jasonwcfan

Yeah the ways AI models have learned to interact with the world are all hilariously skeuomorphic given their capabilities. A model that runs on silicon has to learn English and Python in order to communicate with other models that also run on silicon. And to perceive the world they have to rely on images rendered in the limited wavelengths visible to the human eye.

But I much prefer this approach over allowing models to develop their own hyper-optimized information exchange protocols that are are black box to humans, and I hope things stay this way forever.

The way we plan to handle authenticated sessions is through a secret management service with the ability to ping an endpoint to check if the session is still valid, and if not, run a separate automation that re-authenticates and updates the secret manager with the new token. In that case, it wouldn't need to be stateful, but I can certainly see a case for statefulness being useful as workflows get even more complex.

As for device telemetry, my experience has been that most companies don't rely too much on it. Any heuristic used to identify bots is likely to have a high false positive rate and include many legitimate users, who then complain about it. Captchas are much more common and effective, though if you've seen some of the newer puzzles that vendors like Arkose Labs offers, it's a tossup whether the median human intelligence can even solve it.

Why is it unethical when courts have repeatedly affirmed browser automation to be legal and permitted?

If anything, it's unethical for companies to dictate how their customers can access services they've already paid for. If I'm paying hundreds of thousands per year for software, shouldn't I be allowed to build automations over it? Instead, many enterprise products go to great lengths to restrict this kind of usage.

I led the team that dealt with DDoS and other network level attacks at Robinhood so I know how harmful they are. But I also got to see many developers using our services in creative ways that could have been a whole new product (example: https://github.com/sanko/Robinhood).

Instead we had to go after these people and shut them down because it wasn't aligned with the company's long term risk profile. It sucked.

That's why we're focused on authenticated agents for B2B use cases, not the kind of malicious bots you might be thinking of.

I mentioned this in another comment, but I know from experience that it's impossible to reliably differentiate bots from humans over a network. And since the right to automate browsers has survived repeated legal challenges, all vendors can do is make it incrementally harder to weed out the low sophistication actors.

This actually creates an evergreen problem that companies need to overcome, and our paid version will probably involve helping companies overcome these barriers.

Also I should clarify that we're explicitly not trying to build a playwright abstraction - we're trying to remain as unopinionated as possible about how developers code the bot, and just help with the network-level infrastructure they'll need to make it reliable and make it scale.

It's good feedback for us, we'll make that point more clear!

Thanks! Wasn't familiar with Browserless but took a quick look. It seems they're very focused on the scraping use case. We're more focused on the agent use case. One of our first customers turned us on to this - they wanted to build an RPA automation to push data to a cloud EHR. The problem was it ran as a single page application with no URL routing, and had an extremely complex API for their backend that was difficult to reverse engineer. So automating the browser was the best way to integrate.

If you're trying to build an agent for a long-running job like that, you run into different problems: - Failures are magnified as a workflow has multiple upstream dependencies and most scraping jobs don't. - You have to account for different auth schemes (Oauth, password, magic link, etc) - You have to implement token refresh logic for when sessions expire, unless you want to manually login several times per day

We don't have most of these features yet, but it's where we plan to focus.

And finally, we've licensed Finic under Apache 2.0 whereas Browserless is only available under a commercial license.

Oops! We tested the Oauth flow but forgot to update the email one. Thanks for the heads up, fixing this now.

Proxies are definitely on our roadmap, but for now it just supports stock Playwright.

Thanks for the feedback! I just updated the repo to make it more clear that it's Playwright based. Once my cofounder wakes up I'll see if he can re-record the video as well.

Yep. I used to be the guy responsible for bot detection at Robinhood so I can tell you firsthand it's impossible to reliably differentiate between humans and machines over a network. So either you accept being automated, or you overcorrect and block legitimate users.

I don't think the dead internet theory is true today, but I think it will be true soon. IMO that's actually a good thing, more agents representing us online = more time spent in the real world.

A close friend of mine was an associate of Mr Beast, even living in his house in Greenville for several months. He confirmed a lot of the negative press about him in the media, and was himself ultimately screwed over by Jimmy and has been trying to get recompense for years.

MrBeast has always been clear that his goal is to make the best videos in the world. Not to be the most nurturing place to work, or the most philanthropically minded. This document makes that clear. It shouldn't come as a surprise to anyone that in becoming the best in the world at youtube, he's had to become an extremely toxic individual.

