getmeinrn

This is solvable by augmenting the database schema with comments.

When you integrate your database with an LLM, you'll notice the LLM will produce flawed queries based on wrinkles in your database schema. This is because the LLM relies on conventional understanding of how the schema is probably tied together. When you see the flawed queries, you augment the schema with a comment that explains why the schema has a wrinkle. The LLM takes that into consideration and the resulting queries are improved.

A concrete example[1]: I found that when querying the Sakila movie rental database, the generated query would frequently attempt to join the `rental` table to the `film` table through a nonexistent `film_id` column on the rental table. By adding the linked comment, the LLM stopped doing that.

1. https://github.com/amoffat/HeimdaLLM/blob/dev/notebooks/saki...

It sounds like your concern is that an LLM is not capable of understanding an english explanation of the query and producing the correct query. I personally have not had many issues here, and the time saved by tweaking a generated query is larger than writing the query myself from scratch. And LLMs will only improve so that these tweaks require less time.

You send the produced query through a parser and a validator. The parser ensures it is syntactically correct, the validator ensures it is only accessing allowed data.

Yes! You can also instruct the LLM to constrain the generated statement with specific clauses, so that the notion of "my data" in the english query comes with implicit constraints. This can be made safe using a validation scheme, which is what HeimdaLLM[1] does. It guarantees that generated queries are only joining to allowed tables, that they have required constraints, that they only select allowable columns, and that they only execute allowed functions (and more).

The end result is that you can produce safe SQL queries on behalf of your users, allowing your users to use natural language to query their data in your database.

For example, in this demo[2], I ask it:

   query("how much have i spent renting movies, broken down by month")

And it responds correctly, guaranteed to be constrained to my data.

1. https://github.com/amoffat/HeimdaLLM 2. https://github.com/amoffat/HeimdaLLM/blob/dev/notebooks/demo...

Cool language! I wish there was more standard interpreter support though. Everyone seems to be using a different flavor, which yields different results depending on where the code runs.

They tend not to fail when religion is a pillar in the community. The Amish and Mennonites are great examples.

I just assume everything I do or say is recorded and stored forever on some redundant storage and can be easily searched if I'm ever notable enough to have influence. That shouldn't influence what you do or say though, because if a system like that is corrupt enough to destroy a person for their private conversations, it's not worth participating in or supporting in the first place.

>If their cost in the cloud was about $1 per A100 hour, the training costs for this run alone would be about $63 million.

If someone legitimate put together a crowd funding effort, I would donate a non-insignificant amount to train an open model. Has it been tried before?

Now let's talk about how this extrapolates to the rest of society.

>It’s not possible to extract verbatim text from it

I didn't ask for the whole book, I asked for the first paragraph. It absolutely is possible to get verbatim text from chatgpt.

At one place I worked, I specifically asked in the interview if we would be on call. "Nope." Within 2 months of being hired, everyone on my team went on call, with no talk of compensation bumps. I was put in the awkward position of attempting to renegotiate my salary after just getting hired, or not being a team player. Don't let this happen to you.

It's good that you're thinking about compensation for this work, because it is work, it is extra, and it should never be done for free. Don't let your employer guilt you into it because "everyone else is on call too." Everyone should have a compensation bump then.

You're missing the bigger picture[1]:

  Numerous questions of law or fact common to each Class arise from Defendants’ conduct:
    whether ChatGPT itself is an infringing derivative work based on Plaintiffs’ copyrighted books;
    whether the text outputs of ChatGPT are infringing derivative works based on Plaintiffs’ copyrighted books;

It's not a simple case of "you used our copyrighted materials", it's "you're infringing on our copyright by producing works derived from materials that you used."

1. https://llmlitigation.com/pdf/03223/tremblay-openai-complain...

Yes, I acknowledged that piracy is illegal in my previous post. That's not what the lawsuit is about, according to The Verge:

>In the OpenAI suit, the trio offers exhibits showing that when prompted, ChatGPT will summarize their books, infringing on their copyrights.

>then the use of that content downstream is tainted

What does that mean exactly? That's why I used the "looking at a stolen painting" example.

Sure, pirating materials is illegal. But I don't think that's the big implication that people are getting at here. Is it legal to sell original works derived from perceiving stolen materials? Seems to me that it is.

>You can’t use illegally acquired materials when doing business.

This vague sentence conjures images of a company building products from stolen parts, but this situation seems different. IANAL, but if I looked at a stolen painting that nobody had ever seen, and sold handwritten descriptions of the painting to whoever wanted to buy one, I'm pretty sure what I've sold is not illegal.

You listed every other reason why someone would resist being policed about the word "whitelist", except the "you're being obnoxious" reason. In my experience, that's the far more common reason. Instead, you seem fixated on the racist reasons, since you mention them multiple times.

Believe it or not, people are more likely to be put off by obnoxious word police than they are to be racist.

Wouldn't they first need to prove that OpenAI didn't ingest summaries of the book, and not the book itself?

A more convincing exhibit would have been convincing ChatGPT to output some of the text verbatim, instead of a summary. Here's what I got when I tried:

    I'm sorry for the inconvenience, but as of my knowledge cutoff in September 2021, I don't have access to specific external databases, books, or the ability to pull in new information after that date. This means that I can't provide a verbatim quote from Sarah Silverman's book "The Bedwetter" or any other specific text. However, I can generate text based on my training and knowledge up to that point, so feel free to ask me questions about Sarah Silverman or topics related to her work!

This reads like you enjoy antagonizing people, so I'm sure the feeling is mutual. Who would want to work with/for someone who enjoys agitating their peers or subordinates? Do you ask them their stance on abortion and free speech too?

Have you ever considered that people don't like being told how to behave by unlikable people? And that language policing by definition makes you unlikable? When was the last time you told someone not to use a word and they said "you know what, I like being told how not to speak." ? It's maybe the quickest way to tell everyone that you're insufferable.

So basically you're "proving" to yourself that everyone is either racist or an asshole, by creating a situation where people don't want to listen to you because you're insufferable. But you can pretend to be virtuous because you're doing it for some make believe greater good about the word "whitelist."

BTW I'm actually trying to help you see the reason there's resistance. I use "allowlist" and "denylist" voluntarily in my OSS repos, because it doesn't really matter to me. But I still find people who make it their moral imperative to police these words insufferable. From my perspective, you either don't know that, which shows you have no tact, or you don't care, which shows that you're doing this in bad faith.

I'm interested, can you share?

Who said everywhere did? The link is a table showing which states did which measures. It's meant for the people who try to suggest lockdowns never really happened because they didn't happen "everywhere."

You asked where, I showed you where. You're the only one talking about Wyoming here.

To those of you thinking of debating the above, remember, you cannot win the word game. The more you oppose someone telling you not to use a word, the more racist you appear, no matter the actual origins of the word in question. It simply doesn't matter. The only way to not be seen as a racist is to apologize and use whatever words they tell you to use.

A "stay at home order" is considered a lockdown, per wikipedia. If you disagree, take it up with the editors there. You will find them far more receptive than me.

https://en.wikipedia.org/wiki/U.S._state_and_local_governmen...

Why is that funny?

Don't regret it. People will play word games with you and move the goalposts, but we all lived through it. Don't get demoralized by people who have boundless energy for gaslighting.

>Two years into the pandemic, something shifted. That’s when I was getting inundated with people who were in crisis and on the brink of divorce.

Yet another unintended consequence of lockdowns

Lottery ticket winner explains which platitudes helped them pick the winning numbers

Fellow organic user, I also find the outlook and integrity of Nvidia™ extremely refreshing. Finally a company we can believe in to play the game The Way It's Meant To Be Played™