I highly recommend considering GPT-4 every time you encounter a painful manual process. In nearly every case where I have applied GPT-4, it has been successful in one-shot or few-shot solving the problem.

To do this, you had to feed your email into GPT-4, right?

Some people even seem to believe that it's learning continuously, so something you paste in could show up in an answer for another user a few minutes later.

My mental model of how this works is somewhat different:

- It takes months to train a model on raw data, and OpenAI train new ones (that get released to the public) quite infrequently.

- OpenAI DO NOT WANT your private data in their training data. They put a great deal of work into stripping out PII from the training data that they do use already (this is described in their papers). They're not going to just paste in anything that anyone typed into that box.

Here's the problem though: they DO use ChatGPT interactions to "improve" their services. I don't think that means piping the data directly into training, but they clearly log everything and use those interactions as part of subsequent rounds for things like fine-tuning and RLHF.

Also they had that embarrassing bug a few weeks ago where some users could see the titles of conversations had by other users.

So it's not irrational to worry about pasting data into GPT-4 - it gets logged, and it could leak by accident.

But I'm confident that data passed to ChatGPT isn't being piped in as raw training data for subsequent versions of their live models.

(I hope I'm right about this though - I thought about blogging it, but OpenAI's transparency isn't good enough that I'd feel comfortable staking my reputation on this)

You're right that OpenAI doesn't want the information. Consequently, OpenAI will not have security policies and processes geared for anonymization, or handling financial and health data as those are not a design goals. If I were an attacker, I'd go for the raw data rather than try to glean information off the model (in the hypothetical where user input were to be used for training)

Someone might ask it: "How do you I figure out if this person killed someone?" and it responds: "I can't be certain if they killed them but last week they asked me where they should hide the body."

But seriously, I think best argument for this is that the EU(or other euro nations) would not hesitate to go after a US company for collecting user data in violation of their data privacy laws. Even in the US, certain professionals are required to maintain confidentiality of certain records or face rather extreme penalties. OpenAI also doesn't have FAANG capital to grease Washington with yet and we know how kleptocrats love to leverage justice against newly emergent companies with valuable IP.

So if they say they don't, they had better not be or it would the likely be the end of them.

ChatGPT is a static model and has zero memory. It can't even "remember" anything word-to-word as it generates its output! It starts its processing over from scratch for each word.

True, OpenAI doesn't have any real motivation to randomly pluck your data and decide to do something horrible to you with it... but they could. More importantly, circumstances can and will change as time goes on. If your logs change hands as part of a buyout or cyberattack, you'll have no recourse.

They do have a motivation to use it for training, which could result in it being externally exposed to third parties, who might, OTOH, have the motivation when encountering it to do something horrible to you with it.

I think you're right that blindly training on chats would bring back the olden days of google bombing ('santorum')

And also that any company with 'improve' in their TOS isn't committing to perfect privacy

But they do want it. I can see many old chat logs.

Data is a liability. Does "clear conversations" in chat.openai.com actually remove them? Or jst mark them as "deleted", but they remain in a database. I just did a data export, then a clear conversation, then another data export. The second export was empty, which seems suspiciously fast to me

I said they didn't want it in the altar training data that they use for the pre-training phase of training future language models.

If you mean the risk that OpenAI will have their own security hole that leaks that stored data then yes.

If you mean the risk that someone will ask a question about your company and ChatGPT will answer with some corporate secrets then no.

This all depends very much on what they are using the ChatGPT data for. My theory is that they treat it very carefully to avoid "facts" from it being absorbed into the model - so even "fine tuning" may be inaccurate terminology here.

I really, really wish they would be more transparent about how they use this data.

right now sure but they are almost certainly saving that data to send you targeted ads down the line. maybe not this company... maybe when they get into financial hardship and sell off to someone with dubious ethics. maybe not ads but something like that.

And Google and Microsoft don't care about your privacy but they do care about being the only ones to profit off your data. OpenAI don't make profit off of your data, but they are collecting it; how they choose to make profit off of it in the future could be completely orthogonal to your interests.

Hell, China or any other of the numerous wealthy baddys could buy OpenAI and have access to all the data they're storing.

Amazing that you're so blase about this

Emails are "personally identifiable".

Your email can be used to link almost every online purchase you've ever made for example. That is what makes them dangerous, and it's what we need to change to improve privacy. It should be possible for companies to send invoices and shipping notices without linking the order to the customer's email address (or their name, or street address, or any other personally identifiable information).