My handwriting is barely legible, even to my self. So I'll stick to Notion, thank you very much.

Since you didn't include the link I went ahead and searched for it. Is this it? https://github.com/Storia-AI/repo2vec

There's been a lot of "chat with your x" projects and the value prop always eludes me.

To use the example in the repo, if I want to know what image encoders are supported, I would do a repository search for the "Encoder" keyword to find where they're defined. Then I'd be able to see all the encoders that are supported. That takes me about 10 seconds - why would I want to use a chatbot to do this instead?

Seems like this doesn't solve the problem, it just changes the nature of the problem from a link rot to having a single point of failure.

What assurances does perma cc give that it will continue to maintain its index for the foreseeable future as costs increase? Wayback machine is maintained by a non-profit with a charter and >$30m in annual donations/revenues. As far as I can tell Perma is maintained by a single entity (Harvard Law Library)

Not trying to be negative here but genuinely don't see how a SPOF is better than link rot.

One of the open source vector DBs is probably your best bet. Chroma, weaviate, Qdrant and a few others

From the enterprise perspective, gen AI tools are already incredibly powerful accelerators of productivity. The problem is it's unevenly distributed. Adoption at some companies is probably close to 100%, and at others maybe 1 or 2%.

In 6 months to a year we'll really start to see the outcomes of those employees that know how to use these tools and those that don't diverge, and companies are going to pour a lot of resources into providing training and access to them.

Just ran our deployed cluster through GCP's pricing calculator and it's about $300 USD per month with Llama 2

Yep. We use LangChain's basic text splitter to chunk the documents and the QA chain to stuff it into the prompt. But AFAIK it doesn't check for context length so that's a piece that's still missing.

Upper limit depends on the model, Llama 2 is 4k including the prompt.

Yep the docker containers should run fine on local hardware, but the terraform config only supports GCP right now.

In terms of cost - just ran our deployed cluster through GCP's pricing calculator and it's about $300 USD per month. Definitely not cheap for individual use, but pretty affordable for enterprise use. Running the 40B parameter version will be significantly more.

Not yet, but we can definitely add it. Created an issue: https://github.com/psychic-api/rag-stack/issues/2

In the meantime it uses GPT4all when running locally so you can technically deploy it as well, but it's not very good.

It uses all-MiniLM-L6-v2 from huggingface by default

https://huggingface.co/sentence-transformers/all-MiniLM-L6-v...

You can also specify a specific embeddings model from SentenceTransformers to use in /server/.env

Thanks for the callout! We'll add the local.env instructions to the readme.

Are you using it with input docs or without? Locally it uses GPT4all which isn't nearly as good as Llama or Falcon. I saw a project that is docker for Llama 2 so we might use that instead!

You're right. Either way it's impossible to recreate Llama 2 without the data set so perhaps "free to use model" is a better description than "open source model"

The concept of “source” is nebulous for ML models. If you have the weights you can recreate a model without access to the source code originally used to train it, and similarly just having the source code without the training data won’t allow you to recreate the model.

While it would be nice to have the data set Meta used I think open sourcing the weights is good enough.

We have about 10 other connectors in a separate project at https://github.com/psychic-api/psychic

Thanks for the feedback! We’ll include a demo soon.

OpenAI functions solves this problem for the GPT API. It's only a matter of time before the same functionality is available for open source models

Looks cool! One piece of feedback: It's really tedious to have to type in JSON. Making it a form would save tons of time having to write it out manually.

I had no idea what "unifying" meant at first. The best guess I can make is something that lets humans interact with AI and computer systems, which is super vague.

Maybe I'm just to used to all the marketing speak out there where "unifying" has been co-opted to mean basically nothing. e.g. "We unify technology with human potential!"

Cool - if I'm understanding correctly it's basically an API between different classes of "agents"?

One piece of feedback: the project sounds useful once I read into it a bit more but the headline is confusing since "unifying" can mean many different things.

Hey I've been thinking a lot about your comment. Would you be open to connecting non-anonymously? Would love to pick your brain on API integrations. If so you can email me at jason@psychic.dev so you don't have to doxx yourself.

Exactly our intent :) some people just like to hate

We specifically chose AGPL-3 because we wanted it to be permissive, but we didn't want others to fork our project, take it closed source, and charge for it without adding back anything of value.

We also don't expect companies to customize the functionality, just to self-host it or use the cloud version, or use it for personal projects.