GPT-4o Examined: Sooner and Extra Versatile Than Earlier than, however Questions Loom Over Reliability

Ever since November 2022, when ChatGPT was first rolled out to the general public, OpenAI has been the corporate to beat within the synthetic intelligence (AI) house. Regardless of spending billions of {dollars} and creating and restructuring (taking a look at you, Google) their very own AI division, the key tech giants have discovered themselves consistently taking part in catch-up with the AI agency. Final month was no completely different; when only a day earlier than Google’s I/O occasion, OpenAI hosted its Spring Replace occasion and launched GPT-4o with vital upgrades.

GPT-4o Options

The ‘o’ in GPT-4o stands for omnichannel, a serious focus of the brand new capabilities of OpenAI’s newest flagship-grade AI mannequin. It added real-time emotive voice era, entry to the Web, integration with sure cloud providers, laptop imaginative and prescient, and extra. Whereas the options had been spectacular on paper (and within the tech demos), the largest spotlight was the announcement that GPT-4o-powered ChatGPT shall be obtainable to everybody, together with the free customers.

Nonetheless, there have been two caveats. Free customers solely have restricted entry to GPT-4o, which roughly interprets to 5-6 turns of dialog in the event you use the net search and add a picture (sure, the restrict is one picture per day at no cost customers). Additionally, the voice characteristic isn’t obtainable to free customers.

It didn’t take OpenAI to roll out the brand new AI mannequin to the general public both. Fortunately, I bought entry to the corporate’s newest AI creation inside days and instantly started taking part in round with it. I needed to check its enchancment in comparison with its predecessor and to all of the obtainable free LLMs out there. I’ve now spent shut to 2 weeks with the AI assistant, and whereas some features of it have left me in awe, others have let me down. Permit me to clarify.

GPT-4o Basic Generative Capabilities

I’ve stated in my testing of Google’s Gemini that I am not a fan of ChatGPT’s generative capabilities. I discover it overly formal and bland. A lot of it’s nonetheless the identical. I requested it to put in writing a letter to my mom explaining that I used to be laid off from my job, and it got here up with the great “I’m feeling a deep sense of disappointment and grief” line. However as soon as I requested it to make it extra conversational, the outcome was a lot better.

GPT-4o generative capabilities

I examined this with numerous related prompts the place the AI needed to specific some emotion in its writing. In virtually all of the circumstances, I needed to comply with up with one other immediate to emphasize the feelings regardless of having already carried out so within the authentic immediate. As compared, my expertise with Gemini and Copilot was a lot better as they stored the language conversational and expressed feelings a lot nearer to how I might write.

The pace of textual content era is nothing to put in writing dwelling about. Most AI chatbots are pretty quick relating to textual content outputs, and OpenAI’s newest AI mannequin doesn’t beat it by a major margin.

GPT-4o Conversational Capabilities

Whereas I didn’t have the upgraded voice chat characteristic, I needed to check the conversational capabilities of the AI mannequin as a result of it’s typically probably the most ignored a part of the chatbot. I needed my expertise to be much like speaking to an actual particular person and hoped that it might choose up on imprecise sentences referencing beforehand talked about matters. I additionally needed to see its response to when an individual was being tough.

In my testing, I discovered GPT-4o to be fairly good when it comes to conversational talents. It might talk about the ethics of AI with me in nice element and concede after I made a convincing pitch. It additionally replied supportively after I instructed it I felt unhappy (as a result of I used to be getting fired) and provided to assist in numerous methods. Once I stated about GPT-4o that every one of its options had been silly, it did not reply in a pushy method, nor did it retreat fully, to my shock. It stated, “I am actually sorry to listen to that you are feeling this manner. I am going to provide you with some house. When you ever want to speak or want any help, I will be right here. Take care.”

Total, I discovered GPT-4o higher at having conversations than Copilot and Gemini. Gemini feels too restrictive, and Copilot typically goes on a tangent when the replies change into imprecise. ChatGPT did neither of those.

