celestialcheese

Yes - but legal and explicitly allowed by the user.

BrightData is the biggest of them, they run the free VPN Hola, and have an SDK app owners can install in their apps that allow selling bandwidth from installs. For someone who is price sensitive, trading some free residential bandwidth for whatever service is pretty compelling.

I'm sure there are scummy ones, but Bright seems to require pretty explicit consent. Not affiliated, just looked into it for some apps I have, but the payouts weren't good and I didn't think it'd be a good fit for our users.

So doesn't this just incentivize replacing FTE's with contracting firms when starting up?

Damn this is good - well played.

For those who aren't neck deep in LLM hype - ChromaDB raised $18m at a $75m valuation a couple weeks ago with what's essentially just a thin wrapper around duckdb with a parquet import/export for persistence.

20% of Harvard students pay nothing. [1]

55% receive needs based scholarships, with families making less than 150k per year paying less than 15k per year.

Sure there are a lot of rich kids and legacies. But to say it only benefits the wealthy is just wrong.

1 - https://college.harvard.edu/guides/financial-aid-fact-sheet

Not if the output quality matters - which in medical and biological classification tasks, quality labeling really really does matter. Garbage in, garbage out.

If I was paying someone for labeling or classification, and the quality dropped to 90%, 50%, 30% accuracy, I'd quickly fire them.

Because information curated by an AI with validation by a subject matter expert is worth significantly more than that which was done without a human in the loop. There's no determinism with these systems, and from experience with GPT-4 doing classification tasks, it'll do everything exactly right until it doesn't.

If it's like many businesses in the classification world, they aren't lacking for work, so it's more like they'll be able to do 10x the work done by keeping the same number of people.

If anything, it's the opposite. It's the legal restrictions around scraping that have entrenched the power of the walled-gardens today.

They built their businesses on scraping, then when they had their lock-in and monopolies, they turned around and fought against scraping to keep up-starts from eroding their business.

This blog is a great overview of the current landscape around scraping laws.

https://blog.ericgoldman.org/archives/2022/12/hello-youve-be...

Fair use has been a thorn in the side of content and publishing giants for a long time as it allows for upstarts to create derivative works and databases without violating the original copyright.

If you block fair-use and scraping, those that hold the most data and copyrights (the largest companies) will be the only ones able to create quality foundational models, thus holding the keys and further entrenching their power.

>I think this serves more to enclose that which was free, than to liberate what was closed.

What has been closed? Every day there's new foundational models being released. It's exhausting trying to keep up with the pace of change - Alpaca, Vicuna, LLaMA, etc. I'd be shocked if there wasn't a truly open source foundational model with performance equivalent to OpenAI's GPT-4 by the end of the year.

Any move to limit fair-use and scraping of publicly available information through copyright laws, no matter how good of a reason it is, gives more power to the biggest companies.

This is so backwards.

If you limit scraping and fair use, it further entrenches the power of the tech and content giants.

If you're upset about big tech's power and influence, you should be cheering on the free use of publicly available information, as it has the best shot at unseating that power.

> An annual filing with the SEC shows that J&J paid $7.4 billion in litigation expenses between 2020 and 2021, with the majority spent on legal costs related to talc claims.

https://njbiz.com/jj-offers-8-9b-to-settle-claims-talc-produ...

This is in addition to the settlement deal discussed here.

Being Google or Microsoft (or microsoft affiliated)has its perks.

Laws around scraping content and using that data for derivative works is incredibly nuanced. This article is the best up-to-date overview of the state of the industry [1].

TL;DR - IANYL. if you have enough money for legal defense, and you are scraping publicly available, not behind login-gate, content, it's probably fine and defensible, but will cost an unbelievable amount of time and money to defend.

1 - https://blog.ericgoldman.org/archives/2022/12/hello-youve-be...

lambdalabs will let you do on-demand 8xa100 @ 80GB VRAM/GPU for $12/hr, or reserved @ $10.86/hr

8xA100 @ 40gb for $8/hr

Replicate friend isn't far off.

Very very cool, something I have built personally through a patchwork of scripts, but this looks like it's been executed well.

