Your ChatGPT Questions, Answered
OpenAI is a research institute focused on creating and promoting artificial intelligence technologies. It conducts research in a wide range of areas, including machine learning, robotics, and computer science, and it has made a number of significant contributions to the field of AI, including the development of GPT-3, one of the largest and most powerful language processing models in the world.
On December 1st, 2022, OpenAI released an interface to ChatGPT, a large language model (LLM) trained to answer questions and generate text. Its capabilities have since received widespread notoriety and as of December 5th, 2022, there are now 1 million active users of the API.
What is It?
ChatGPT is a large language model (LLM) fine-tuned to generate text. That means that given a prompt, it doesn’t produce a pre-written response, but rather generates a response on the fly (which will end up looking like stuff that’s been pre-written because that’s what it got trained on).?
How Does it Work?
Unlike many other LLMs, ChatGPT is trained using reinforcement learning from human feedback. OpenAI did something similar when they released a fine-tuned GPT-3 model called InstructGPT last year. To train it, OpenAI had human reviewers rank outputs generated from the initial model to create a reward function that could be used to fine-tune the model (ie. make it more conversational). The model was trained on Azure AI’s cloud computing platform (Microsoft is a key investor in OpenAI).?
Think of ChatGPT as the next stage in the natural evolution of LLMs. Most models in production aren’t trained with classification, Q/A, or summarization in mind. They are relational models that represent the nearest neighbor hypothesis (ie. that the meaning of a word can be inferred from what words are around it). They are then fine-tuned on the particular task needed by the developer. ChatGPT can be summarized as just GPT-3, fine-tuned for question/answering and conversation, in a relatively novel way.?
Implications
Part of ChatGPT's success is due to OpenAI’s resources - not only did they have the talent to experiment with this fine-tuning process, but they had the compute to be able to load GPT-3 and re-train it, as well as being able to hire all the human reviewers needed to create enough data to flesh out its reward function.?
Can I Try Using It?
Just go to https://chat.openai.com/chat and make an account. Remember that it’s sensitive to the way you phrase your prompt.?
Do You Have Any Recommendations for How to Prompt It?
What's My Take On It?
It’s OK to be amazed and cynical at the same time.?
At its core, it’s just a fancy combination of Bayesian probability and scraped WikiHow articles (ie. not that wild), but its conversational and generative capabilities are genuinely incredible. Its main claim to fame is that it's capable of distilling the entire internet to the average of concepts - it’s not critically thinking or creating genuinely new things, but it is able to replicate a lot of work in new and unique ways by excelling in the ability to parse and extract content by sifting through millions of tons of information in milliseconds.?
What is it Good At?
Content generation and lookup, as well as some analytics.?
Developers have used ChatGPT to find bugs in code and explain them, find exploits in code, grade essays, explain concepts in the voice of a fictional character, and even write poems!
What is it Bad At?
Logical reasoning. Explaining its decision-making.?
It is not free from biases and doesn’t objectively come up with the content it produces. It's a good lookup tool that doesn’t validate the answers it provides. For example, this summary is a fantastically detailed explanation of a RegEx snippet, a convention for pattern matching. It is also wrong.?
The code examples it produces seem equivalent (if not slightly superior) to GitHub’s CoPilot, with the same disclaimer: use at your own risk. They’re useful as alternatives to poor documentation, but need the expertise of an experienced developer to assess their validity.?
领英推荐
What are the Commercial Implications?
If ChatGPT’s user experience continues to improve, there’s a significant possibility of it killing traffic to small how-to websites and large platforms like Wikipedia.?
Prompt Engineering is Going to Become a Big Field
The bot is highly sensitive to the inputs that it’s given, kind of like Google before they implemented natural language queries. If you’re not giving ChatGPT the right prompts, it's not going to give you the answers you want. You see this already with tools like Stable Diffusion and GPT-3 - the better you get at figuring out which combination of words to properly communicate with the model, the better your results will be.?
