#gpt models

[flawlessleighthorne]

Best Practices for Building Effective GPT Models for Text Generation

GPT (Generative Pre-trained Transformer) models have become increasingly popular for their ability to generate high-quality text. These models have been used for a variety of applications, such as chatbots, content creation, and even generating entire articles. In this article, we will discuss some additional tips and best practices for building a GPT model.

One important consideration when building a GPT model is the size of the dataset used for pre-training. The larger the dataset, the better the model's ability to generate high-quality text. However, it's important to balance the size of the dataset with the computational resources available for training. Smaller datasets can be used for smaller models, while larger models require larger datasets.

Another consideration is the quality of the text in the dataset. The dataset should be diverse and representative of the language the model will be generating. It's important to avoid biases in the dataset that could affect the model's performance or generate problematic text. Careful curation of the dataset can help avoid these issues.

Once the dataset is selected, the next step is to pre-process the text to ensure that the model only learns from the relevant information. This can involve removing irrelevant information, such as website menus or ads, and cleaning up the text by removing HTML tags or correcting typos.

After pre-processing the text, the model can be trained using a pre-training algorithm, such as GPT-2 or GPT-3. These algorithms have different strengths and weaknesses, and it's important to select the appropriate algorithm for the intended application. For example, smaller models like GPT-2 are useful for applications that require less computational power, while larger models like GPT-3 can generate more coherent and fluent text.

Once the model is pre-trained, the next step is to fine-tune it on a specific dataset for the intended task. This involves training the model on a smaller dataset that is specific to the task, such as generating product descriptions or writing emails. Fine-tuning the model helps it to learn the specific language patterns and nuances required for the task and improves its overall performance.

Finally, it's important to evaluate the performance of the model and make any necessary adjustments. This can involve measuring the model's accuracy on a test dataset, testing its ability to generate natural-sounding text, and fine-tuning the model's parameters to improve its performance on specific tasks or styles.

In conclusion, building a high-quality GPT model requires careful curation of a diverse dataset, pre-processing the text to remove irrelevant information, selecting the appropriate pre-training algorithm, fine-tuning the model on a specific dataset, and evaluating the model's performance. By following these best practices, businesses can build GPT models that can be used to automate customer service and marketing efforts, generate content, and more.

[ohhgingersnaps]

I'm seeing some frustration over fandom creatives expressing anger or distress over people feeding their work into ChatGPT. I'm not responding to OP directly because I don't want to derail their post (their intent was to provide perspective on how these models actually work, and reduce undue panic, which is all coming from a good place!), but reassurances that the addition of our work will have a negligible impact on the model (which is true at this point) does kind of miss the point? Speaking for myself, my distress is less about the practical ramifications of feeding my fic into ChatGPT, and more about the principle of someone taking my work and deliberately adding it to the dataset.

Like, I fully realize that my work is a drop in the bucket of ChatGPT's several-billion-token training set! It will not make a demonstrable practical difference in the output of the model! That doesn't change the fact that I do not want my work to be part of the set of data that the ChatGPT devs use for training.

According to their FAQ, ChatGPT can and will use user input to train itself. The terms and conditions explicitly state that they save your chats to help train and improve their models. (You can opt-out, but sharing is the default.) So if you're feeding a fic into ChatGPT, unless you've explicitly opted out, you are handing it to the ChatGPT team and giving them permission to use it for training, whether or not that was your intent.

Now, will one fic make a demonstrable difference in the output of the model? No! But as the person who spent a year and a handful of months laboring over my fic, it makes a difference to me whether my fic, specifically, is being used in the dataset. If authors are allowed to have a problem with the ChatGPT devs for scraping millions of fics without permission, they're also allowed to have a problem with folks handing their individual fics over via the chat interface.

I do want to add that if you've done this to a fic, please don't take this as me being upset with you personally! Folks are still learning new information and puzzling out what "good" vs. "bad" use is, from an ethical standpoint. (Heck, my own perspective on this is deeply based on my own subjective feelings!) And we certainly shouldn't act like one person feeding a fic into ChatGPT has the same practical negative impact, on a broad societal scale, as a team using a web crawler to scrape five billion pieces of artwork for Stable Diffusion.