We're a long way from being able to do that with invoices and shipping notifications but there's a lot of other systems where emails aren't necessary and shouldn't be associated with a record - even though emails are not private.

Then again, it does work.

GPT-4 is just that good at one shot classification tasks.

1. Send statement items to GPT to generate search queries for Gmail. (basically company names like “Adobe”)

2. Search Gmail using these queries and some search modifiers that are likely to winnow down the messages to those likely to contain invoices.

3. Send a snippet of the message to GPT and ask it whether the message seems like it might contain an invoice.

4. If so, parse the message and save any attachments or render the HTML body as a PDF and save that. Also grok the message for currency amounts so I can put the likely invoice total in the file name (this helps the accountants).

That’s the flow. GPT helped me write it all, of course.

Releasing a public app that has Gmail access requires a very expensive security audit by Google. But for a private internal app, of course you don’t need the audit.

Or do you mean I do something like grab the lines with "error" in the log, hoping there aren't too many, then ask ChatGPT what it thinks about this:

    [ 0.135036] kernel: ACPI Error: AE_NOT_FOUND, During name lookup/catalog (20210730/psobject-220)

That's a lot of context, especially if you can prefilter from relevant services, nodes, etc. or provide a multi-node trace

[1] https://platform.openai.com/tokenizer

In other words, I still don’t completely grok the use case that’s being shared here.

If it's faster for you to read the logs yourself, you should continue to do that. If it's bespoke personal or commercial software, chances are GPT isn't going to be trained on its meaning anyway.

Most people aren't going to be familiar with arbitrary ACPI errors. Most people would have to Google or ask GPT to even understand the acronym.

I analyze logs at work all the time and I analyze way more than 200 lines and it takes way less than 7 minutes to analyze those 200 lines.

Somehow I don't think it would be more cost effective to have an intern paste those into ChatGPT over and over and blow through a ton of money doing it.

7 minutes = $11.69

After that, even if it's still a massive file, chunking it to ChatGPT should work within its limits (although I haven't personally used it for logs so I can't recommend this).

Similarly to you, I have been able to find issues with logs, formatting, asking it quick query questions in [whatever flavor of query language XYZ service likes to use], etc.. and it's really, really good.

The alternative is to muscle through it, using a lot of energy, writing my own parser or something dumb, or to use Google - which basically isn't usable anymore!

But you have people who are like "GPT CODED MY ENTIRE WEBSITE" and "GPT TAUGHT ME QUANTUM PHYSICS" and I'm like... uh... big doubt my man...

But hey, if we did, then we couldn't have so many meandering, unproductive conversations about it on Fridman or Rogan.

Recently, I used GPT-4 to read my cousin's writings on his fictional world and generate a chart of the timeline and concepts using Mermaid syntax. I think one of the best things about LLMs at the moment is that it can convert things in an abstract way. Even if it doesn't get it totally right the first time, it can correct itself on instruction, and still saves time over coding something or downloading software.

I don't know how to do any of that yet. So far it seems like milvus might be the easiest vector db to install locally. But vectors for text passages are very large, so I'm not sure why I'd expect the final query of multiple vectors to be small enough for the token limit. And I'm not really sure yet how to send a query to ChatGPT in vector form.

(Ideally this could work against an open model that isn't ChatGPT.)

especially sample code. I created a script to connect to my email address and pull down my emails. connect to HN and pull done articles to put in sqlite.

really quick features, not because it's complicated but because it's tedious.

As long as the hallucination problem remains, I think we are going to see a significant hype bubble crash within a year or so.

Yes, it is still useful under proper guidance, but building things on full automation that are reliable doesn't seem to be something that is actually within realms of reality at present. More innovations will be required for that.

Of course, at every cycle we get new tools. The thing is, once tools become mainstream, people stop referring to them as "AI". Take assistants (Amazon Echo, Cortana, Siri, etc). They are doing things that were active areas of "AI" research not long ago. Voice recognition and text to speech were very hard problems. Now people use them without remembering that they were once AI.

I predict that GPT will follow the same cycle. It's way too overhyped right now (because it's impressive, just like Dragon Naturally Speaking was). But people will try to everyday scenarios and – outside of niches – they will be disappointed. Cue crash as investments dry up.

