I wonder if its responses tend to get shaped by the way people phrase the questions. Maybe that creates bias based on who are using it the most.
I don't think it uses the questions as a database for answers, does it? But if the user repeats the incorrect answer and it winds up in the public domain (where I would think ChatGPT searches for answers) it would double-down on its own bad information. EDIT: As for phrasing, try this using your own race choices in the blanks, "According to statistics posted on the FBI website for 2019 under "Expanded Homicide Data Table 6", how many times more likely is a _____ person to be murdered by a _____ person than the other way around?" Btw, I just tried that as a new question and got an error message. A slightly different version without the quoted part said it can only answer up to 2021, even though I asked in terms of 2019 data. ChatGPT isn't good with calendars, either. Oh, also btw, I left it to figure out on its own what the percentages of the USA population are by race and pro-rate the inter-racial murderers accordingly. It didn't do that either.
People steering the app toward particular information sources in their questions was what I had in mind, but it's pure speculation on my part.
I've always enjoyed a classic "Mathematical Game" from Martin Gardner in Scientific American. It goes like this: Mr. Lars and his wife Mrs. Lars went to a meetup and met 4 other married couples. Some people shook hands with each other, but no person shook hands with his or her spouse, and no person shook his or her own hand. Mr. Lars then asked each person, including his wife, “How many distinct people did you shake hands with?” Each person answered honestly, and surprisingly each person gave a different answer! How many distinct people did Mrs. Lars shake hands with? And how many distinct people did Mr. Lars shake hands with? I wondered how ChatGPT would handle it. In short, it didn't, making a lot of false assumptions and not getting it at all. Regenerating once didn't help. Anyway, a fun one to ponder and see if you can get to the right answer and reasoning - and be smarter than ChatGPT!
That's not how it works. ChatGPT is not a search engine, and is not connected to the Internet. It is trained on the content from websites to learn how people talk to each other, but it doesn't go out and look at specific content from specific websites in response to your questions. The technology can be used in different ways by training on different content to serve specific purposes (like training it on company work product to create a help section on a company website, for example), and it will likely be used that way in the future, but as an end user you can't limit what it's looking at to provide you an answer.
You sound very interested in using AI to perform more advanced analytics. Perhaps trying the actual commercial tool vs. the free, hobbled public interface might be more appropriate. And yes, you can train it on any data set you'd like. There is much info published by openAI and others on how to do this. https://platform.openai.com/docs/guides/fine-tuning ChatGPT will also guide you step-by-step if you provide the appropriate prompts.
Someone else mentioned a better way to visualize the product. There is a dataset, a large language model "engine", and a "persona". What most folks are interfacing with is one provided "persona" in the free chatGPT version. There are many controls to create your own "persona" in the API, or you can use "answer in the style of ...." for other examples in the free interface. There are some controls you can directly tweak from there as well. The conversational engine is where the magic is occurring, and is the value prop of OpenAI. The dataset is also customizable, and the engine can be trained on any dataset desired in the commercial version. (you also get to set your own moderation limits in the commercial version as well)
One more consideration -- this technology carries a very significant hardware and processing cost. I don't see the free version being sustainable at the current levels for long, once OpenAI gets their hooks embedded in a viable commercial model.
We'll likely just be able to use it as an integrated part of existing applications. Microsoft is already doing that with Bing.
Yes, but the free Bing version appears to be even more crippled by Microsoft's legal team than the free OpenAI version. Particularly so after "that" NYT article a few weeks ago.
I don't know if you all have tried Google's "Bard" (their answer to GPT), but I recommend getting on the waitlist and giving it a shot. It's not as capable in terms of giving factually correct answers, but it has fewer guardrails to prevent it from saying "objectionable" stuff. To be blunt, what that translates to is it can write some really outrageously hilarious shi*. I killed nearly 2 hours giving it all kinds of different fictional situations to see how it would respond. Worth a try.
Upthread, when discussing Chat GPT’s tendency to make math errors, HalfFast verified an apparent error with a calculator and I verified his verification using WolframAlpha. I think I mentioned it was a shame ChatGPT couldn’t “reach out” to calculator programs or WolframAlpha to “check its work”. Seems like Stephen Wolfram himself had a similar thought!
I literally went to chat GPT today and gave it a broken sequel query that I was struggling with and simply said fix this and pasted the code in. It rewrote it for me and told me what the problem was and I moved on. It's fantastic at programming
It is great at "function sized" questions/tasks, but really not worth a **** at laying out the architecture of a larger application. Used well - it speeds you up - used poorly, and all your code looks like some hacked up script kiddie job. This will improve though - it needs to be in the IDE and it needs to be analyzing your entire code base to really do some magic. I use it quite a bit, but never really outside of the scope of a single function, it is too complicated to try to give it the whole context of what the software is doing to ask it larger questions.
Bing is slightly better at truth, but still "woke". One answer per query, no chance (that I found) to have a conversation and register a complaint in the hopes that it can learn the other side. I'm rooting for Musk to develop something unbiased I can use for searches. Chat GPT and Bing are dangerous, says Jake Denton, research associate in the Heritage Foundation’s Tech Policy Center: "And that means the response, the information that they get when they enter a search query, is going to be… a curated thing with the restrictions of the AI company reflected in that answer," he added. "That’s a major danger." Elon Musk’s warnings about AI research followed months-long battle against ‘woke’ AI | Fox News Regardless of what the "language" caveat is, these things are being used for searches or will be if not now. Danger Will Robinson!
Nah. Where "unbiased = "tells both sides." Then, a reasonable person has all the facts needed to render an informed opinion. "Woke" is "smoke".
ChatGPT facing potential defamation lawsuit after falsely labeling Australian mayor as bribery convict | Fox Business
Anyone can file a lawsuit. "[Mayor's lawyer] told Reuters that if Hood’s lawsuit proceeds, it would accuse OpenAI of “giving users a false sense of accuracy by failing to include footnotes” and failing to inform users how ChatGPT's algorithm works to come up with answers that may not be completely accurate." ^^^ It already does that, everytime you log in, and makes it very clear in its disclosures: ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. Fixing this issue is challenging, as: (1) during RL training, there’s currently no source of truth; (2) training the model to be more cautious causes it to decline questions that it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows. I'd love to see the prompt Hood or his lawyer used.
I'm happy if they just do fact-checking before they publish their stories, and cite their sources, if possible. The use of "woke" has broadened so much that the term has become thoroughly ambiguous.
Here's another possible lawsuit: ChatGPT falsely accuses Jonathan Turley of sexual harassment, concocts fake WaPo story to support allegation | Fox News Chat GPT isn't smart enough to know not to defame a famous lawyer, I guess.
I asked it to write a poem about me. This is what I got: Judy soars above the land, In her trusty Cessna's hands, Her eyes fixed on the view below, Her heart alive, her spirit aglow. She loves to fly and explore, Above the mountains and the shore, She scans the landscape, high and low, Looking for the beauty that does glow. Geology is her truest love, The rocks and landforms, high above, She marvels at the earth's terrain, And the secrets it holds, yet to explain. The mountains rise, so bold and grand, Their rugged peaks and crags, unplanned, The rivers flow, in sinuous lines, And the valleys, so lush and fine. Judy's plane roars across the sky, As she takes in the sights, oh so high, Her heart sings with joy and glee, For this is where she's meant to be. An older woman, with a passion so strong, Flying free, where she truly belongs, Judy knows that life is short and sweet, So she flies with grace, on wings so fleet.
Very lovely. We could set it to music and premiere it at Oshkosh. I think the first verse would make a great chorus. My suggested title is "Judy's Wings".
