in this case, it wasn't given any examples of the exact voice in combination with text, it is just using the prompt voice + a prompt text to generate new audio.

[0] https://en.wikipedia.org/wiki/Zero-shot_learning

Have you considered adding an autocomplete? Maybe add in three most likely completions to your word and have three separate signals to select each respectively? Maybe that's cheating, but that could up your words per minute drastically!

I think there is also an aspect of willful disregard. This technology may change a lot, and it may be easier to dismiss that idea rather than process it.

They do, the catch is that 250k of equity has a high chance of ending up as $0

I'm curious, how did you figure this out? Do you search texts of random HN comments? Did you remember that comment? I see these posts on here sometimes, and just don't understand how the commenters so consistently find these types of things.

In your first paragraph, you are describing Hypothetical Document Embeddings (HyDE) [0]. I've tested it out, and in certain cases, it works amazingly well to get more complete answers.

[0] https://python.langchain.com/en/latest/modules/chains/index_...

I came here to say the same thing. It's either that or Parkinson's Law [0], usually expressed as: work expands so as to fill the time available for its completion. I find that for myself it is true, when I have more time, I take my foot off the pedal. When I am running late, I work harder to meet the deadline.

[0] https://en.wikipedia.org/wiki/Parkinson%27s_law

It's intimately tied to the rest my repo, but I'll spend some time tonight and try to pull it out into it's own library.

I have implemented a scaled down version of this that just identifies the selectors needed for a scraper suite to use. for my single use case, I was able to optimize it to nearly 100% accuracy.

Currently, I am only triggering the GPT portion when the scraper fails, which I assume means the page has changed.