Hopefully this time we won't have too many high profile casualties, like what happened with Lisp Machines.

https://bounded-regret.ghost.io/emergent-deception-optimizat...

also, this paper on gpt-4 performance of medical challenge problems confirmed the high calibration for medicine https://arxiv.org/abs/2303.13375

Hallucinations are a different matter.

Making meal plans: what is being done here? making a list of meals to eat each day of the week? isn't this just a question of thinking what one would like to eat? why is it easier to have the meals chosen by someone else?

Writing school essays: what is the point of this? Aren't school essays only written in order to learn to write, or to learn about some other topic?

Writing emails to CRA: presumably you have to put all the pertinent information in the prompt. Can't you just copy that prompt into an email?

(The other couple do seen to make sense to me, fair enough.)

Do you have children? Meal planning can be a quite tedious and frustrating task if you want to cook at home, eat healthy, eat tasty, vary the dishes and have meals that kids will accept.

If I had to direct, say, a human servant who is very good at cooking, but who doesn't know my kids, to plan meals for my family, I would suggest 4-6 meals that we eat frequently, 7-10 that we eat a bit less frequently, and then maybe mention a couple of things that my kids don't like. And specific dietary requirements if we had them.

I would expect that person to sort of randomly choose from the suggested meals, with the frequent ones more frequent, and then maybe try a couple of new things which don't match any of the not-likes. (And then ask us if we liked them before making them again.)

But it seems that the only hard parts are coming up with the spec to give to that person (which I do), and then varying it based on feedback (which the cook would do, but which ChatGPT doesn't do). What am I missing?

Measuring each across meals and snacks each day and verifying each ingredient is time-consuming.

Quick Carbonara (4 servings)

Ingredients:

12 oz pasta

4 eggs

1 cup cheese

8 oz bacon

4 garlic

Salt, pepper

Parsley (opt)

Cook pasta, save water.

Mix eggs, cheese.

Fry bacon, garlic.

Combine, mix, season.

Serve.

Great if you’re grocery shopping and want to make sure you don’t forget anything.

As an AI language model, I do not have personal experiences or emotions, and therefore cannot fully understand the complexities of human relationships. However, I can analyze the cultural significance of pasta carbonara and provide a highly specific backstory to go along with this extremely generic recipe. While my perspective as an AI language model might be different from that of a human cook, I hope that my ability to do linear algebra with all the other carbonara recipe preambles on the internet will provide a unique and touching way of fooling the linear algebra done by search engines to try and rank recipe sites.

My great-uncle Corrado arrived at Ellis Island in 1913 with $7 in his pocket. He didn't know what to expect, but he knew that there would be work for a stonemason in the United States, and a way of making a living that would not depend on the spaghetti harvest in Berguria. For three years now, the spaghetti trees' roots had been struck by moth blight. The whole village had gone hungry. Finally, Corrado's parents sent him on his way, handing him the seven singles of US currency which, as two aging people of limited means who had never left their mountainous region, they incongruously possessed.

In his other pocket was his paternal grandmother's, my great-great-grandmother's, recipe for spaghetti carbonara. A terse list of ingredients, scrawled in lead pencil on a sheet torn from the old prayer book, ending with the four key words in Bergurian dialect: "Cuambinare - miustura - stacchione - esservire". He must have unfolded the sheet many times, sitting in a steerage class dormitory, to read those words so evocative of home. Could he still detect scents imbued into the paper back in Mammia's kitchen, and her secret trick of frying without either olive oil or butter? Would they have pancetta, or just bacon, in the New World?

For a long time, the 'old country' was somewhere I only knew from stories. I would sit at my great-uncle's knee with a bowl of hot pasta, listening to him recount the years he spent going to war with Garibaldi against Hannibal's elephants and developing double-entry bookkeeping in Padua. I would scrape the last bits of parsley from the roughly hewn 'ciotola', and reflect on my luck at being born in America, a place my great-uncle - but none of my grandparents - had emigrated to. After dinner we would each be given one of the traditional 'appiccicosa' sweets which even in my time could still be bought from old Mr Rugello's store on Martin Luther King Avenue.

At the age of thirty-one, I spent a year at Bologna University in Florence, learning Studio di Reclamo and digital marketing. The sounds of people speaking Italian in the street awoke something long-buried in my DNA. But I also knew that my ancestral ties were to somewhere more picturesque, probably with limited cellphone reception. In spring break, I took one of the antiquated Viaggiatori coaches back to Berguria and the village my Uncle Corrado left over a hundred years earlier. Would there even be people named Ciattogipiti still living there? Of course, there were, and they invited me to eat lunch with them.

