Introducing ChatGPT

ChatGPT will no longer slip into your DMs. OpenAI recently fixed an issue where the AI would message people unprompted, which led to some users feeling a little creeped out. On September 15, user SentuBill posted a screenshot on the ChatGPT subreddit showing the generative model asking them about their first week at high school. After a bit of back and forth, ChatGPT stated it wanted to check in, and if they prefer to initiate conversations, to let the AI know. Another Reddit user described a similar experience. The AI model asked them how their symptoms were progressing a week after they used the platform to check on an illness. Also: Every iPhone model that will support Apple's upcoming AI features There was speculation on what caused this strange behavior to occur. At first, people thought it was a feature of OpenAI's new o1-preview and o1-mini models, collectively known as Project Strawberry. The company has promoted this technology by claiming that the two models think "through problems before [responding], much like a person would." People assumed ChatGPT's sudden ability to message users proactively was a newly trained behavior. Others believed the whole thing was faked. AI developer Benjamin de Kraker on X (formerly Twitter) demonstrated that it's possible to instruct the AI to prompt users before starting a conversation and delete the initial input.

ChatGPT can help you update your resume or draft a cover letter — but don’t forget its conversational abilities. Once you’ve got the interview, GPT can even help you prepare for potential questions and test out your answers. Since it generates responses based on the most likely word to come next, it’s the perfect tool for working out the most likely thing the average interviewer will say to you. And because this is ChatGPT, you can feed it information so it takes into account all the specifics of your unique situation. The best way to begin is by pasting the job post into the chat and asking it to generate some likely interview questions. You might want to work on answers to these questions yourself, but if you’re low on time, ChatGPT can also help you with your research. For example, if you paste the company’s “About Us” or “Company Values” web pages into ChatGPT, you can ask it to pull relevant information that you could use in your answers. Keywords are also important during interviews — ask ChatGPT to analyze the post for keywords and make sure your interview answers cover the keyword topics. According to this TikTok post, this will ensure that you’re staying on point and keeping your answers relevant to the job. Once you’ve drafted what you want to say, you can put your answers to the test. You can ask ChatGPT to generate likely responses, evaluate your answers, give suggestions, or all of the above. Just remember to take ChatGPT’s responses with a grain of salt — you’re the professional here, so if something sounds off, just ignore it or do some further research. While this isn’t the same as getting a real person to review your answers, it’s a useful free alternative to paid services that provide mock interviews. If you want the prompts, check out this Reddit post! Using ChatGPT as a makeshift website chatbot Nowadays, many product websites have helpful little chatbots that can instantly answer your questions about shipping and returns, so you don’t have to search the site for the right page. It’s super useful, but these bots aren’t just available everywhere. Luckily, ChatGPT can help you out with text-heavy web pages in a similar way — even though it can’t access the internet itself. Simply paste the entire web page into the chat (you don’t need to hand-select the information — just Ctrl+A to select the entire page, buttons and all) and ask your questions. This can help you find the right section of a terms and conditions agreement or even help you find a certain section of text that you remember vaguely but can’t recall the exact wording of. If you’re asking for paraphrased information, don’t forget to ask it to show the sections it took the information from. This way, you can see the information for yourself and know that there are no hallucinations going on. Teaching ChatGPT to talk just like you (or someone else) While the practical applications of this might be limited, people have been making tools to train LLMs like ChatGPT to mimic the speech patterns of different people. By feeding it data from WhatsApp conversations, you can task ChatGPT with learning your mannerisms, emoji usage, and talking style to generate original messages that sound like you wrote them. The designer of said tool said on Hacker News that it only managed to fool 10% of his friends but that they believe it can do much better with more work and commuting power. If someone can get this work with enough accuracy, it’s easy to imagine a product that can quickly and effortlessly answer messages and emails on your behalf without sounding out of place. The idea can also be used for more fantastical purposes — a new website recently popped up that uses an LLM trained to respond just like Albus Dumbledore (RIP Michael Gambon) from the Harry Potter series. Its sole purpose is to answer your questions in a mysterious and whimsical manner — and it does so pretty well! If you’re interested in carrying out your own experiments, check out the GitHub repository here.

Understanding the mechanics of ChatGPT requires a closer look at the underlying technology, the training process, and the way it generates responses. Let's delve deeper into these aspects. The Underlying Technology: Transformer Models ChatGPT is based on the GPT (Generative Pretrained Transformer) architecture, specifically GPT-4. Transformer models, introduced in the paper "Attention is All You Need" by Vaswani et al., have become a cornerstone of many natural language processing tasks. Transformers use a mechanism called "attention" to weigh the influence of different words when generating a response. For instance, while generating a response to the sentence "The cat chased its tail," the model would understand that "cat" is the subject and is more important than "tail." The GPT architecture extends on transformers with a focus on generating coherent and contextually accurate text. Training ChatGPT ChatGPT's training is a two-step process: pre-training and fine-tuning. Pre-training During this stage, the model is exposed to a large corpus of text data from the internet. However, it's important to note that the model doesn't know specifics about which documents were in its training set or have access to any specific documents or sources. The goal of this phase is to learn the statistical patterns of the language. For instance, if the phrase "The sky is..." is often followed by the word "blue" in the training data, the model will learn to predict "blue" as a likely next word. Fine-tuning After pre-training, the model goes through a fine-tuning process, where it is trained on a narrower dataset generated with the help of human reviewers following certain guidelines provided by OpenAI. This dataset consists of various hypothetical inputs and responses. This stage helps the model respond better to specific inputs, maintaining a safer and more useful interaction with users. Generating Responses Once trained, ChatGPT generates responses using a method known as "autoregression." It begins with an input message, then predicts the next word, and the next, until it forms a complete sentence. It chooses each word based on the probabilities of what it predicts as the next likely word, considering all the previous words in the input. This process continues until a specific condition is met, such as reaching a maximum number of words. ChatGPT sometimes introduces randomness in its word selection process to generate diverse and creative responses. By understanding these mechanics, we can better understand how ChatGPT mimics human-like conversation, thus making interaction with technology more seamless and natural.

14 December