With two new versions of GPT-4, o1, a new search engine, and internal drama, ChatGPT and its parent company OpenAI have had a remarkable second year. What are the next steps?
On November 30, 2022, OpenAI launched its first ChatGPT model globally.
What was initially intended to be a test for OpenAI models quickly became a chatbot synonymous with development of generative artificial intelligence (genAI).
In 2024, OpenAI launched several new versions, including GPT-4, which offers faster intelligence for text, voice and vision, and o1, a new series of models capable of reasoning through complex tasks in the domains science, coding and math.
A few weeks ago, OpenAI launched SearchGPTa browser extension that provides “quick and timely responses” to user queries using relevant web sources, bypassing search engines altogether.
The company has faced a number of internal problems, with the resignation of its co-founder Ilya Sutskever, the dissolution of its superintelligence research team and a series of lawsuits filed by information companies based in the United States. United for alleged copyright infringement.
What awaits OpenAI’s ChatGPT, which is trying to take up the challenge in its third year of existence? Euronews Next takes stock.
OpenAI’s next “giant breakthrough”
During an Ask Me Anything (AMA) on Reddit last month, OpenAI founder Sam Altman and his colleagues gave an overview of their priorities for their third year.
Kevin Weil, product manager at OpenAI, told the AMA that one of the “big themes” of 2025 will be whether the ChatGPT can perform tasks autonomously.
Mr. Altman suggested it could be like an autonomous agent, which he saw as the company’s next “giant breakthrough.”
“AI agents,” called “agentic AI,” will enable businesses to design the large language models (LLMs) that run their systems to automate tasks in the workplace.
This is a feat that some of OpenAI’s rivals are already working on, such as Google Cloud Vertex AI agentsLinkedIn and Microsoft. According to media reports, Google’s next Gemini update, Project Jarvis, could also include autonomous agents.
Kate Devlin, professor of artificial intelligence and society at King’s College London, said AI agents already launched by OpenAI’s competitors had received mixed reactions.
“Some people are very positive about it and consider it a complete game changer, because it allows you to create a situation much closer to that of a personal assistant, where you can entrust a lot of his tasks at AI,” she said.
“Some people are wary of it and don’t like the idea of giving so much information or control to the model.”
ChatGPT’s Next Model: What We Know So Far
Some predict that OpenAI will release an entirely new model before the end of the year.
Altman told the Reddit AMA in October that there were “really good releases” coming later in the year, but there were “none that we’ll call ChatGPT-5.”
The priority, Altman said, is to “prioritize shipping” of other models, like GPT4.0 and 0.1, which were released this year.
A few weeks later, an article from The Verge indicates that the release of Orion, believed to be the successor to GPT-4o and o1, is planned for December for some companies that work closely with OpenAI, so that They can create their own products and features.
Tadao Nagasaki, CEO of OpenAI Japan, announced in September a future model of ChatGPT that will be “100 times more powerful” than GPT-4, according to local media.
Devlin said it was possible that AI agent activity could be included in a future version of Orion.
Other updates could include Sora, a text-to-video AI model, which has been delayed until now.
The software, still being “perfected” according to Mr. Weil, was leaked this week.
Mr. Altman said the next version of DALL-E, OpenAI’s image generation software, is “worth waiting for” but does not yet have a release plan.
Should OpenAI consider specializing or scaling down?
In 2025, OpenAI and other companies will particularly need to be wary of general shortages affecting the sector, Mr. Devlin continued.
“AI companies need more compute, more energy and more data,” Devlin said. “So it’s a question of what they can do with the limits placed on these elements.
One option that OpenAI could consider is reducing the size of LLMs to small or medium-sized models, which are less resource intensive.
These smaller models would be able to “curate data” to be more specific or useful in specific areas, like law or health, Mr. Devlin said.
“Instead of just increasing and increasing, it’s time to reflect on what we’ve achieved so far, because there are definite benefits… but we know there are risks and perhaps we should take the time to assess these risks,” Mr Devlin said.
Abdul Sadka, director of the Sir Peter Rigby Digital Futures Institute at Aston University, said OpenAI should keep ChatGPT “generic” so that individual companies or industries can refine it based on specific datasets for which they wish to use it.
However, Mr Sadka said he could see OpenAI expanding to give ChatGPT more “modalities”, such as the ability to recognize medical images to “potentially give you diagnostic reports on any underlying conditions” that ‘a patient might have.
To help ChatGPT become more specialized, Sadka said companies using the software could create an external knowledge base that AI has never seen before in order to “reduce the likelihood of hallucinations”, a term used to explain the AI’s hypothetical answers to questions it doesn’t recognize.