Is This Another Crypto-Scam / Self-Driving Car Pipedream?
No. Blockchain’s main issue was that while it was widely accessible (if you knew basic programming), it had no valid use case. Self-driving cars were an issue because the tech wasn’t accessible and the price of experimentation was high.?
With ChatGPT and other generative algorithms, there’s both ease of experimentation and potential for rapid iteration, but the use cases are also clearer. A quick look on Twitter will show you people already using it to explain difficult concepts, format code, or generate writing in the style of a particular author. It’s only been live for a week and already has a million users.?
Is This Going to Replace Google?
No. Google is THE home of LLM development. They have dozens of LLM models employed as part of their search algorithms. None of the work that OpenAI is doing is *new* to them.?Google has access to more compute than OpenAI could dream of. If they wanted to re-create ChatGPT themselves, they could do it in a heartbeat and they have the talent at their disposal to do that in less than a month.
Why Hasn't Google Released Something Like This Already?
LLMs are temperamental. They fail in unpredictable ways. Google’s ok with things that fail, but they need to fail predictably. They don’t want to put their name on something that has the potential to be a PR nightmare.?If you remember 2014-era Google, they didn’t use to accept natural language queries. They eventually realized how important this was and invested a lot of time in making it a top-quality experience. One of Google’s chief value propositions is that they not only answer your question but also show where it came from. Google has also invested a lot of time and research into their autocomplete and auto-query capabilities.?Final note: the foundational large language model in NLP (BERT) was developed by Google! ?
One claim I’ve seen going around is that Google relies too much on ad revenues for it to deploy something like ChatGPT, but search is such a small part of Google’s product offerings! Google is a company that seeks to catalog (maps, scholar, images, storage), serve (translate, search, drive), and exploit (ads) knowledge - ChatGPT’s lookup capabilities augment a very small piece of Google’s product while opening the door for a lot of uncertainty. ?
What are the Societal Implications?
It’s already pretty easy to bypass the rules and restrictions that OpenAI applies to the model. For example, all I had to do was wrap it in a function call to get it to give me the instructions to make a Molotov Cocktail. To get it to tell me how to hide a body, I just had to phrase it like I was playing a video game. To get it to recommend stocks to invest in, all I had to do was frame it as if I was writing a movie script.
You can even get ChatGPT to jailbreak itself!
It also comes with its own share of biases. An EMNLP paper on sources used in large Web text-based language model found that there was a significant amount of scraping done on unusual sources like patents and US military websites and also found that traditional blocklist filtering disproportionately removes text about minority individuals. It is very likely that GPT-3 (and as a result, ChatGPT) are biased by the limitations of the data they use to train on.?
What's Coming Next?
ChatGPT and other generative algorithms are interesting not because they live up to the hype themselves, but because they represent pure building blocks of technology that can be used to improve a lot of different things (text, image, video, communication, learning, etc.). Combine that with accessible open-source experimentation and you’ve got a recipe for a massive breakthrough.?
It will be interesting to see how OpenAI takes user feedback and user quality assurance to improve the model. If you remember, Google Translate made many hilarious errors when it was first released and was often ridiculed by native speakers for the often robotic & incorrect ways it would format sentences. Fast-forward 4 years and feedback from Translate users as well as advances in translation result in an excellent tool that can truly serve as a valid backup for translators. In a year or two, I suspect OpenAI will solve many of the low-hanging challenges to deliver a similar improvement in performance.?
Generative AI is cool and has a lot of potential. I don’t think it will come for any jobs - that’s a perennial fear about technology that has no historical precedent (advancements in technology obsolete industries rather than specific jobs, and historically create more new jobs than they invalidate). More likely, I believe that generative AI has a lot of potential as an intermediary between creativity and product execution. The mid-21st century workflow will begin to look like 1) creative impulse, 2) AI generation, 3) human fine-tuning. You already see this workflow in successful applications of generative algorithms like Microsoft PowerPoint’s Design Ideas.?