"It also shows why smaller AI start-ups are relevant despite the headline dominance of Open AI. Focused companies create better products. And the organizational separation of product and foundation model gives these start-ups more leeway than Open AI, Google or Microsoft. The biggest challenge, however, remains that potential users and customers discover the product and then have to change their habits.
It was a bang. The young search engine start-up Perplexity receives $75 million in a new round of financing alongside IVP from, among others, Amazon founder Jeff Bezos and many others such as Shopify CEO Tobias Lütke, ex-Github CEO Nat Friedman and Angel-List- Co-founder Naval Ravikant. The financing round reads like a who's who of the US tech industry.
Perplexity, which was only founded in August 2022, is valued at $500 million in this second round of financing in its young existence. According to sources at The Information, Perplexity has about 10 million monthly users. Sales are $3 million per year.
These numbers are still manageable, but should not hide the potential of Perplexity. Far more than Bing's Copilot or Open AI's GPT-4 with Websearch, Perplexity shows how Large Language Models are revolutionizing web search.
This is how perplexity works
The company uses the licensable Bing index, i.e. the websites indexed by Bing, and its own index infrastructure, which is not publicly specified. When a user searches with Perplexity, the search query is passed to Copilot.
The Copilot, based on GPT-4 and Claude 2, analyzes the search query, asks any clarifying questions and often generates several, slightly reformulated searches, which are then passed on to the Bing index in parallel. Think of Copilot as a search expert who tries to understand the user's intent and then enters several, better-formulated searches into the search engine.
After this intermediate step, the most relevant results are then passed on to the LLM, which synthesizes an answer from them. In the free version, this answer provides GPT-3.5. Paying customers can choose between GPT-4, Claude 2, Google's Gemini and Perplexity's first proprietary model.
Perplexity forgoes the chat interface with speech bubbles and instead tries to produce a structured results page. This interface decision alone leads to a big difference in everyday use compared to Bing's Copilot. This is despite the fact that Bing also offers similar functions, such as automatically generated follow-up questions that are relevant to the topic.
Perplexity has continuously increased the range of functions in 2023. You can search for images and videos. Paying customers can have images generated from the search results page. A function that relies on Dall-E3 in the background and, among other things, specializes in illustration images for publishers. If the paying user is unsure whether the chosen LLM really synthesized the best result, he or she can have the result rewritten using a different model. You can upload documents for analysis or start searches with an uploaded image.
This is where the advantage comes when all major models are available under one roof. The image-based search runs on GPT-4 Turbo with Vision, while the document analysis is best carried out with Claude 2 because this model currently has the largest "memory" and can therefore also analyze entire books.
For power users, Perplexity also offers additional functions to personalize the search. Users can enter information about themselves in their profile, which will be included in every search. Folders can be created for projects. You can not only organize your own searches in these folders. A prompt can also be added to each folder, which will be included in every search along with the search query. This makes it easy to implement very different ways of using Perplexity. However, this enormous setup effort is not relevant for the mainstream.
Perplexity also has different focus types that correspond to specialized searches: Academic only searches academic papers, Wolfram Alpha can be targeted directly, and Perplexity searches can also be limited to YouTube or Reddit. The “Writing” focus stands out: This switches off web searches and allows Perplexity to be used for text production like the classic ChatGPT.
Additional types of focus are likely to drive Perplexity further this year. The first Pro users reported about being temporarily free in the associated Discord forum switched, new search focuses: Yelp, Shopping, Shopify, Klarna, News, Product Opinions and others. This opens up obvious new revenue streams for Perplexity, affiliates at Shopify, for example, and direct collaborations like Klarna. The potential of a dedicated shopping search in the style of Perplexity is also great.
The author has been using Perplexity since 2022 and has been a paying customer since May 2023. Perplexity, like any LLM product, has the problem that information is not always correct. You have to be aware of this fact when using it and always check information if in doubt. However, the prominent indication of the linked sources makes this easy. While the classic Bing looked virtually identical to Google, the Perplexity search differs fundamentally from classic search engines in terms of how it looks and works.
Perplexity is the first real challenger for Google because it breaks new ground and thus achieves a new level of quality in web searches. More than GPT-4 with Websearch or Bing's Copilot, Perplexity shows the enormous new space of options that has also opened up in web search with LLMs.
What we can learn from Perplexity about the AI sector
Perplexity showed two important aspects as early as 2022. First, that LLMs are not structurally synonymous with the continuous text black box ChatGPT. ChatGPT also felt like magic at the end of 2022 because Open AI refrained from citing sources, even if they could be viewed. Perplexity and the other LLM-driven web searches like Bing's Copilot thrive on the fact that they synthesize the result, but the sources of their information can be checked by the user. Because they are stated.
The second aspect we learn about the AI sector at Perplexity is the division of labor. With GPT-4, Open AI is still the market leader and quality leader for LLMs. But thanks to the special relationship with Microsoft, GPT-4 is available via Azure to the entire economy via an interface for a user fee. This also applies to Anthropic, whose Claude 2 is available on AWS. This means that companies like Perplexity can use the best models available on the market to build their own products on top of them.
And what's more, Perplexity and Co. can combine the models from the different model suppliers. And: You can also combine these proprietary models with your own and third-party open source models. In November 2023, Perplexity introduced two of its own models, built on the basis of Metas Llama and Mistral from Paris. Perplexity cleverly calls these web search-focused models 'Online LLMs'.
In contrast, everything in Open AI's offering is based solely on the Open AI models. This is not a problem as long as Open AI is at the forefront when it comes to the quality of the models. You can see this on Google right now. Everything Google builds with AI is of course based on Google's own AI models. As long as Google's Gemini is not at the forefront in terms of quality, Google's AI-native products will not be ahead either.
The organizational separation of product and foundation model gives start-ups in the current cloud computing environment of LLMs the structural advantage of being able to select and combine the best models. The disadvantage of this licensing is of course the associated dependence on the model suppliers when it comes to model properties and cost structures. Here too, Perplexity shows what companies can do about it. They build their own models in parallel in the hope of being able to reduce the dependency little by little.
The biggest challenge for Perplexity is also the biggest challenge for the AI sector as a whole. AI-native services often offer superior quality to established services. But this quality often only becomes apparent to users when they get used to a different way of solving the respective task. This necessary change in behavior, and at the company level the necessary process change, is the biggest hurdle to using AI. The costs of training and inference, which mean that these services cannot be offered free of charge and financed by advertising, do the rest; that is, if you're not Microsoft. As with Perplexity, this cost structure always means that the best version of the product is behind the payment barrier. And it is still quite high, as is usual with consumer AI today thanks to the high costs: 22 euros per month or 229 euros annually. Very few customers jump over this very high hurdle.
Perplexity beats Google in almost all types of web searches. Often even by dimensions. Perplexity also allows new types of searches that are not yet possible with Google or even Bing's Copilot are. This is the advantage of start-ups over corporations. The entire organization is focused on one product and works to implement the best customer experience there. Too few people know how good the product is as a result of this work. And nothing will change in the foreseeable future.
Like any new type of product, Perplexity also fights against the usual final enemy: customer inertia, which plays into the hands of the established competition. Habits don't change overnight. Even with AI, the classic challenges of young companies with new products remain the same." [1]
1. Europa übernimmt Führung bei Investitionen in Climate Tech. Frankfurter Allgemeine Zeitung (online) Frankfurter Allgemeine Zeitung GmbH. Jan 9, 2024. Von Holger Schmidt
Komentarų nėra:
Rašyti komentarą