Known to be one of the most complete language models in the world, GPT-3 will see several improvements. The company specializing in artificial intelligence, OpenAI, explained on November 28 that it was working to update one of the versions of the model, InstructGPT. It made it possible to improve the rendering desired by the user.
From InstructGPT to text-davinci-003, OpenAI seeks to improve GPT-3
GPT-3 is a language model. It is primarily used to train or operate machine learning or deep learning models for natural language processing (NLP). Currently, this tool is used by several hundred applicationsincluding Neuron Writer, an SEO tool or the suite Microsoft PowerApps. The latter has also facilitated access to GPT-3 on its Azure cloud service.
Since its release in July 2020, the language model with 175 billion parameters has gone through several versions. One of its previous releases was InstructGPT, a language understanding model that brings more precision to the renderings a user wants to achieve by querying the artificial intelligence derived from GPT-3. In order to satisfy developers and researchers, OpenAI has decided to perfect InstructGPT in order to offer more realistic, more refined renderings, and therefore, closer to what a human being could offer.
Called text-davinci-003, this new version of the language comprehension model will make it possible to manage complex instructions. A real breath of fresh air for one of OpenAI’s tools the most used currently, the image generator DALL-E 2. Users of this software will be able, thanks to text-davinci-003, to generate much more accurate renderings than before.
Reinforcement learning to boost GPT-3
To develop text-davinci-003, OpenAI used reinforcement learning, a branch of machine learning where the AI remembers the mistakes it has made to avoid repeating them. In the same way, the model will remember the correct answers in order to resume them if a similar request is requested again.
As its name suggests, text-davinci-003 is the third version of this model offered by OpenAI. Its previous version, text-davinci-002 did not necessarily seek to go into detail, unlike the brand new version developed. In a press release relayed by arstechnica, OpenAI presented the differences between these two models. A user questioned an AI using text-davinci-002 by asking a very advanced tech question. He answered him in a few lines by offering him three different solutions.
With text-davinci-003, the artificial intelligence answered him with much more detail. He certainly proposed the same three solutions as text-davinci-002, but for each of them, he provided clarifications of several lines. The user ended up with three different paragraphs explaining the different uses of these solutions so that he could make his choice more easily without having to do further research.
In another register, this new model is also better to write poems or song lyrics on request depending on the needs of the user. Simple or rich rhymes, diverse and varied themes, figures of speech, sonnet or ode, text-davinci-003 will undoubtedly delight lovers of poetry.
We would like to give thanks to the writer of this write-up for this incredible material
