OpenAI Launches Strawberry AI, Pushing ChatGPT Closer to Human Intelligence
The preview version is now available to ChatGPT Plus and Team users, with Enterprise and Edu users set to gain access next week.
Amidst recent rumours of a new AI model in development, OpenAI has officially pulled back the curtain on its long-rumoured AI model, code-named "Strawberry."
This model family, formally also known as "01," promises significant advancements in reasoning and problem-solving where traditional large language models (LLMs) have often fallen short.
At the heart of this release are two variants: o1-preview and o1-mini. The former is a full-fledged powerhouse aimed at complex, multi-step reasoning, while the latter is a more nimble, cost-effective option optimized for tasks like coding.
The o1-mini may lack some of the sophisticated reasoning skills of its sibling, but its 80% lower price tag makes it an attractive option for developers seeking efficient solutions without breaking the bank.
Emmanuel Oyedeji
OpenAI in its announcement claims that o1-preview outperforms its predecessor, GPT-4o, across multiple benchmarks, particularly in competitive programming, mathematics, and scientific reasoning.
OpenAI says o1-preview can handle multi-step problems, including complex math and coding questions. This is so, thanks to its new optimization algorithm and a tailor-made training dataset, helping to reduce the dreaded "hallucination" problem, which has long plagued AI models.
What's more is that it also uses a "chain of thought" technique, which allows the model to think through multiple prompts before generating a response. This enables more thoughtful and accurate outputs, though it results in slower responses compared to GPT-4o.
What truly sets o1 apart is its ability to go beyond pattern recognition, solving complex problems autonomously. Using reinforcement learning, it not only arrives at answers but also explains the reasoning behind its conclusions, a skill its predecessors lacked.
This capability is perhaps best demonstrated in its stellar academic performance—OpenAI reports that o1-preview performs on par with PhD students in subjects like physics, chemistry, and biology. For example, it ranked in the 89th percentile in Codeforces competitive programming and scored an impressive 83% on the International Mathematics Olympiad qualifier, far outpacing GPT-4o's 13%.
However, there's a catch. As an early-stage model, o1 lacks some of the features that have made GPT-4o so useful, such as the ability to browse the web and upload files. Nevertheless, OpenAI is betting that these trade-offs are worth it for users who need enhanced reasoning capabilities.
The timing of o1's launch couldn't be more critical. OpenAI is in the midst of securing a funding round that will send it to a $150 billion valuation as competition in the AI space intensifies. Rivals like Anthropic and Google are also advancing their models' reasoning abilities, but OpenAI hopes o1's early release will give it an edge in AI reasoning capabilities. The company says, cracking reasoning is an important next step toward human-level intelligence.
The preview version is now available to ChatGPT Plus and Team users, with Enterprise and Edu users set to gain access next week.
