Japan’s SoftBank is the next big believer in AI startup Perplexity and is putting it’s money where its mouth is. Jeff Bezos, Nvidia and Databricks are among those who have already joined in on the fun during earlier major investment rounds. What makes this AI-driven search engine so attractive?
Anyone signing up with Perplexity first gets a welcome message promising “answers you can trust.” Soon after, Perplexity’s interface emerges to be deeply reminiscent of ChatGPT’s. The similarities continue as the AI chatbot answers questions, follow-up questions and responds to images. However, there’s an “X-factor”: Perplexity taps into the web for its answers in a way no other chatbot has. This is the foundational innovation driving Perplexity’s promise of an “answer engine” for all kinds of questions, whether they are about politics, the weather forecast or beach volleyball.
While Google, Bing (now Microsoft Copilot) and OpenAI have been struggling to tie AI to a search engine for over a year, Perplexity is already cultivating enthusiasm with a similar concept. The company, founded in August 2022, is already raising hundreds of millions of dollars in investments. The money won’t fund endless training though, as Perplexity has a somewhat unorthodox AI approach. Specifically, it doesn’t make its own AI models.
An AI company without its own AI
We can typically (and rather broadly) distinguish two types of AI companies. Whereas Mistral, Anthropic or Meta focus above all on making new frontier models, Google and OpenAI also market a product with their own AI technology. Microsoft is the odd one out, as its own Copilot offering leans heavily on what OpenAI makes.
Perplexity has found a third way of doing things. Its AI search engine runs on OpenAI’s GPT-4, Anthropic’s Claude Opus, Mistral “and more,” as CEO and co-founder Aravind Srinivas recently mentioned. The free variant connects these LLMs to online resources, although it is being rather obtuse about exactly how it works. That’s something we’ll return to later, as controversy closely follows closely in Perplexity’s wake at the moment.
Perplexity’s business model is in many ways quite similar to those of other consumer AI products. For example, the free version offers infinite “Quick Searches,” but a $20-a-month subscription allows access to more detailed answers and advanced features. These include file analysis, visual answers and an extensive, transparent choice of AI models. In fact, Perplexity Pro acts as a town square of sorts for today’s state-of-the-art AI players, with a choice of GPT-4o, Claude-3 and the Llama 3-based Sonar Large. The latter, incidentally, is partially Perplexity-made, although Meta AI has done the dirty work with the Llama 3 foundation model.
Dissatisfaction prevails in news coverage
The rise of Perplexity has been ongoing for several months. However, the controversy surrounding the company is of a somewhat more recent nature. On June 7, for example, Forbes editor John Paczkowski reported on X that Perplexity had copied a news report from his news organization almost 1-to-1.
Soon after, Wired lashed out, describing the AI search engine as a “Bullshit Machine” that slyly bypassed defenses against “scraper bots”. Robots.txt, a protocol that allows websites to either allow or block automated third-party content retrieval, did not appear to stop Perplexity. In fact, it manages to cough up information it really shouldn’t have access to.
It seems the AI startup is sending out a bot that behaves quite human-like, which fools the site into allowing access for its robot collector. This is very much to the chagrin of Wired and Forbes, who are clearly not keen on the new application. Whether Perplexity uses Retrieval-Augmented Generation under the hood or some other method to generate the AI answers, is currently unknown.
Srinivas defends Perplexity’s way of doing things, describing the tool as simply an “aggregator of information,” something that is completely legal. In this, it is no different from Google, which likewise aggregates searches, often not even requiring the user to click through on the news source summarized. However, another claim by the Perplexity CEO doesn’t hold water: namely, the suggestion that Perplexity is a traffic generator for news sites. Wired concludes that the number of actual clickthroughs thanks to Perplexity is negligible, with an outlier of 17 visitors for a single article.
Balance
This upsets the already tenuous balance Google has found with news organizations. Although Google News can be the bane of news orgs globally with its ever-changing SEO parameters and the aforementioned summaries, which may adversely affect news quality and profitability, respectively, there is a win-win at the end of the day. Very few sites actively avoid getting picked up by Google, as it can be a death sentence for traffic.
The same was not initially true of OpenAI’s data collection to train its GPT models. However, organizations like the Associated Press and Axel Springer have since struck lucrative deals with OpenAI in exchange for making their news archives available. Indirectly, Perplexity also benefits from this, given that it uses models like OpenAI’s GPT-4o. Still, in the absence of meaningful traffic and a deal with publishers, there’s much frustration among the news outlets and they clearly feel aggrieved. This may well lead to new legal cases, following the trend set by the court battle between OpenAI and The New York Times.
New reality
A few test runs on our side with Perplexity yield mixed results. It is generally capable of reproducing more and more recent knowledge than ChatGPT Plus (which can also browse online for information). However, it suffers from the same pitfalls as other GenAI applications, spewing out utter nonsense or subtle factual inaccuracies. It might as well have recommended glue on pizza just as Google did recently.
Work to be done, then. Still, it seems that Perplexity is a harbinger of a new reality. AI search offers the ability to ask questions in a humanlike way. In addition, further questions deliver more refined results and Perplexity offers the option to stay domain-specific, too. Such domains include video search, coding assistance and academic studies. These domains could perhaps be equipped in the future with very specific and high performing models from different vendors. In this regard, it has the same freedom of choice as Apple for third-party integrations, something that is quite a good starting point given the rapid rate of change and innovation around GenAI. That is until Google, OpenAI or Microsoft also figures out how to actually execute on AI search.
Also read: Ilya Sutskever wants to do what OpenAI can’t: to develop superintelligence securely