I am interested in knowing more about how this part works. Most approaches I have seen focus on basic RAG pipelines or some variant of that, which don't seem practical or scalable.

Edit: and also, what about procedural memory instead of just storing facts or instructions?

Comments URL: https://news.ycombinator.com/item?id=45491368

Points: 14

# Comments: 0

I built Nitpicks initially because I was tired of PMs sending me small screen recordings with changes they wanted, so that it automatically implemented them for me.

Then, I was very surprised by the good results that it was producing so I decided to convert it into an actual product others can use. It is really useful for non-technical people in a product team. The whole team can now contribute to the product even if they have no idea of how to code, it's just a click and record the screen.

I would love to see you trying it out and sharing your feedback. Feel free to reach out directly to the email at the footer of the page.

Comments URL: https://news.ycombinator.com/item?id=43218610

Points: 1

# Comments: 0

I believe there are many issues for devs and users with the current approach of copy-pasting AI provider API keys into applications.

I have built brainlink.dev as a solution to this and would love to get your feedback.

It is a portable AI account that users can connect with one click to every application that integrates the SDK. It works as follows:

1. The user clicks the connect button to link his brainlink account with the app. 2. The app obtains an access token to perform inference on behalf of the user, so the user pays for the usage.

Behind the scenes, a secure Auth Code Flow with PKCE takes place so that the app obtains an access and refresh token instead of directly an API key. When the application calls a model providing a user access token, the user pays for the inference.

I believe this approach offers multiple benefits to both, developers and users.

As a developer:

- I can build and test my app against a specific model without being tied to what API key the user provided, ensuring that everyone gets the same UX.

- I can easily move my app to a different model at any time. Without brainlink, if users add, let's say, an OpenAI API key, to change to Claude I would need to ask every user to update their API key.

- Asking for API keys goes against the ToS of most providers

As a user:

- The initial friction of configuring API keys disappears, especially for non-technical users who don't know what's an API key.

- My privacy increases because AI providers can't track my usage as it goes through the proxy.

- I have a single account that I can connect to multiple apps and see how much each app is consuming

- I can easily revoke connections (tokens)

I tried to make it very simple to integrate with an embeddable button, but you can also create your own button. Here is a live demo with a very simple chat: https://demo.brainlink.dev

I would love to hear your feedback and would be happy to help anyone who wants to integrate it.

Comments URL: https://news.ycombinator.com/item?id=43050733

Points: 8

# Comments: 1

Also, I was surprised by the amounts of people that have GPUs and how well SLMs perform in many cases, even those with just 1B parameters.

Comments URL: https://news.ycombinator.com/item?id=41947574

Points: 3

# Comments: 2

Comments URL: https://news.ycombinator.com/item?id=41796830

Points: 2

# Comments: 0

I am aware of the legal and tax burdens that many entrepreneurs suffer because of the exit tax. That's the reason why many of them (me included) decide to open their companies directly in the U.S., which makes Spain to continue becoming more poor.

I really don't understand how we passed from being the biggest world empire to where we are right now in barely 500 years.

Cost of compute will continue decreasing and we will reach that point where it is feasible to have AI everywhere. I think with this particular technology we have already reached a no return point

Also, to provide some feedback, it would be awesome to make it automatically appear on the text areas and highlight errors like Grammarly does, it creates a much better UX.

For new projects that, with luck, will have a couple hundred users at the beginning it is just overkilling (and also very expensive).

My approach is usually Vercel + some AWS/Hetzner instance running the services with docker-compose inside or sometimes even just a system service that starts with the instance. That's just enough. I like to use Vercel when deploying web apps because it is free for this scale and also saves me time with continuous deployment without having to ssh into the instances, fetch the new code and restart the service.

I feel like there could be 2 major advantages: costs at scale and privacy.

1. When talking about the cost, GPT-4o-mini is inexpensive and if we continue in that path, the cost for inference will become negligible soon. Unless your company makes huge use of the model (or uses huge contexts), like those running thousands of autonomous agents, investing in the hardware, does not seem like the best alternative.

2. Privacy. I would say this is more relevant for some industries that work with highly sensitive data. However, I can see how big companies simply engage in private cloud contracts with Azure or other cloud providers. They provide that peace of mind and scalability and at the same time, depending on the contract, some guarantees.

So my big question is, do you know use cases or companies deploying LLMs on their data centers, or looking to do it or is this just for hobbyists?

Comments URL: https://news.ycombinator.com/item?id=41522549

Points: 2

# Comments: 0