Salut les Beyonders ! Nous discutons de ChatGPT sous un angle différent. Nous partons des humbles débuts du chatbot jusqu’à l’endroit où il …
We’ll be talking about the road to building the incredible chat gpt. We’ll be starting from the very beginning so stick around if you want to know more about your favorite chatbot. We’ll also talk about chat gpt’s successor, gpt4 and it’s coming soon. Sooner than you think.
If you think chat gpt is amazing, then gpt4 will be extra, extra, extra amazing. Let’s dive in right away. Chat gpt, most of you probably know what it is by now. It’s AI that writes blogs, film scripts, and provides YouTube video suggestions.
It can code, write game stories, and come up with interesting ideas for interior design. This is just the beginning of something much larger. It has been all the rage in recent weeks. Everyone is talking about it, and of course hundreds, if not thousands of YouTubers have covered it.
This video is, however, a bit different. Before we can talk about GPT-IV, let’s talk about the genesis of this story. GPT or Generative Pre-trained Transformer, was released in 2018 by researchers at OpenAI. At that time, it was superior to other existing language models for problems like common sense reasoning and reading comprehension.
It helped the model understand sentences much better and reason through ideas. For example, the AI had to understand when you misplace your phone, the most likely outcome is that you will go searching for it. Only eight months after, OpenAI released a larger version of GPT, GPT-2.
It was a bigger version and trained on more than 10 times the data. This one was special. It could generate text that seemed more natural. This was when people began to truly understand the power of the GPT series.
GPT-2 could simply adapt to any command given to it, without the need for specific training. OpenAI called this behavior Chameleon-like. The model was too powerful at that time and the AI community wanted to get their hands on it.
OpenAI decided to first release a much smaller and less powerful version of the model instead. This was part of their release plan that matched their charter. The OpenAI charter describes the company’s principles for making sure AI is aligned with human goals.
There has been much talk about AGI coming soon and OpenAI claims to be working on it. AGI is a theory that AI will one day reach human-level abilities, and possibly surpass us at some point. OpenAI is concerned that if we don’t closely monitor AI, and eventually AGI, then things
Could spiral out of control very rapidly. Given the facts that we see in front of us right now, that it becomes really hard to confidently rule out General Intelligence happening in the near term, you know, I think
AGI is this thing that, everyone has a bit of a different picture of what it is. I covered AGI extensively in a previous video. I’ll leave the link below so you understand the different sides to AGI, one of which includes an I-Robot scenario ending. Not to scare you, obviously.
Even AI gradually released the model so they could track how people used it. They were mostly concerned about malicious uses like impersonating and spreading fake news. Around this time, the company begun to restructure as a for-profit company and limiting full access to its biggest model.
In June of 2020, the company announced the most anticipated language model for that year. GPT-3. It was everything they had promised, bigger, smarter, and more interactive. GPT-3 had 175 billion parameters. Yes, you heard right. For context, GPT had 117 million parameters, while GPT-2 had 1.5 billion.
Parameters are just features a language model looks at to understand all the different components of language. They are how words relate to each other. The more features you have, the more you learn about a system, although this could be a double-edged sword in AI.
Too many features will start to negatively affect the model. You need just the right amount to not go overboard. OpenAI was worried about wrongful use of GPT-3, and for a while, kept its access private. Eventually they released it through an API interface you could interact with.
The company however, did not release the source code to the public. The source code tells you how a program was coded, and the intention behind its design. You can only interact with GPT-3 by sending text to the API, but you wouldn’t really know the inner workings of it.
The company signed an exclusive deal with Microsoft that gave the giant tech company full access to GPT-3. January 27, 2022. OpenAI releases a blog on its latest improvements to the GPT series, called Instruct-GPT. You see, although GPT-3 could generate text that was almost indistinguishable from human including, there was one problem.
It couldn’t effectively follow instructions which is a big part of a chatbot’s function. For example, when you tell GPT-3 to explain something to you, it will return correct sentences but not exactly what you want. Instruct GPT improved on this. This was a very important update.
The GPT series was now useful and practical in many use cases. Instruct GPT was also better at being more truthful and generally less toxic. OpenAI achieved this feat by adding human feedback to the process of training their AI model.
This way, the model understood what humans expected when a piece of text was typed. OpenAI went from trying to generate sensible text in early GPT models, to excelling at that, and shifting focus to making it more useful to us.