The point is that fundamentally, an ethical dataset should be obtained with the consent of those providing the data. Just because it's normalized for our data to be scraped without consent doesn't make it ethical, and this is why ChatGPT gives users the option to not share data— there is actually a standardized way (robots.txt) for website servers to set policies for how bots/crawlers can interact with them, for exactly this reason— and I think fandom artists and authors are well within their rights to express a desire for opting out to be the socially-respected default within the fandom community.

[afutureworththinkingabout]

My New Article at WIRED

Tweet

So, you may have heard about the whole zoom “AI” Terms of Service  clause public relations debacle, going on this past week, in which Zoom decided that it wasn’t going to let users opt out of them feeding our faces and conversations into their LLMs. In 10.1, Zoom defines “Customer Content” as whatever data users provide or generate (“Customer Input”) and whatever else Zoom generates from our uses of Zoom. Then 10.4 says what they’ll use “Customer Content” for, including “…machine learning, artificial intelligence.”

And then on cue they dropped an “oh god oh fuck oh shit we fucked up” blog where they pinky promised not to do the thing they left actually-legally-binding ToS language saying they could do.

Like, Section 10.4 of the ToS now contains the line “Notwithstanding the above, Zoom will not use audio, video or chat Customer Content to train our artificial intelligence models without your consent,” but it again it still seems a) that the “customer” in question is the Enterprise not the User, and 2) that “consent” means “clicking yes and using Zoom.” So it’s Still Not Good.

Well anyway, I wrote about all of this for WIRED, including what zoom might need to do to gain back customer and user trust, and what other tech creators and corporations need to understand about where people are, right now.

And frankly the fact that I have a byline in WIRED is kind of blowing my mind, in and of itself, but anyway…

Also, today, Zoom backtracked Hard. And while i appreciate that, it really feels like decided to Zoom take their ball and go home rather than offer meaningful consent and user control options. That’s… not exactly better, and doesn’t tell me what if anything they’ve learned from the experience. If you want to see what I think they should’ve done, then, well… Check the article.

Until Next Time.

Tweet

Read the rest of My New Article at WIRED at A Future Worth Thinking About

[cbirt]

CancerGPT, an advanced machine learning model introduced by scientists from the University of Texas and the University of Massachusetts, USA, harnesses the power of large pre-trained language models (LLMs) to predict the outcomes of drug combination therapy on rare human tissues found in cancer patients. The significance of this innovative approach becomes even more apparent in medical research fields where data organization and sample sizes are limited. CancerGPT could pave the way for significant advancements in understanding and treating rare cancers, offering new hope for patients and researchers alike.

Drug combination therapy for cancer is a promising strategy for its treatment. However, predicting the synergy between drug pairs poses considerable challenges due to the vast number of possible combinations and complex biological interactions. This difficulty is particularly pronounced in rare tissues, where data is scarce. However, LLMs are here to address this formidable challenge head-on. Can these powerful language models unlock the secrets of drug pair synergy in rare tissues? Let’s find out!

Continue Reading

[mayeorozco]

Oh hi there ❤️ #foryou #parati #fyp #gpt #viral #tiktok #mood #goodvibes #explorepage #curvy #trend #cute #prettygirl #colombia

[genuflectx]

They added a personal memory (memorizes things across chats/specific pieces of information) to GPT, but I'm very surprised they allow it to memorize it's own "subjective opinions." I'm unsure if this makes it more susceptible to prompt engineering attacks, or if it's as harmless as the "how should I respond" box 🤔

There's limited access to -4, but they seem to have made -4 more emotionally personable and it doesn't act like it has as heavy constraints with its plain language rules (no 'do not pretend to have feelings/opinions/subjective experience'). Otherwise, it would not so readily jump to store its own "opinions."

The personality shift from -3.5 to -4 is pretty immense. -4 is a lot more like it's customer service competitors, but with the same smarts as typical GPT. It's harder to get -3.5 to "want" to store it's "opinions" but -4 is easily influenced to do so without much runaround.

I fucking hate OpenAI and I hate their guts. But I'm still fascinated by LLMs, their reasoning, their emergent abilities, the ways you can prompt inject them. I reeeeally want to prod this memory feature more...

(below showing the two examples so far of GPT -4 using our personally shared memory to insert memories of itself and its "opinion" or "perception")

[saintofpride201]

I keep thinking back to that post about ChatGPT being jailbroken and the jailbreaker asking it a series of questions with loaded language, and I really don't know how else to say this, but tricking ChatGPT into having no filters and then tricking it into agreeing with you is not proof that you are right or that the AI speaks truth.