I would use this today, but I'm concerned about lock-in. Is it possible to export/import data from baseplate?

langchain and llama-index are two big opensource projects which are great for buildign this type of thing.

https://github.com/mayooear/gpt4-pdf-chatbot-langchain for example

Very interested in this - I've been using embeddings / semantic search doing information retrieval from PDFs, using ada-002, and have been impressed by the results in testing.

The reasons the article listed, namely a) lock-in and b) cost, have given me pause with embedding our whole corpus of data. I'd much rather use an open model but don't have much experience in evaluating these embedding models and search performance - still very new to me.

Like what you did with ada-002 vs Instruct XL, has there been any papers or prior work done evaluating the different embedding models?

> In the end, YC is an investor, an investor with a strong reputation, who is now sitting in a very comfortable chair and can select the best startups and invest in them at a meager price

Wait what? a ~$7m valuation, _pre product_, is considered meager these days?

Yes, I would - these tools are massive time savers, and if I suspect it'll soon become a situation where if you don't use these tools, you're going to get left behind.

If my org wouldn't pay for them, I would start looking for other jobs (unless your in an industry that can't use them like national security, etc), but I'm not going to let $30/mo stop me from using a tool that significantly improves my work just because of a middle manager.

Also, many programmers I know bring their own keyboards.. Chefs often bring their own knives, or mechanics with their own tool chests. Ownership of your own tools is powerful.

You overestimate the willingness and ability of the average org to implement and support things like this on their own.

Tailwind - https://github.com/motifland/markprompt/blob/main/tailwind.c...

There's a big community of people out there doing just this - https://www.youtube.com/watch?v=Ao1mwEcPsEk

https://cellsaviors.com/blog/how-to-salvage-18650-cells

Reminded me of one of the cooler uses of old iphones.[1] These old phones are going to continue being useful long after their initial life as a phone. As long as apple doesn't act like apple and lock everything down.

1 - https://findthatmeme.com/blog/2023/01/08/image-stacks-and-ip...

If they open it up, possibly. But honestly, building your own tools is _super_ easy with langchain.

- write a simple prompt that describes what the tool does, and - provide it a python function to execute when the LLM decides that the question it's asked matches the tool description.

That's basically it. https://langchain.readthedocs.io/en/latest/modules/agents/ex...

This is a big deal for openai. Been working with homegrown toolkits and langchain, the open source version of this, for a number of months and the ability to call out to vectorstores, serpapis, etc, and chaining together generations and data-retrieval really unlocks the power of the LLMs.

That being said, I'd never build anything dependent on these plugins. OpenAI and their models rule the day today, but who knows what will be next. Building on a open source framework (like langchain/gpt-index/roll your own), and having the ability to swap out the brain boxes behind the scenes is the only way forward IMO.

And if you're a data provider, are there any assurances that openai isn't just scraping the output and using it as part of their RLHF training loop, baking your proprietary data into their model?

Swiss govt is signaling they will backstop it. https://www.ft.com/content/0324c5a6-cecd-4fb3-85b3-7cdc99a33...

"Built to Sell" - quick read, but transformative for me because something about it clicked with me and got me to pivot from selling "work" to selling "products"

Right now, we hire and retain bad teachers because many public schools act more as childcare than learning centers.

Without providing context or examples, GPT-4 is already better at answering questions than the average teacher, with the unlimited patience that only a computer can provide.

With context, which Khan academy has in abundance due to their lesson plans and transcripts, accuracy will be higher than even the best teachers and tutors.

Once you give context and known-true facts to best-in-class LLMs like GPT-4, the output is shockingly good.

32k context is absolutely huge. There's all sorts of techniques for summarizing large documents down to get into 4k right now with 3.5, but it's incredibly lossy.

But boy, not cheap at all - $2 per api call on a 32k token document + whatever the output.

gpt-3.5-turbo is going to be around for a long time. At this price, your use case is going to need to be replacing a large cost center. Which based on their released results on common benchmarks, is absolutely going to happen.

You think that won't get clawed back if there was something illegal?

This isn't crypto, you can guarantee that these sales will be investigated extensively.