Have you tinkered around with Chat GPT? Did you like it? From what I’ve heard, it provides mixed results.
There are a variety of factors that go into your experience with Chat GPT, one of the key aspects is your communication skills.
Are you able to provide good prompts? Do you know how it works to have the right outputs? But even if you knew how to operate it, it may have provided subpar results, depending on your query. This isn’t likely to change with GPT-4, you will have to ask the right questions and prompt it in the right way. Still, what you are supposed to expect with GPT-4 is better results.
The engineers at Open AI took notice of previous issues and rigorously worked on improving it.
They came back with a new version, GPT-4.
What is GPT-4? Is it any good? Can you access it right now? Here is what you need to know about GPT-4.
What is GPT-4?
According to Sam Altman, the CEO and co-founder of OpenAI, this the latest “capable and most aligned model yet.” It is an extensive multimodal model that can take in image and text inputs and provide text outputs.
OpenAI notes that it is not as capable as humans in many real world situations but it can be on par with humans on different professional and academic domains.
When it comes to examples of how it fares in the academic domain, we have already seen its proficiency in passing legal tests and other common standardized tests. From the Bar Exam to the LSAT to the GRE, AP tests, and the SAT’s, we see that GPT-4 has been able to surpass it all.
Is GPT-4 Better?
If you look at GPT-4 and its proficiency in these standardized tests and compare it to GPT-3, then you would state that it is better. The latest version has been able to score higher on these tests. OpenAI notes that the team has been working on improving the model by using their adversarial testing program in addition to ChatGPT.
The organization further notes that the model is better on factuality, steerability, and refusing to go out of preset boundaries.
Open AI notes that individuals may not be able to see stark differences between GPT3.5 and GPT-4. But it assures users that GPT-4 will be better in handling more complexity. It will be able to shine with creativity, reliability, and absorb more nuance in inquiries.
What is interesting with GPT-4 and the exams, noted earlier, OpenAI didn’t conduct significant training for the exams, but it was able to pass many exams and go further than GPT3.0.
At the same time, Open AI notes that it has been using GPT-4 internally for a wide variety of tasks that range from support to programming.
The creativity aspect is important because it shows that GPT-4 is not regurgitating answers from an archive of information but may be heading in the direction of applying theory to more novel questions and situations.
Is GPT-4 Perfect?
No. GPT-4 is not perfect. It still has issues, it may still make up information to a question that you are answering. It does not seem to be as proficient in coding or mathematics as of yet. Indeed, it does have reasoning errors.
It may not be able to provide great performance on leetcode type of problems.
OpenAI warns that organizations should take great care when using it for high stakes situations. If that is the case, then it would be wise for organizations to use it with great caution.
Another important point to watch out for is the fact that GPT-4 is still restricted by data cut offs, it doesn’t have information after September 2021.
“GPT-4 generally lacks knowledge of events that have occurred after the vast majority of its data cuts off (September 2021), and does not learn from its experience. It can sometimes make simple reasoning errors which do not seem to comport with competence across so many domains, or be overly gullible in accepting obvious false statements from a user. And sometimes it can fail at hard problems the same way humans do, such as introducing security vulnerabilities into code it produces.”
New Risks with GPT-4
In addition to previous risks, OpenAI notes that it is optimizing for safety priorities, specifically on refusing to reply to certain requests. These requests can range from synthesizing dangerous chemicals to other inquiries that are harmful in nature.
While OpenAI is striving to implement safeguards and act cautiously, it does note that attempts to jailbreak the model can take place and may be successful.
Implications of the Improvement of these Models and the Growth of AI
While some are not impressed with the fact that the model can answer questions and ace many exams, it is important to sit back and think about what is happening. If it can answer questions on legal exams, and a wide variety of other exams, and this is how we certify ability at the present stage, then how safe are white collar workers?
For instance, we rely on specific professionals to conduct taxes, the GPT-4 demo showed that it can conduct taxes, if that is the case, what does that mean for those accountants or tax professionals that are paid at least $60,000 to conduct these activities?
It would be interesting to see how AI can help with medical research, reading medical research papers, helping to synthesize information and helping others to move faster in their own research.
On Growth of AI
Recently, an AI agent flew a VISTA X-62A training aircraft more than 15 hours, in another instance, it tricked a human into solving a captcha.
The times are wild, buckle up.
OpenAI Partners with Be My Eyes and Open Sources AIEvals
The organization has partnered with Be My Eyes to prep their image input capability. At the same time, to continue to improve the model, the organization is open sourcing AI Evals so others can evaluate and provide feedback.
How to Sign Up for The New and Improved ChatGPT
If you are looking to tinker around with the GPT-4 through ChatGPT, simply go to this link and get on the waitlist.