According to OpenAI, ChatGPT now has a new model called ChatGPT o1. The new model is designed to solve complex problems across a small handful of fields, but OpenAI says the model achieves similarly to Ph.D. students in those tasks.
The new model ChatGPT o1, which currently includes both o1-preview and o1-mini versions, stands at the 89th percentile in Codeforces competitive programming contests %ranks among top 500 American students on Math Olympiad tests and “outcompetes Ph.D.-level specialists in terms of benchmark Q& chemistry problems” (OpenAI).
“It still hallucinates less than the previous two,” OpenAI head of research (and my interview subject) Dr. Jerry Tworek is quoted as saying in Chamberfied “New words from an AI,” published by The Verge. It is fed with a new optimization technique using a special training dataset. While earlier models in the field tried to synthesize patterns like their training data, O1 does reinforcement learning — using rewards and punishments.
The aspect that distinguishes o1 from past technologies manifests through its intelligence, an insightful report from The Information revealed yesterday. The model refrains from instantly spouting responses and requires between ten to twenty moments to thoughtfully compose an answer.
Referred to as “Strawberry” by curious spectators, possibly acknowledging the viral trend of influencers quizzing bots concerning how numerous “Rs” reside in the word, the innovative o1 implementation expels the necessity for “chain-of-thought prompting” in which supplementary inquiries must be made to comprehend interim reasoning. Alternatively, the version is intended to naturally display its logic from the inception.
While o1 shows promise as a precursor to more advanced conversational AI, its current limitations leave much to be desired compared to competitors. Unlike its future iterations, o1 lacks a robust connection to external data sources, restricting its abilities. It cannot ingest uploaded files or leverage the full power of developer APIs. In contrast, the o1-mini model aims to provide fast, focused answers for STEM inquiries through a streamlined design.
Across tech titans, the race is on to develop AI assistants that can truly understand and react to human needs. Earlier this year, Google showcased an enhanced version of Gemini at I/O with newfound conversational agility. More responsive than before, Gemini allows back-and-forth exchanges unconstrained by turn-taking.
Even mid-speech, it gracefully accommodates interjections without losing its train of thought. As technologies progress, the capabilities separating experimental prototypes from polished products continue narrowing at an exponential pace.
At Apple’s recent iPhone extravaganza, they showcased not only their latest handsets but also their progress in artificial intelligence. Powered by collaborations with heavy hitters like OpenAI, Apple Intelligence is the moniker for a suite of AI-assisted features coming to future iPhones.
However, while AI has long driven tech stock euphoria to new highs, cautiousness may now be creeping in. Nvidia, the chipmaker supplying neural processors for the world’s largest AI data centers, took a substantial hit last week. Broadly across the sector, investors could be taking a wait-and-see approach until tangible results from AI services become clearer.
Still, OpenAI continues to reach astronomical private valuations despite the slow materialization of their vision. The future remains unclear, but what is evident is that expectations for AI are as diverse as the companies tackling its many open questions.
For ChatGPT Plus and Team users, the ChatGPTo1-preview model is rolling out now. ChatGPT Initiative and Edu users will gain access next week. Creators can also use the API for prototyping.