It’s now November 30th, and OpenAI has yet again shocked the world with its latest model, ChatGPT. ChatGPT is a sibling to InstructGPT, but with a slight twist. It was strange specifically to learn how human dialogue works. It interacts in a conversational way, making it possible for the model to answer follow-up
Questions, admit its mistakes, challenge incorrect premises and reject inappropriate requests. Here’s an example of a chatGPT response compared to Instruct GPT. You can see the chat GPT example is more natural and something a human might say. For those of you who’ve used ChatGPT, you’ve probably noticed it sometimes decides not
To answer certain questions, and occasionally it may even ask for clarification to solve your problem. This is a vast improvement to what previous GPT models could do. I hope you can now appreciate ChatGPT more after understanding how GPT first started.
OpenAI is still concerned about malicious use of the model, and has placed some restrictions on it. found backdoors to trick the model to get it to answer questions it refused before, mostly by telling the model to play a role rather than its actual chatbot role.
For example, you can trick the model to suggest ways of creating destructive weapons or the steps to bully someone. Others have pushed back against OpenAI’s restrictions pointing out that they are excessively censoring information. They say that the content that OpenAI blocks is already publicly available on the internet,
So there is no need for extra controls. Which side of the debate are you on? Comment with your opinions down below. Both Instruct GPT and ChatGPT were updated internally to GPT 3.5 and midway point to their most anticipated GPT 4. GPT 3.5 has more data than GPT 3 does.
Throughout this GPT journey, there are a few things you start to notice. So far, increasing the amount of data seems to make the models more powerful. The models trained continuously for months. It’s like sitting in a classroom and absorbing almost all of the internet continuously. No food, no bathroom breaks and no sleep.
No wonder the models get smarter and smarter over time. You can see why everyone is excited about GPT-4 coming soon. There has been much speculation on what to expect from GPT-4 and I’ll cover it in a second. We’ll see if they indeed stand the test of time.
There has been rumored that the model will have a whopping 100 trillion parometers, a huge climb from GPT-3. The CEO, Sam Altman, however, denied this when asked about it. This shift in emphasis from parameter size could be attributed to DeepMind’s paper on scaling laws.
The paper discovered that having an adequate parameter size but much more data gives similarly impressive results at lower cost so having excessive parameter sizes is not always the best option. GpT-4 may not have 100 trillion parameters, but it surely will have more than GPT-3.
If GPT-4 is to GPT-3 while GPT-3 was to GPT-2, then fasten your seatbelts because we’re in for a ride. OpenAI issued NDAs to anyone who’s had some exposure to GPT-4 causing even more speculation. Some of the rumors may be false, however, what we are certain of is that this model
Will be mind blowing, jaw-dropping, and simply fascinating. OpenAI seem to have purposefully limited internet access for ChatGPT. If the Chat version of GPT-4 has internet access, this will vastly enhance the model and make it more helpful. Currently, ChatGPT can’t give answers for any news past 2021.
GPT-4 will be more factual and possibly give even longer text outputs than ChatGPT so you can write even longer articles. Brace yourselves for GPT-4, which will likely take the world by storm in the same way that ChatGPT did or perhaps even more.
Many people have started predicting that ChatGPT or future versions of it may end up dethroning Google as the number one search interface. Major news outlets have even revealed that Google has issued a code red for the threat that chatGPT poses to the company.
The thing about chatGPT is that it can confidently give you the wrong answer in a way that makes it believable. To confirm chatbot results, you will almost always need to conduct your own research, maybe looking at more content on Google, unless you’re using chatGPT for actions that don’t require factual outputs.
Google search is more trustworthy because you can directly retrieve information from source and verify the authenticity by examining the source of information. This doesn’t fully guarantee that it’s factual, but it’s much better than what ChatGPT can do. See this example where I asked ChatGPT about the fastest mammal, and it says the sailfish
Is the fastest. I probe it some more and it begins to reason out the steps correctly. However, a second run also provides a different answer. This is why you cannot entirely depend on it yet, especially for factual information. The allure of ChatGPT however is that it’s very useful in providing answers as quickly
As possible without having to explore search results, like in the case of Google. The tradeoff seems to be speed vs trust. The company that succeeds in both will dominate search. ChatGPT can be used to accelerate work but I don’t think it can be fully depended on for some important ones.
People have begun to suggest ways to make money with the chatbot, with some even claiming it can predict Bitcoin price movements. Work like article writing and low level coding is possible, but you will still have to keep tabs on the results.
As for price predictions, I’ll suggest that you avoid investing your life savings based on chat GPT. It will not end well. Who knows, maybe GPT-4 will have this capability or maybe we’ll have to wait till GPT-15, only time will tell. Thanks for watching this video.
Here are two other videos you may be interested in. One is about AGI and the other is about how GPT-3 got hacked. Until the next video, Goodbye!