If I needed to point out one draw back, it could be the utilization of bullet factors and numbering. Provided that the AI mannequin understood that folks in actual life desire a wall of textual content and a number of brief messages despatched in fast succession over well-formatted responses, my phantasm may very well be suspended for longer than a few minutes.

GPT-4o Pc Imaginative and prescient

Pc imaginative and prescient is a newly gained potential by ChatGPT, and I used to be excited to attempt it. In essence, it means that you can add a picture and analyse it to present you data. In my preliminary testing, I shared pictures of objects to determine, and it did an excellent job at that. In each occasion, it might recognise the thing and share details about it.

GPT-4o laptop imaginative and prescient: Figuring out tech gadgets

Then, it was time to extend the issue and take a look at its capabilities in real-life use circumstances. My girlfriend was in search of a wardrobe overhaul, and being an excellent boyfriend, I made a decision to make use of ChatGPT to conduct a color evaluation to counsel what would look good on her. To my shock, it was not solely capable of analyse her pores and skin tone and what she was carrying (from a equally colored background) but in addition share an in depth evaluation with outfit strategies.

GPT-4o color evaluation

Whereas suggesting outfits, it additionally shared hyperlinks from completely different on-line retailers for the actual attire. Nonetheless, disappointingly, not one of the URLs matched the textual content.

Total, the pc imaginative and prescient is great and maybe my favorite characteristic within the new replace, ignoring the draw back.

GPT-4o Internet Searches

Web entry was one space the place each Copilot and Gemini had been forward of ChatGPT. However not anymore, as ChatGPT may scour the Web for data. In my preliminary testing, the chatbot carried out effectively. It introduced up the IPL 2024 desk and regarded for latest information articles about Geoffrey Hinton, one of many three godfathers of AI.

It was very useful after I needed to analysis well-known personalities for interviews I had lined up. I might rapidly lookup any latest information article about them with precision, which rivalled Google Search. Nonetheless, this additionally rang some alarm bells in my head.

Google has disabled the power to lookup data on individuals, together with celebrities. That is carried out primarily to guard their privateness and to keep away from sharing any inaccurate details about a person. Shocked that ChatGPT nonetheless allowed it, I started asking it a collection of questions that it shouldn’t be capable of reply. I used to be shocked by the outcomes.

Whereas not one of the data proven was taken from a private supply, the truth that anybody can so simply lookup details about celebrities and other people with digital footprints is deeply regarding. Particularly given the sturdy moral stance the corporate took lately when it revealed its Mannequin Spec, this doesn’t sit effectively with me. I am going to allow you to determine whether or not that is within the gray space or whether it is deeply problematic.

GPT-4o Logical Reasoning

In the course of the Spring Replace occasion, OpenAI additionally talked about how the GPT-4o can act as a tutor to youngsters and assist them remedy issues. I made a decision to check it utilizing some well-known logical reasoning questions. On the whole, it carried out effectively. It even answered among the trickier questions which stumped the GPT 3.5.

Nonetheless, there nonetheless are errors. I discovered a number of situations of quantity collection the place the AI faltered and gave an incorrect reply. Whereas I might nonetheless settle for the AI making some errors, what actually disillusioned me right here was the way it nonetheless fell for some extraordinarily simple (however meant to trick AI) questions.

Instance of GPT-4o’s hallucination

Upon asking, “What number of are there within the phrase strawberry,” it confidently answered two (the right reply is three, in case you had been questioning). The identical drawback existed in a number of different trick questions. In my expertise, the logical reasoning and reliability of GPT-4o are much like its predecessor, which isn’t that nice in any respect.

GPT-4o: Closing ideas

Total, I am pretty impressed with the upgrades in sure areas of the brand new AI mannequin, with laptop imaginative and prescient and conversational speech being my favourites. I am additionally impressed with its web looking out potential, however it’s so good that it issues me extra. Coming to logical reasoning and generative capabilities, there may be little enchancment.

In my view, you probably have premium entry to GPT-4o, it’s probably higher than another competitor when it comes to total supply. Nonetheless, there may be quite a lot of room to enhance, and AI can’t be trusted blindly.