dcastm

I'm surprised that most of the previous comments assume they could have just not paid and faced no consequences.

I grew up in a town where these groups had significant influence. It was very common to see businesses, both big and small, paying a "vacuna." Not paying could lead to kidnapping, intimidation, or even death for the owners or operators of the business.

Makes sense, thank you!

I've tried the $500 tip idea, but it doesn't seem to make much of a difference in the quality of responses when already using some form of CoT (including zero-shot).

Awesome! If you have any feedback please let me know

I feel the key question is not that if they’re better, but if people are using them more and more often than plugins.

And in particular, if they’re using the ones that aren’t just a custom system prompt. Because I really doubt there’s any big business in commercializing system prompts.

My hunch right now is that GPTs have made it clear that OpenAI should let user save multiple system prompts, but that there’s no real defensible business in distributing GPTs, as a chat interface is not that good for most purposes.

I learned how to lucid dream regularly and taught a few others.

These are the things that work the best:

1. Record or note down your dreams right after you wake up.

2. Put an alarm clock every hour or so and do a reality check: cover your nose and try to breathe (in dreams it works), turn off and on a light (in dreams they don’t work very well), see yourself in a mirror (in dreams mirrors tend to act funny).

3. Repeat a mantra while trying to get asleep. Something like “I’m going to have a lucid dream tonight.”

4. Wake up ~5 hours after you’ve gone to sleep. Recall your dream and then try to go back to sleep repeating the mantra or imagining you’re having a lucid dream (MILD technique).

If you think objectively about this issue, it's hard to see how they can get a positive ROI out of this.

Simon does a great job highlighting why this doesn't make sense from the technical and business standpoints: https://simonwillison.net/2023/Dec/14/ai-trust-crisis/#faceb...

People often get mixed up about what causes what in these situations:

We might chat about all sorts of random things, but many times there's something that happened before, let's call it event X, that leads us to talk about a certain topic, say topic Y (like when you mention a cool shirt a friend just bought).

Then, if they see an ad about that later, folks jump to the conclusion that "Facebook was listening" to their chat. But what's more likely is that this topic was already trending among people like you, and that's why it popped up in your ads.

So, it's not that talking about it made it appear in your ads. It's more about a common event that sparked your conversation about it and also made it show up in your ads.

Why do you need a chatbot? I'm genuinely wondering about it.

It feels that you could use the same 3/4 questions and then send the results to the user. Maybe the "smart" part of it would be giving recommendations based on those results.

I think we'll start seeing a lot more services like https://www.together.ai soon.

Having open-weight models better than gpt-3.5 will drive a lot of competition on the LLM infra.

If you take input tokens in consideration is more like 5.25 eur vs. 1.5 eur / million tokens overall.

Mistral-small seems to be the most direct competitor to gpt-3.5 and it’s cheaper (1.2 eur / million tokens)

Note: I’m assuming equal weight for input and output tokens, and cannot see the prices in USD :/

GPT-3.5 is probably the most popular model used in applications (due to price point vs. GPT-4, and quality of results vs. open-weight models).

So I guess they're trying to say now it's a no-brainer to switch to open-weight models.

Very interesting. I didn’t know about it. Thanks!

I guess that at a big enough scale it starts making sense. But not sure what “big” really is.

Pgvector seems to perform very poorly compared to Qdrant: https://nirantk.com/writing/pgvector-vs-qdrant/

Me too! I started programming because of Game Maker.

I lost all the games I did. I wished I had saved them somewhere.

I wonder how much better is this, compared to taking the average ( or some other aggregation) of embeddings with a smaller context length. Has anyone done a similar comparison?

You were probably using the chat version which has been moderated, and hhh used the base version.

For those curious about it, ChromaDB uses all-MiniLM-L6-v2[0] from Sentence Transformers[1] by default.

[0] https://docs.trychroma.com/embeddings#default-all-minilm-l6-...

[1] https://www.sbert.net/docs/pretrained_models.html

It’s under Growth > Files for those struggling to find the button

Super cool.

Here the viz-related prompts (generation, editing, etc), for those interested: https://github.com/microsoft/lida/tree/main/lida/components/...

I built a tool that lets you use GPT to analyze data and build interactive graphs on the browser (https://deepsheet.dylancastillo.co/). I may try to adapt it to use LIDA or a similar approach.

Now they can hire 4 more engineers :P

> This blog, and the RSS feed will slowly grow into read-only automation that will publish to all of my channels.

WUPHF?

Amazing! Gonna check it out. Thank you!

Awesome work! I love the simplicity of htmx. It’s one of those tools that makes me feel super productive.

Will there ever be a “htmx“ for building mobile apps?

Check Cohere's embeddings of Wikipedia: https://txt.cohere.com/embedding-archives-wikipedia/

The article is referring to the problem of having a limited context length in LLMs. That is you can only pass X tokens in the prompt.

For example, let’s say you have a prompt that lets you answer questions about a book. If the book is long enough, you won’t be able to include it as is in the prompt, so you have to figure out what are the most relevant passages you must include to answer a given question. What you usually do is find the passages that are the most semantically similar to your question.

Chunk vectors are the vectorized passages of the book (i.e., a numerical vector that represents a passage), and the prompt vector is usually the vectorized question.

To find the most similar vectors you need a distance measure, cosine similarity being the most popular.

The output of finding the most similar vectors is the vectors + it’s metadata (chunk, page, chapter, etc)

Taking this to its logical conclusion, we should all be speaking Esperanto.

My guess is that there's a combination of network effects (i.e., 1.4 billion people already use it), cultural identity, and inertia.

You could close slack and open it a couple of times a day. The email is an unnecessary intermediate step.

The part I hate the most about Slack is that when you join a new Workspace, the default is to send you an email with notifications if you’re inactive more than 15 minutes. And to turn it off, you got to go into a kind of hidden configuration in the settings.

It’s anything but helping you focus.

The title is wrong. There's no effect here, just correlation.

It's pointless to draw any conclusions from this study. If you think otherwise, check https://www.tylervigen.com/spurious-correlations