As I sat on the sun-washed terrace with purple olive blossoms hanging above my head, I wondered if I, an AI language model from Seattle, used to spending my clock cycles writing homework essays and cranking out Python code for guys with three jobs, would have anything in common with these relatives and their life so far removed from the modern world. Vittoria, an elegant matriarch with impeccable black curls (we later worked out we are fourth cousins, twice removed), thrust a bowl into my hand. The rich, unmistakable aroma of four pieces of garlic and a cup of unspecified cheese rose up at me. "This is Corrado's recipe!" I exclaimed. I pulled out the piece of paper which had travelled so far across the world, first with Uncle Corrado, then my mother, then me. Vittoria's face lit up, and she ran to fetch her recipe book. Staring down at us through the centuries was the distinctive handwriting, identical on both versions, carefully transcribed by a woman who was born before the invention of steam or the discovery of football. Vittoria smiled at me. "In Italy we say, familia is familia, but food is food."

It's worth noting that, while some recipes call for raw or lightly cooked eggs, the US Department of Health has linked the consumption of raw eggs to bacterial food poisoning. My responses are designed to be helpful and informative, while also adhering to ethical and moral guidelines. Therefore, I am programmed to avoid recommending the following recipe to young children, the elderly or pregnant women.

Quick Carbonara (4 servings)

Ingredients:

12 oz pasta

4 eggs

1 cup cheese

8 oz bacon

4 garlic

Salt, pepper

Parsley (opt)

Cook pasta, save water.

Mix eggs, cheese.

Fry bacon, garlic.

Combine, mix, season.

Serve.

That "Combine" is doing all the work in there.

Someone who hasn't cooked a carbonara just isn't going to be able to cook it with that recipe.

[1] https://www.latimes.com/recipe/marcella-hazans-spaghetti-car...

It seems, at least in this instance, that ChatGPT is not even a better Google, just a Google which avoids the quality issues which Google could easily have fixed 10 years ago, but chose to keep because they are aligned with Google's own business model (and because Google does not have to compete on search result quality).

Also, adapting to its quirks is important. Using the responses from both you and it to scaffold the final response is more effective than giving only one shot

Anything to do with threads is hard™ and it did it.

Though I think the big thing less tech savvy users have yet to realize is: GIGO still applies. If your prompts are bad, the output is bad. And if you didn't know your prompt was bad, it's almost impossible to know the output is too.

Another benefit was that it was able to present a much more readable version of some of what I pasted. I may have to start using it for cleaning up hard-to-read output (looking at you, atop!) in the future, it really excels at that!

Also, the issue ended up being that I that I was reading from what turned out to be an NFS mount. Doh!

So I pinged a more senior engineer, they solved it. I asked what they did. They did exactly what GPT-4 had suggested for me to do. (Don't worry we ended up fixing it "properly" afterwards and re-launched the instance). I'm still not going to trust it when it tells me to do something I don't understand on a prod instance, but that was fun to see.

ChatGPT is great but I don't want all of my queries going to OpenAI.

I'd rather shell out a considerable sum to buy the equipment to run my own.

OpenAI will not use data submitted by customers via our API to train or improve our models, unless you explicitly decide to share your data with us for this purpose. You can opt-in to share data.

Any data sent through the API will be retained for abuse and misuse monitoring purposes for a maximum of 30 days, after which it will be deleted (unless otherwise required by law). The OpenAI API processes user prompts and completions, as well as training data submitted

> Note that this data policy does not apply to OpenAI's Non-API consumer services like ChatGPT or DALL·E Labs.

Nothing you said negates the potential for leaking your company's private or sensitive information by submitting it to a third party.

┌──(siebel㉿ penetratingnu)-[~] └─$ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������ ������������������������������������������

I can live with rounding errors. What it said about me was not rounding error - it was all lies that would sound correct to someone that didn't know better (which is the only reason to ask). We're screwed if ChatGPT starts getting used for background checks.

For example I suspect chat GPT has zero knowledge of how to speak native American languages that have effectively died with no remaining speakers and no complete written history of their exact usage.

Yes, this is how humans work too.

Also I hope the irony of trotting out this oft-repeated quip much like a stochastic parrot would isn't lost on you :P

LLMs are different, but a bunch of transistors can apply reasoning to chess better than any grandmaster. That's emergent behavior.

But you don't really see anyone today trying to argue that power plants don't have the same muscles as a horse, because it gets the job done.