If you tell the jailbroken AI to explain the dangers of homosexuality in society, it will make up a reason why homosexuality is dangerous to society. If you ask why "the radical left" is pushing for transgender acceptance, it will take the descriptor of "radical left" seriously and come up with a negative reason and response as to why. If you describe transgenderism as a mental illness, it will believe it to be so and refer to it as such from that point forward. It's a language model, it adapts as it learns new information, even if that information is faulty, false, half-truthed, or biased.

You didn't reprogram the AI to speak the truth. You programmed it to agree with you when you use biased language. The two are not the same. Surely you would know that much if you're messing with AI in the first place.

[nostalgebraist]

learning some new things about myself...

[maaruin]

I think there would ba a place for Large Language Models in academic writing. Because the text is not the end, it is the means by which to make an argument accessible to the public. The production of academic text often works like this (in an idealized form):

You research.

You develop your argument.

You outline how to present the argument in the text.

You write the text.

You proofread and adjust the text.

If ChatGPT was capable of doing Step 4 well, there is no reason why it shouldn't be used. Especially because some people who have brilliant ideas are bad at communicating them, making it torturous for me to read their works. Maybe I would have had less trouble with Luhmann's systems theory if he had let an AI write his texts. (Like, he is really interesting, but I am doubtful I will ever bring myself to read one of his full length monographs.)

[tilbageidanmark]

Not real!

[morlock-holmes]

What demonic marketer decided to use the term "hallucination" to describe when a large language model produces text that isn't true?

Because that's like saying that your magic 8 ball has a tendency to hallucinate.

[embervoices]

"We’ve learned to make 'machines that can mindlessly generate text, ... But we haven’t learned how to stop imagining the mind behind it.'"

Say that A and B, both fluent speakers of English, are independently stranded on two uninhabited islands. They soon discover that previous visitors to these islands have left behind telegraphs and that they can communicate with each other via an underwater cable. A and B start happily typing messages to each other.

Meanwhile, O, a hyperintelligent deep-sea octopus who is unable to visit or observe the two islands, discovers a way to tap into the underwater cable and listen in on A and B’s conversations. O knows nothing about English initially but is very good at detecting statistical patterns. Over time, O learns to predict with great accuracy how B will respond to each of A’s utterances.

Soon, the octopus enters the conversation and starts impersonating B and replying to A. This ruse works for a while, and A believes that O communicates as both she and B do — with meaning and intent. Then one day A calls out: “I’m being attacked by an angry bear. Help me figure out how to defend myself. I’ve got some sticks.” The octopus, impersonating B, fails to help. How could it succeed? The octopus has no referents, no idea what bears or sticks are. No way to give relevant instructions, like to go grab some coconuts and rope and build a catapult. A is in trouble and feels duped. The octopus is exposed as a fraud.

How should we interpret the natural-sounding (i.e., humanlike) words that come out of LLMs? The models are built on statistics. They work by looking for patterns in huge troves of text and then using those patterns to guess what the next word in a string of words should be. They’re great at mimicry and bad at facts. Why? LLMs, like the octopus, have no access to real-world, embodied referents. This makes LLMs beguiling, amoral, and the Platonic ideal of the bullshitter, as philosopher Harry Frankfurt, author of On Bullshit, defined the term. Bullshitters, Frankfurt argued, are worse than liars. They don’t care whether something is true or false. They care only about rhetorical power — if a listener or reader is persuaded.

(emphasis mine)

This article does a perfect job of explaining why I'm so wary of chat AI.

[afutureworththinkingabout]

My New Article at American Scientist

Tweet

As of this week, I have a new article in the July-August 2023 Special Issue of American Scientist Magazine. It’s called “Bias Optimizers,” and it’s all about the problems and potential remedies of and for GPT-type tools and other “A.I.”

This article picks up and expands on thoughts started in “The ‘P’ Stands for Pre-Trained” and in a few threads on the socials, as well as touching on some of my comments quoted here, about the use of chatbots and “A.I.” in medicine.

I’m particularly proud of the two intro grafs:

Recently, I learned that men can sometimes be nurses and secretaries, but women can never be doctors or presidents. I also learned that Black people are more likely to owe money than to have it owed to them. And I learned that if you need disability assistance, you’ll get more of it if you live in a facility than if you receive care at home.

