OpenAI reveals new and improved GPT-4o model – but can’t quite explain why it’s better

There is a new model of OpenAI’s GPT-4o mannequin on the town. However what it will probably exactly do appears to be a thriller, even to OpenAI. In an X put up on Monday, the corporate spilled the beans, saying: “there is a new GPT-4o mannequin out in ChatGPT since final week. Hope you all are having fun with it and test it out if you have not! we predict you may prefer it.”

In any other case, OpenAI was mum about what enhancements this new mannequin presents. In updates to its X put up, the corporate stated that the brand new GPT-4o mannequin is out there for paid subscribers in addition to these on the free tier (with a message cap). However it’s not GPT-4o-2024-08-06, which was additionally launched final week and is now operating on Microsoft Azure.

Some ChatGPT customers chimed in earlier than Monday’s announcement, claiming they observed a distinction within the chatbot’s dealing with of requests and duties. In line with VentureBeat, a number of folks felt that GPT-4o was behaving in another way and higher than previously. Others stated that GPT-4o’s native picture technology expertise via ChatGPT gave the impression to be kicking in. A number of stated that the improve improved multi-step reasoning.

In a single  X put up, an account named @misaligned_agi stated, “Wow, GPT-4o now makes use of multi-step reasoning. It is spectacular to see this in motion. Seems the replace wasn’t a brand new mannequin however a brand new methodology.”

With multi-step reasoning, an AI breaks down complicated issues and questions right into a smaller collection of sequential steps, tackling every step individually, after which comes up with the response. The very best instance is a math downside that requires a number of calculations. The AI solves every equation to reach on the general reply.

Nevertheless, a spokesperson for OpenAI advised me that the hypothesis about multi-step reasoning missed the mark.

After a lot theorizing amongst ChatGPT customers, OpenAI lastly shed some gentle in regards to the replace, now often known as ChatGPT-4o-latest. The one factor is that the corporate’s clarification continues to be obscure.

“Bug fixes and efficiency enhancements … we have launched an replace to GPT-4o that we have discovered, via experiment outcomes and qualitative suggestions, ChatGPT customers are inclined to desire,” OpenAI stated in its newest launch notes on Tuesday. “It is not a brand new frontier-class mannequin. Though we would wish to let you know precisely how the mannequin responses are completely different, determining granularly benchmark and talk mannequin habits enhancements is an ongoing space of analysis in itself (which we’re engaged on!).”

This implies that OpenAI conjured up a brand new and improved mannequin however does not actually understand how or why it is higher. Hmm, OK. Additional particulars within the launch notes nonetheless did not reply the query.

“Typically we are able to level to new capabilities and particular enhancements — and we’ll strive our greatest to speak that each time doable,” OpenAI added in its notes. “Within the meantime, our group is consistently iterating on the mannequin by including good knowledge, eradicating unhealthy knowledge, and experimenting with new analysis strategies primarily based on person suggestions, offline evaluations, and extra. That is the case with this mannequin replace.”

Right here, it feels like OpenAI is ready for customers to outline the brand new mannequin so that everybody can work out what it truly does. In different phrases, OpenAI says to its customers, “You inform me, after which we’ll each know.”

On its ChatGPT fashions web page, the corporate supplied just a few specifics on ChatGPT-4o-latest. Described as a dynamic mannequin repeatedly up to date to the present model of GPT-4o, it is supposed for analysis and analysis.

Educated on knowledge as much as October 2023, this newest mannequin can deal with 128,000 tokens, or 96,000 phrases, in a single dialog, the identical quantity as its predecessors. Nevertheless, it will probably output as much as 16,384 tokens, or 12,288 phrases, the identical as GPT-4o-mini, however with an enchancment of over 4,096 tokens within the authentic GPT-4o mannequin.

No matter new mannequin or methodology OpenAI has added to GPT-4o, the outcomes actually appear definitely worth the effort. The most recent model landed on the prime of the pack in testing at Chatbot Area, a web site that pits one AI chatbot mannequin towards one other.

Listed beneath “anonymous-chatbot,” ChatGPT-4o-latest earned a rating of 1315 primarily based on greater than 11,000 group votes, serving to OpenAI reclaim the highest spot from Google’s Gemini 1.5. Based mostly on its efficiency, the brand new mannequin confirmed a notable enchancment in such technical domains as coding, following directions, and onerous prompts.

If you wish to see for your self, taking ChatGPT-4o-latest for a spin your self is straightforward sufficient. The brand new expertise are already baked into the model of GPT-4o out there with the ChatGPT web site and cell apps (in addition to the API). ChatGPT Plus subscribers ought to be sure the mannequin is about to GPT-4o, whereas free customers can use the usual ChatGPT.

Strive asking extra complicated and nuanced questions and see how the AI fares, particularly in contrast with its previous efficiency. Then, perhaps collectively, we’ll work out what this new mannequin truly does.