AI News Flash | AI Search Rising Star "Perplexity" Hits 780 Million Monthly Searches — Next Target: 1 Billion a Week

Hello from the world of AI! 🚀Today's newsletter covers the following hot topics.

• AI search rising star "Perplexity" hits 780 million monthly searches — next target: 1 billion a week
• EleutherAI releases a massive 8TB AI training dataset, combining copyrighted material with public content
• Meta may be making its largest-ever private AI investment — huge talks underway with Scale AI

PERPLEXITY

AI Search Rising Star "Perplexity" Hits 780 Million Monthly Searches — Next Target: 1 Billion a Week

Fast-growing AI search service "Perplexity" says its monthly search volume hit 780 million in May 2025. CEO Aravind Srinivas shared this at the Bloomberg Tech Summit on May 30, with month-over-month growth running at an astonishing 20% or more.Srinivas said that at the current pace, the company could hit 1 billion searches a week within a year — noting that in early 2022, daily searches numbered just 3,000, showing how explosive growth has been.Behind this is Perplexity's next-generation browser, "Comet," aimed at redefining the browser itself and deepening its connection with users.In Comet, search works from virtually any UI — search field, tab screen, sidebar — creating a distinct experience for every user.Srinivas said, "As long as users keep operating inside our own browser, there's almost no limit to extending their time on it." The search field, new tabs, side menus, and any page browsed are built to prompt further searches — a new touchpoint for people tiring of browsers like Chrome, and a likely growth driver ahead.

Source: Perplexity

Comet, he said, is built around a vision beyond a mere search tool — realizing "proactive AI." A typical AI response aggregates multiple searches; but if the AI could judge and act directly, one instruction could complete an entire browsing chain.That requires rethinking browsing from the ground up, not building on a conventional browser — "processing must be coordinated across server and device." The very idea of a "browser" needs redefining, he argued.Comet goes beyond the conventional browser category, and Perplexity positions it as a "cognitive operating system.""Comet is always beside the user, supporting daily life and work, and can act autonomously on the web when needed," he explained.He added: "The relationship between people and the internet is fundamentally changing. We used to 'use' the internet — now we 'live inside' it. Digital space has woven itself into our lives. True personalization needs something present with the user in every context — that's why the browser must be rebuilt."Comet's details remain largely under wraps, but Srinivas has said it's designed to track behavior outside the app too, enabling targeted ads — echoing Google's past ad-network strategy.No release date for Comet has been announced, but Srinivas posted on X that it would launch "soon."

ELEUTHER

EleutherAI Releases a Massive 8TB AI Training Dataset, Combining Copyrighted Material with Public Content

【Reuters, June 2, 2024】The Wall Street Journal reported, citing people familiar with the matter, that Meta Platforms (NASDAQ: META) is working to let brands freely produce ads and choose their target audience through its own AI systems by the end of 2025.Meta, which runs social apps used by more than 3.4 billion monthly users worldwide, already offers AI features that automatically optimize ad creative, generate background imagery, and adjust video formats — letting advertisers run high-performing campaigns efficiently.According to the report, brands need only provide a product image and a budget; Meta's AI then automatically creates the ad assets (images, video, text) and proposes optimal targeting and spend for placement on Instagram and Facebook.The methods used to collect AI training data haven't changed because of copyright lawsuits — but companies and researchers have grown noticeably more reluctant to disclose information. EleutherAI's Stella Biderman pointed out this chilling effect in a piece published on Hugging Face on May 31."Some researchers working on data-related research say they can't publish their findings for fear of litigation risk," she wrote.

Source: GettyImages

EleutherAI's "Common Pile v0.1" is now available on Hugging Face and GitHub, built with legal counsel to ensure lawful data use throughout. The dataset includes more than 300,000 public-domain books digitized by the Library of Congress and the Internet Archive.It also draws on OpenAI's speech-recognition model "Whisper" to transcribe audio into text as additional training data.According to EleutherAI, two models trained on just a small portion of Common Pile v0.1 — "Comma v0.1-1T" and "Comma v0.1-2T," each with 7 billion parameters — matched the performance of Meta's original LLaMA model across programming, image recognition, and math benchmarks.AI research group EleutherAI has released a new, legally cautious dataset, "Common Pile v0.1," on Hugging Face and GitHub. Built with legal guidance, it draws mainly on more than 300,000 public-domain books collected and published by the Library of Congress and the Internet Archive — supplemented by text generated from audio using OpenAI's speech-recognition model, "Whisper."Two models built from this data, "Comma v0.1-1T" and "Comma v0.1-2T" (7 billion parameters each), matched Meta's early LLaMA model on various AI benchmarks despite being trained on only a portion of Common Pile v0.1.EleutherAI's Biderman rejects the conventional view that "high-performing AI can't exist without unlicensed data," instead emphasizing the potential of public-domain and openly licensed data.The release of "Common Pile v0.1" is also being seen as something of an atonement, given the controversy over "The Pile," EleutherAI's earlier dataset that included copyrighted material.Going forward, EleutherAI says it plans to deepen ties with research institutions and infrastructure providers to release lawful open datasets more frequently.※Update at 9:48am Pacific Time: Biderman revealed on X (formerly Twitter) that, besides EleutherAI, several institutions including the University of Toronto also collaborated on this dataset and the models.

META

Meta May Be Making Its Largest-Ever Private AI Investment — Huge Talks Underway with Scale AI

According to Bloomberg, Meta is negotiating a potentially massive investment — possibly exceeding $10 billion — in AI startup Scale AI. It would be Meta's largest AI-related investment to date, and one of the biggest private funding deals in tech.

Source: GettyImages

Scale AI, built up under CEO Alexandr Wang, provides data-labeling services to companies including Microsoft and OpenAI. This step — essential for training AI models — is often carried out by external contract staff. The US Department of Labor previously investigated the company over the fairness of its labor contracts, though the inquiry didn't lead to legal action.In 2023, revenue reached $870 million, and 2024 is expected to bring growth to $2 billion. Meta already participated in the company's Series F round, which valued Scale AI at $13.8 billion.Scale AI has also been involved in developing "Defense Llama," a military-focused LLM built on Meta's "Llama 3," deepening the technical ties between the two companies.

In Closing

We hope today's content helps you put AI to work and stay on top of the latest information.If you found this useful, feel free to share it with interested colleagues or friends.We'll keep curating the topics that matter most.Watch for our next newsletter.Thanks for reading.