The next step for an A.I. closer to how our brain works
GPT-3.5
While everyone is waiting for GPT-4, the updated version of the generative transformer model GPT-3, OpenAI has made GPT-3.5 available in the form of the previously discussed ChatGPT AI chatbot.
ChatGPT is a refined version of GPT-3.5, an update that the company had not announced. The chatbot demonstrated its ability to generate text in a conversation format, which the company claims enables it to respond to follow-up inquiries, acknowledge its mistakes, reject incorrect premises, and reject unsuitable requests. The model is a kin of InstructGPT, a refined GPT-3 model taught to follow an instruction prompt and produce a thorough response.
The development of digital content, the authoring and debugging of code, and providing customer support are all use cases for ChatGPT. GPT-3.5 is a group of models developed by OpenAI using text and code from before Q4 2021. The company has created this collection of models for various tasks rather than making available a fully trained GPT-3.5. According to the developers, ChatGPT is based on text-DaVinci-003, which is an upgrade from text-DaVinci-002 in GPT-3.
Similar to InstructGPT, the model was trained using Reinforcement Learning from Human Feedback (RLHF), but with different data collection methods: “We trained an initial model using supervised fine-tuning: human A.I. trainers provided conversations in which they played both sides—the user and an A.I. assistant. We gave the trainers access to model-written suggestions to help them compose their responses”, OpenAI said in a blog post.
“To create a reward model for reinforcement learning, we needed to collect comparison data, which consisted of two or more model responses ranked by quality”. The researchers used chatbot dialogues with AI trainers to randomly choose a model-written sentence while sampling multiple possible completions and then had the AI trainers rate the quality of each completion.
ChatGPT can also use rhymes and syllables to compose songs and poetries, therefore it appears to be able to grasp word and sentence structure.
While there is still work to be done before an AI steals anyone’s job or lesson plans, some people may be concerned about unemployed writers or the demise of essay writing in schools. Nonetheless, OpenAI detailed the new model’s limitations along with its specifications. The researchers observed that the model occasionally provides responses that may appear plausible but are inaccurate or illogical.
Additionally, ChatGPT responds inconsistently to simple phrasing adjustments, claiming ignorance in response to one inquiry but accurately responding to another. The model occasionally overuses certain words and phrases, which according to OpenAI may be a result of biases in the training data caused by trainers’ preferences for longer, more detailed responses. According to the researchers, the model occasionally infers a user’s purpose when responding to an unclear question when, in reality, it ought to be asking clarifying questions.
According to ChatGPT’s developers, the biggest drawback is that, even though OpenAI has trained the model to reject incorrect requests, ChatGPT may still react negatively to commands or behave biasedly. The company acknowledges that, unless enough user feedback is gathered to enhance it, the OpenAI Moderation API may result in false negatives or positives for some content that is flagged as dangerous.
The company says it intends to make frequent model updates to address these limitations while also gathering user feedback on undesired model results, with a focus on harmful results that “could occur in real-world, non-adversarial conditions, as well as feedback that helps us uncover and understand novel risks and possible mitigations.” To that aim, the company also said that it will run a ChatGPT Feedback Contest, with $500 in API credit as the top prize.
GPT-4
After GPT-3 and chatGPT, GPT-4 is the next-generation large language model, and it is anticipated to use 100 trillion parameters. As a point of comparison, the human brain has an average of 86 billion neurons, therefore this staggering parameter count is made possible by recent developments in artificial intelligence supercomputing from the cerebral brain scale chip cluster hardware. The use of sparsity at the core of the model’s design distinguishes GPT-4 and OpenAI from earlier large-scale models.
As a result, even though the model has a huge parameter space, the compute cost is probably lower than one might anticipate. Compared to its predecessor, which cost $12 million for every training run, it is anticipated to have training costs of roughly $6 million.
This is due to the model’s neurons having a high percentage of inactive neurons, which lowers the computational resources needed to operate the model. In Layman’s terms, this means that the model will be able to retain more potential next words, next sentences, or next emotions depending on the context of the input; in other words, GPT-4 is likely to be more like human thinking than GPT-3 because it will be able to take into account a wider range of options when producing outputs.
The adoption of sparsity in GPT-4 might enhance its functionality and increase its efficiency because it would consume less computational power to operate.
There are several ways to accomplish that, one of which is if OpenAI has improved its software-level optimization, enabling it to train the model more effectively. Another option is that GPT-4 can benefit from faster hardware or chips, which would lower the price of the computer power needed to train the model. Another explanation for the lower training expense of GPT-4 could be that OpenAI and Cerebras (an American artificial intelligence company that builds computer systems for complex artificial intelligence deep learning applications), collaborated to train their model on the latter’s most recent supercomputer CS-2, the fastest AI accelerator available, which cuts inference latencies from milliseconds to microseconds and training periods from months to minutes. Compared to AI computation based on graphics processing units, it uses a tiny fraction of the resources. In this way, performance-harming optimization strategies could be avoided by developers.
GPT-4 is also far bigger than previous sparse models like the Switch Transformer and Wu Dao, which only have a few trillion parameters.
The fact that GPT-4 is planned to be a multimodal language model, able to receive a range of inputs including text, audio, image, and potentially even video, is another significant feature. This is a promising advance because it will enable GPT-4 to provide outputs in a range of formats.
Given the recent developments in audiovisual generative artificial intelligence and language models’ already outstanding capacity to produce text that is human-like, it makes sense for OpenAI to carry out further research in this area using GPT-4. Although the precise handling of these various input modalities by GPT-4 is still unknown, the use of multimodal input and output is an important direction for the development of AI because not only does the human brain process information from a variety of senses, but also because the world is multimodal.
Artificial intelligence will be able to efficiently engage with the outside world and grasp the complexity of the human experience once it can analyze inputs relating to sight, sound, touch, smell, and even taste.
Another revolution is going to happen. With GPT-4 we could get the first AI algorithm closer to our brain and it could possibly be the first algorithm to be implemented in a robot brain. Such huge knowledge of the world could give us answers to something we can’t understand yet. Imagine if you can elaborate in your brain all the knowledge of the world and be able to make connections between all of these data. What kind of answers could you get? For a person, this knowledge could look like that of an oracle or a God. Add that you can be answered by audio and video media as well, this will change the way we could use and access information. Anyway, if AI’s answers will be more complex than what we can understand how will we discern whether an answer will be reliable or not?