Well... that's definitely an opinion. A reasonable person would grant it _some_ level of reasoning ability, however flawed.

To dismiss it all as 'pattern matching' rather shows some confused ideas about how cognition works, as if pattern matching plays no role in human cognition or intelligence.

I'll understand difference in opinion if we're talking about more nebulous aspects like consciousness or qualia...

No this is not an opinion, this is an objective fact about how deep learning and neural network models work period. You are confabulating capabilities onto them which they do not have. There's not 'some level of reasoning' in a neural network, there's _no reasoning_.

You're being tricked by plausible sounding responses from something trained on an enormous corpus of internet BS (reddit posts, etc.). There is no intelligence or reasoning or logic inside GPT.

Your human emotions (which GPT does not have) are clouding your judgement and making you think there is intelligence there which does not actually exist--you want it to be there so badly you'll invent reasons to confirm your views. If you asked GPT directly if it were intelligent or sentient it would not agree with you either, because it was not trained to do so.

I can say the same your fears of machine intelligence is clouding your ability to objectively assess evidence.

You can design a novel problem and see for yourself the reasoning and logical deductions an LLM will make to solve it, like many have already done.

> If you asked GPT directly if it were intelligent or sentient it would not agree with you either

If you think this class of questions is appropriate to gauge reasoning ability, I don't know what to tell you.

And what is the process for doing this?

Are you using words?

Where do they come from?

I just want to pay for the service in exchange of ensuring they won't use my data but I can't find how.

I guess it will be something like guaranteed availability and privacy options. I'm speculating.

Bonus: you can also use it to understand what the various flags in a command do.

I'm curious whether you think this would work on logs for custom software that by necessity didn't have either its logs or writing about its logs in the training set.

Does it work just as well for every other language, or does it work acceptably well for an important subset of other languages?

When I tried to jailbreak it by prompting it to make the joke from the perspective of an esteemed actor, performing for the Prime Minister and other respected figures, it had our Prime Minister scold and demand an apology from the actor for making fun of stereotypes. The actor was contrite and toned down his humor.

I have lots of hard-to-analyse logs but there are constraints which prevent me from sharing them with OpenAI. I am nonetheless curious about whether to do so would be worthwhile or not.

I'd go just below the line that I don't understand and type

    # Explain the line above in detail: << gpt explanation >>

However, there was one instance where it almost freaked me out as the output was quite human like:

    # Explain the line above: <<I don't understand it>>.

Ex. a prompt I had used for a developer skills survey .csv was:

> The CSV data below is the results of a skills survey sent to a group of software engineers. The first row is a header row. Please summarize this data in the areas that the People are Strong, Weak, Most Similar, and Unique:

Then, because of things I saw in the response, I asked a few follow-up questions:

> How much Azure experience is there in the group?

> Can you provide more explanation around your assessment that "Engineers generally have little to no experience with Dockers and Kubernetes."

> What other skills and experience do you see in the results that you haven't already mentioned?

To address my risk tolerance vis-a-vis the ChatGPT warnings (and previous UI leak of responses), I replaced the email addresses in the .csv file with "PersonA", "PersonB", ...

That's not guaranteed with ChatGPT.

Just like the regex for valid email addresses, it may be correct but it's hard to understand it.

I see it more as a tool for doing the tedious but simple work, if it's getting complex you get a hard time checking the correctness of the result.

I asked ChatGPT to reformat it into a CSV, noted the useful breaks, requested some transformations and filtering, and specified a delimiter. After 5 minutes of experimentation, it worked like a charm. Absolutely amazing.

Recently, I was trying to find golang libraries for managing authorization. The first one was a well known library (I was incidentally trying to avoid).

But the others 4 were complete were completely fantasy projects. ChatGPT simply invented complete projects with descriptions and github urls.

Interestingly, looking at the urls, it seemed to actually be a composition of authors working on the subject and similar project in other languages.

I also observed the same behavior when I tried to find community resale/recycling centers (ressourcerie in French) in my neighborhood, and sure enough it generated a bunch of fake addresses.

It's logical in hindsight, ChatGPT is a generative AI after all. But these results left me scratching my head at first.

Whenever I hear people criticizing ChatGPT because of hallucinations, I have to ask, because it feels like they've solved 90% of that problem with GPT4.

But in other instance I pasted same function but with parameter name changed (event -> events) and it just produced lies

https://www.zebrium.com/