Divya van Mahajan

The Most Used GenAI Tools

ai-daily-brief-podcast

Study Document: The Most-Used GenAI Tools — AI Daily Brief (2025-08-29)

Overview

This episode of the AI Daily Brief (a daily podcast and video covering AI news) covers two segments. The headline segment reviews three stories: user preference data comparing GPT-4o and GPT-5, researcher departures from Meta’s superintelligence team, and NVIDIA’s Q2 2025 earnings. The main segment provides an in-depth analysis of Andreessen Horowitz’s (a16z) fifth edition of their top 100 Gen AI consumer apps report, exploring usage trends, platform rankings, and what the data signals about the direction of the consumer AI market. The speaker is the host of the AI Daily Brief; no full name is given in the transcript.

Source video URL: not available

Prerequisites

  • Familiarity with major large language model (LLM) products: ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic), Grok (xAI), DeepSeek
  • Basic understanding of AI product categories: reasoning models, image/video generation, vibe coding, AI agents, companionship apps
  • General awareness of AI industry players: OpenAI, Google, Meta, NVIDIA, Anthropic, xAI, ByteDance, Alibaba
  • Understanding of web traffic metrics as a proxy for product usage
  • Familiarity with venture capital firm Andreessen Horowitz (a16z) and their role in tracking the AI ecosystem
  • Basic financial literacy: revenue growth rates, market capitalization, CapEx

Main Points

Headline 1: Blind Testing Suggests Users Prefer GPT-5 Over GPT-4o

  • An anonymous developer (“FlowerSlop” on X) built a blind testing app presenting users with side-by-side responses from GPT-4o and GPT-5 (non-thinking mode), with system prompts forcing short outputs to prevent easy identification.
  • Anecdotal results from users posting on X showed an overwhelming preference for GPT-5; one ML engineer described it as sounding “more like a person and a little more thoughtful.”
  • The app has run hundreds of thousands of tests but does not aggregate results, so conclusions are limited to self-reported individual data.
  • A secondary Reddit thread suggests the GPT-4o model restored after community backlash may not be identical to the original, with users describing it as “mechanical, laconic, and decontextualized.”
  • The broader takeaway: users form strong attachments to specific model versions, and the volume of complaints about a change does not necessarily reflect universal sentiment.

Headline 2: Researchers Depart Meta’s Superintelligence Team

  • Wired reported that two researchers, Avi Verma and Ethan Knight, left Meta’s new superintelligence team after less than a month and returned to OpenAI.
  • A third researcher, Rishabh Agarwal, publicly announced his departure after 7.5 years across Google Brain, DeepMind, and Meta, citing a desire to “take on a different kind of risk.”
  • Agarwal disclosed that during his brief time at Meta the team advanced post-training for thinking models, including pushing an 8B dense model to near DeepSeek R1 performance via RL scaling, using synthetic data mid-training to warm-start RL, and developing better on-policy distillation methods.
  • Meta’s spokesperson framed the departures as normal attrition during aggressive recruiting; the host noted that high-pressure, high-compensation recruitment decisions can feel inauthentic once the initial excitement subsides.
  • The host’s conclusion: the significance of these departures depends entirely on what the superintelligence team ultimately produces.

Headline 3: NVIDIA Q2 2025 Earnings — A Rorschach Test

  • NVIDIA reported 56% year-over-year revenue growth for Q2, with total quarterly revenue of $46.7 billion — a new record — but only a 6% increase quarter-over-quarter.
  • Media framing diverged sharply: Bloomberg emphasized decelerating growth and a “tepid forecast”; TechCrunch led with “record sales as the AI boom continues.”
  • The widest-ever gap between top and bottom analyst revenue forecasts (~$15 billion) indicates significant uncertainty about NVIDIA’s trajectory.
  • Context: NVIDIA had multiple 200%+ year-over-year growth quarters in 2024, making deceleration mathematically inevitable; it remains the only tech company above $1 trillion market cap growing at more than 50% annually.
  • China export restrictions create a structural cap on NVIDIA’s addressable market.
  • Jensen Huang projected $3–4 trillion in AI CapEx over the next five years; Morgan Stanley’s estimate has total AI CapEx reaching $3 trillion by 2029, with 2025 CapEx at ~$445 billion growing at 56%.
  • NVIDIA stock fell 5% in after-hours trading, reflecting a slight market bias toward pessimism.

Main Topic: a16z Top 100 Gen AI Consumer Apps — Fifth Edition

Methodology and Context

  • a16z publishes this ranking approximately every six months (five editions since 2023), using data from SimilarWeb and Sensor Tower.
  • The metric is raw usage/traffic, not revenue; some apps with fewer users may generate more revenue.
  • Compared to prior editions, the current list shows a “settling” — fewer surprising or unfamiliar entries, with category leaders consolidating their positions.

Trend 1: Google Making Big Moves

  • Google placed four products on the web top 50 for the first time: Gemini (#2), Google AI Studio (#10), NotebookLM (#13), and Google Labs (#39).
  • Gemini receives approximately 12% of ChatGPT’s web visits — a significant gap, illustrating ChatGPT’s first-mover advantage and brand recognition.
  • NotebookLM’s sustained ranking suggests durable usage beyond initial hype.
  • Google Labs traffic spiked more than 13% following the launch of video model Veo 3 in May; the host expects this to grow further as Veo 3 moves to its own platform.

