Join our daily and weekly newsletters for the latest updates and exclusive industry-leading AI content.
Perplexity AI, a startup founded by former Google AI researchers Andy Konwinski, Aravind Srinivas, Denis Yarats, and Johnny Ho, is making waves in the web search arena. It combines a web index with real-time information, and a conversational AI chatbot style. Their chatbot, Perplexity Copilot, has mainly used AI models like OpenAI’s GPT-4 and Anthropic’s Claude 2 until recently, which subscribers could switch between.
Perplexity AI has now launched its own AI large language models (LLMs) called pplx-7b-online and pplx-70b-online, reflecting their parameter sizes of 7 billion and 70 billion, respectively. These models are enhanced versions of open-source models from Mistral and Meta. In AI, parameters indicate the number of connections between the model’s neurons, generally meaning higher parameters equate to more powerful and intelligent models.
Perplexity’s new LLMs stand out because they offer updated, factual information, something other leading LLMs often miss. These new models are available for others to use through Perplexity’s API. As stated by Perplexity CEO Aravind Srinivas, these models are the first live LLM APIs grounded with web search data without a knowledge cutoff.
Previously, GPT-3.5 and GPT-4 had limited knowledge up to September 2021, though they were updated recently. For current events, ChatGPT relies on Microsoft’s Bing search. Elon Musk’s xAI is also integrating real-time information into its chatbot Grok, drawing from the X (formerly Twitter).
Other AI companies, like Cohere, also aim to ensure recent knowledge in their models through web browsing and retrieval augmented generation (RAG). Perplexity employs its in-house search and indexing to keep its LLMs updated. Their sophisticated ranking algorithms prioritize high-quality, non-SEOed sites, and provide responses with the most relevant and recent information.
Perplexity hired human evaluators to score its new LLMs based on helpfulness, factuality, and freshness. They compared responses from these models with Meta’s Llama 2 and OpenAI’s GPT-3.5 Turbo. Perplexity’s models performed better than others in freshness and factuality, though GPT-3.5 still led in helpfulness.
The new Perplexity models are now accessible through their API, moving from beta to public availability. Although trained on free, open-source models, Perplexity charges $20 monthly or $200 annually for its Pro subscription, which includes a $5 monthly API credit. Additional API usage costs extra, and pricing details are provided upon request via email.
While the uptake of these new models is yet to be seen, some users already prefer Perplexity over Google search, citing a better overall experience. Venture capitalist Jeremiah Owyang highlighted Perplexity’s innovative approach as a disruptive force to traditional search models.
With Google Bard facing challenges and delays in their next-gen AI, Perplexity has a unique opportunity to position itself as a prominent search alternative, offering conversational AI that retrieves answers from the web directly. Subscribe to stay updated with the latest news.