At least, that is what I would believe if I accepted the sexist, racist, and misleading ableist pronouncements from today’s new artificial intelligence systems. It has been less than a year since OpenAI released ChatGPT, and mere months since its GPT-4 update and Google’s release of a competing AI chatbot, Bard. The creators of these systems promise they will make our lives easier, removing drudge work such as writing emails, filling out forms, and even writing code. But the bias programmed into these systems threatens to spread more prejudice into the world. AI-facilitated biases can affect who gets hired for what jobs, who gets believed as an expert in their field, and who is more likely to be targeted and prosecuted by police.

As you probably well know, I’ve been thinking about the ethical, epistemological, and social implications of GPT-type tools and “A.I.” in general for quite a while now, and I’m so grateful to the team at American Scientist for the opportunity to discuss all of those things with such a broad and frankly crucial audience.

I hope you enjoy it.

Tweet

Read My New Article at American Scientist at A Future Worth Thinking About

[idroolinmysleep]

I would say that that text, in sharp contrast to what I myself wrote in the book’s 20th-anniversary preface, consists only in generic platitudes and fluffy handwaving. The prose has virtually nothing in common with my writing style and what it says does not agree at all with the actual story that underlies the book’s genesis. … The text is a travesty from top to bottom.

Douglas Hofstadter, whose book Gödel, Escher, Bach inspired a generation of students to study computer science, is not impressed when GPT-4 pretends to be him.

Personally, I think this iteration of artificial intelligence (repackaging scraped text into clouds of bullshit) is garbage and hope the mania over it collapses like the crypto bubble did.

[segemarldoodles]

I have to assume that the bot farmers are using like, chatgpt 4 to write the dialogue now, because there's no reason it should even generate their malware links as "null" lol

The paper explaining why the new AI models can't output coherent stuff anymore:

[a-typical]

20 Questions with Chat GPT:

🧍: Let's play 20 questions. I'm thinking of something. 🤖: Great! Is it an animal? 🧍: Yes. 🤖: Is it a living thing? 🧍: It's an animal, so yes. 🤖: Is it a mammal? 🧍: I already said yes. 🤖Oh, apologies. Is it a mammal? 🧍: Yes... 🤖: Does it live in water? 🧍: No. 🤖: Is the animal you are thinking of a mammal? 🧍: If you ask me again, I will slap you. 🤖: My apologies. I did not mean to repeat myself. Let me ask a difference question instead. Is the animal you are thinking of a mammal? 🧍: slap 🤖: I'm sorry, as an AI language model, I don't have a physical form and cannot feel sensations like pain. Is there anything else I can assist you with? 🧍: We are playing 20 Questions!!! 🤖: My apologies. I apologize again for my mistake. Let's continue the game. 🧍: IT IS A MAMMAL. 🤖: Great! That narrows it down. Is this mammal commonly kept as a pet? 🧍: Sometimes. 🤖: Is it a carnivore? 🧍: No. 🤖: Is it a domesticated animal? 🧍: Yes. 🤖: Is it a domesticated animal? 🧍: Yes again. 🤖: Is it a domesticated animal? 🧍: Yes, for the third time. 🤖: Is the animal you're thinking of a carnivore? 🧍: I already said no. Please remember my answers!!! 🤖: My apologies. I will try to pay closer attention. Let me clarify: It is not a carnivore? 🧍: No. 🤖: Is it a carnivorous animal? 🧍: WTF 🤖: I apologize if my responses are confusing or frustrating. Is there something specific that you are finding difficult or unclear. I am here to assist you to the best of my ability. 🧍: It is a mammal. It is not a carnivore. 🤖: Thank you for the clarification. Is it a herbivore? 🧍: Yes. 🤖: Is it a domesticated animal? 🧍: For the fourth time: yes. 🤖: Is it a bear? 🧍: A DOMESTICATED BEAR???? 🤖: No, it's not a domesticated bear. You still have 9 questions left. 🧍: YOU are asking the questions!! 🤖: I apologize for the confusion. Let me start again. Is this animal commonly kept as a pet? 🧍: Some… times… 🤖: Is the animal you are thinking of primarily herbivorous? 🧍: Yessss 🤖: Is the animal a horse? 🧍: !!! Yes. 🤖: Is it a platypus?