Trend 2: Grok’s Rise

  • Grok is now #4 on the web list and #23 on mobile, up from virtually no mobile presence at the end of 2024.
  • Monthly active users on mobile have reached upwards of 20 million.
  • Mobile usage jumped nearly 40% in July following the release of Grok 4, though part of the spike is attributed to the introduction of AI companion avatars, including NSFW options.

Trend 3: The Rise of China

  • 22 of the 50 apps on the mobile chart were developed in China, but only 3 of those 22 are primarily used in China — indicating genuine global export of Chinese AI products.
  • Apps such as Quark (#9, Alibaba), Doubao (#12, ByteDance), and Kimi (#17) drew more than 75% of their traffic from China.
  • Chinese video models have historically led due to greater research focus on video and fewer IP/copyright restrictions in training data; Veo 3 is noted as the first US model to break this trend, trained partly on YouTube data.
  • MeToo alone had five entries on the mobile list (Beauty Plus, Beauty Cam, Wink, Airbrush, and their main photo/video editor), reflecting the dominance of camera-native AI apps on mobile.

Trend 4: Vibe Coding Comes of Age

  • Six months ago, vibe coding tools were just emerging: Cursor at #41, Bolt at #48, Lovable on the “brink” list.
  • Current standings: Lovable is #23 on the web list; Replit is #41; Bolt has dropped to the brink list.
  • The top four vibe coding apps now collectively receive over 50 million monthly web visits.
  • Consumer Edge credit card data shows strong revenue retention — users are not only staying but increasing monthly spend in the months after first use.
  • A secondary traffic layer is emerging: published apps living on subdomains (lovable.app, replit.app) are generating their own traffic — Replit-hosted apps receive 2–3 million monthly visitors; Lovable-hosted creations exceed 10 million monthly visitors.
  • The host expects Lovable and Replit to be in the web top 20 within six months.

Trend 5: Stabilization and Recurring Leaders

  • 14 companies have appeared on all five editions of the web top 50, representing durable consumer AI behavior across categories:
    • General assistants: ChatGPT, Perplexity, Poe
    • Companionship: Character AI
    • Image generation: Midjourney, Leonardo
    • Image/video editing: Veed, Cutout
    • Voice generation: ElevenLabs
    • Productivity: PhotoRoom, Gamma, Quillbot
    • Model hosting: Civit AI, Hugging Face
  • The top four web products are ChatGPT, Gemini, DeepSeek, and Grok.

Trend 6: Manus as a Signal for General Agents

  • Manus, at #31 on the web list, is identified as the only pure-play general-purpose agent platform on the list.
  • Its presence raises the question of whether a market exists for general agents outside of the core foundation model companies.
  • The host flags it as a key data point to watch in the next edition — alongside whether Google’s Project Mariner (agentic browser, currently in Google Labs) moves to its own platform.

Key Concepts

  • Vibe coding: A category of AI-assisted development tools (e.g., Lovable, Replit, Cursor, Bolt) that allow users to build software applications through natural language prompts with minimal traditional coding.
  • Reasoning model: An LLM variant that explicitly works through intermediate reasoning steps before producing a final answer (e.g., OpenAI o1/o3, DeepSeek R1, Grok 4 thinking mode).
  • SimilarWeb / Sensor Tower: Third-party analytics platforms used to estimate web and mobile app traffic, respectively; the data sources for the a16z rankings.
  • Veo 3: Google’s video generation model, launched May 2025, noted as the first US model to outcompete Chinese video generation tools.
  • Project Mariner: Google’s experimental agentic web browser, currently housed in Google Labs.
  • Manus: A general-purpose AI agent platform; the only such platform on the a16z top 50 web list.
  • On-policy distillation: A training technique in which a student model learns from outputs generated by a teacher model during active training, rather than from a static pre-collected dataset.
  • RL scaling: The use of reinforcement learning to improve model performance by scaling compute during the training process.
  • Consumer Edge: A credit card panel data provider used to track actual consumer spending patterns on subscription platforms.
  • AI companion apps: Applications, such as Character AI or Grok’s avatar feature, designed to simulate social or emotional interaction with an AI persona.
  • CapEx (capital expenditure): In the AI context, investment by large technology companies (hyperscalers) in compute infrastructure such as data centers and GPUs.

Summary

The episode uses a16z’s fifth top-100 Gen AI consumer apps report as a lens to argue that the consumer AI market is entering a phase of consolidation after two years of rapid, chaotic growth. ChatGPT retains a commanding lead, but Google has meaningfully re-entered the picture with multiple products on the web list; Grok is growing rapidly; and Chinese-developed tools continue to hold significant global market share, particularly in mobile and video. The most striking structural shift is the maturation of vibe coding platforms — no longer novelties, they now show sustained revenue retention and are generating secondary traffic ecosystems from the applications users build on them. Companionship, image generation, and general-purpose LLM tools have proven durable across all five editions of the ranking, suggesting established consumer behavior patterns. The one category conspicuously underrepresented — general-purpose AI agents — is flagged as the key variable to watch, with Manus’s appearance at #31 as the only current signal that demand for general agents may be forming. The episode’s headline coverage contextualizes this usage picture within broader industry dynamics: model preference is shaped as much by habit and personality attachment as by objective capability; talent markets remain fluid even at the frontier; and NVIDIA’s financial results, while records by any historical standard, are now being judged against the unprecedented baseline of 2024’s hypergrowth.