In a bold challenge to the status quo, Perplexity AI, a nascent startup founded by ex-Google AI virtuosos—Andy Konwinski, Aravind Srinivas, Denis Yarats, and Johnny Ho—has unfurled a compelling contender that could potentially overshadow Google Search. The linchpin of their audacious move is the fusion of a web index and real-time data with an engaging AI chatbot interface, epitomized by their chatbot, Perplexity Copilot. Unlike its predecessors, which hinged on existing AI models like OpenAI’s GPT-4 and Anthropic’s Claude 2, Perplexity Copilot now steps into the limelight with its own proprietary AI large language models (LLMs), denoted as pplx-7b-online and pplx-70b-online.
These LLMs, distinguished by their colossal parameter sizes of 7 billion and 70 billion, emerge as fine-tuned iterations sourced from Mistral and Meta’s open source models, mistral-7b, and llama2-70b, respectively. Parameters, denoting the connections among artificial neurons, serve as a yardstick for a model’s potency and intelligence, with larger parameters correlating to enhanced knowledge and performance.
Why do Perplexity’s new online LLMs stand out, and what sets them apart from counterparts like ChatGPT and others? Notably, they extend beyond being mere tools for customization; they aspire to deliver “helpful, factual, and up-to-date information”—a realm where prevailing LLMs, such as GPT-3.5 and GPT-4, have historically faced challenges.
Perplexity’s CEO, Aravind Srinivas, asserts that their PPLX LLMs are trailblazers—the first live LLM APIs infused with web search data and devoid of any knowledge cutoff. In stark contrast, competitors like GPT-3.5 and GPT-4 have been shackled by knowledge cutoffs, with their last update restricted to September 2021, albeit recently.
The quest for real-time knowledge in LLM chatbots intensifies, with Elon Musk’s xAI flaunting Grok, promising seamless integration with sibling company X’s real-time data. On another front, companies like Cohere leverage web browsing capabilities and retrieval augmented generation to infuse recent knowledge into their LLMs.
PPLX online LLMs distinguish themselves further through Perplexity’s proprietary approach. Their in-house search, indexing, and crawling infrastructure act as a dynamic conduit, augmenting LLMs with the most pertinent and current information. A sophisticated ranking algorithm ensures prioritization of high-quality, non-SEOed sites, bolstering the LLMs with snippets—concise website excerpts for generating responses with the latest information.
To validate the prowess of their LLMs, Perplexity conducted evaluations where human contractors assessed responses based on criteria such as helpfulness, factuality, and freshness. The results, gauged through an Elo scoring method, indicated that PPLX models excelled in freshness and factuality compared to both GPT-3.5 and raw Llama 2 models. However, GPT-3.5 retained an edge in terms of perceived helpfulness.
For those eager to harness the potential of PPLX online LLMs, Perplexity’s API is now open for use, transitioning from beta testing to general public availability. Notably, the integration of search and web indexing technology in these models comes at a cost. While the base models are trained on free, open source models, the addition of search and web indexing tech incurs a subscription fee of $20 USD monthly for the Pro tier or $200 annually. Users gain a $5 monthly credit for the Perplexity API, providing access to PPLX models.
As Perplexity stakes its claim as a potent alternative in the realm of web search, the timing seems opportune, especially with the reported hiccups in Google’s offerings. With Google Bard facing controversies and Gemini, the purported GPT-killer, encountering delays, Perplexity’s innovative approach could indeed herald a new era—a vision where AI assistants seamlessly interact and draw answers from the web, redefining the user experience in the landscape of search.
Software Engineer | AI Tech Enthusiast | Content Writer