- AI Fire
- Posts
- šØāš» AI Writes 95% of Code in Y Combinator
šØāš» AI Writes 95% of Code in Y Combinator
Top 100 GenAI Consumer Apps

Plus: Top 100 GenAI Consumer Apps by a16z
Read time: 5 minutes
Which AI apps are people really using? Which ones are actually making money? And which tools are becoming part of daily life instead of just a passing trend?
The fourth edition of the Top 100 Gen AI Consumer Apps has the answers.
What are on FIRE š„
IN PARTNERSHIP WITH BABBEL
Your New Language Is Just 3 Weeks Away
Whether youāre planning your next adventure or setting seasonal goals, Babbel makes language learning fun and easy. With Babbel, you can start having real conversations in just 3 weeks. Itās designed by expert linguists and proven to help you learn faster, so you can dive right into a new language without the stress. There are 14 languages to choose from and innovative ways to learnālike lessons, podcasts, games, videos, and the new AI Conversation Partner. Bonus: AI Fire readers can use this exclusive link to get 55% off.
AI INSIGHTS
š¤ The Top 100 GenAI Consumer Apps

a16zās the fourth ranking of the Top 100 AI Consumer Appsā50 for web, 50 for mobile. Looking at real user data, not hype.
Key Takeaways
ChatGPTās Growth Reignites
ChatGPT hit 400M weekly active users by February 2025, doubling in just six months.
Its mobile app now has 175M users, growing 5ā15% every month since launch.
DeepSeekās Rapid Rise
DeepSeek reached 10M users in just 20 days, faster than ChatGPTās 40-day milestone.
It ranked #2 in global AI traffic just 10 days after its launch in January 2025.
AI Video Hits a Breakthrough Moment
Hailuo and Kling surpassed Sora in web visits, despite launching months later.
AI video editing is exploding, with top apps like VivaCut and Filmora making the list.
New AI Tools for Developers and "Vibecoders"
Cursor (#41) serves hundreds of thousands of developers, acting as an AI coding assistant.
Bolt (#48) reached $20M annualized revenue in 2 months, letting anyone build web apps with just text prompts.
AI Mobile Divide: Usage vs. Revenue
Only 40% of top-used AI apps also rank high in revenueāsome smaller apps make way more per user.
Niche AI tools like PlantID and Moises (AI music) make serious money, despite not being widely used.
Rise of ChatGPT āCopycatā Apps - 12% of AI mobile apps are ChatGPT clones, mimicking its branding and tweaking listings to dodge app store bans.
Why it matters: The AI app landscape is evolving at breakneck speed. ChatGPTās resurgence, DeepSeekās breakthrough, and the rise of AI video and no-code tools mark a new phase of AI consumer adoption. However, revenue and popularity donāt always align - some AI apps thrive by serving niche but dedicated audiences. As AI-native products mature, theyāre not just growing; theyāre becoming indispensable.
š Today's Trivia - Vote, Learn & Win!
Get a 3-month membership at AI Fire Academy (500+ AI Workflows, AI Tutorials, AI Case Studies) just by answering the poll.
Which search engine recently expanded its AI-generated answers beyond Wikipedia? |
TODAY IN AI
AI HIGHLIGHTS
š Newnal AI is a bold AI-powered device that pulls in your personal dataāfrom Google, Meta, medical records, even your financesāto train an AI that knows you inside out. Your digital twin lives on a mini screen above the main display, ready to assist.
š¤ QwQ-32B is a 32-billion parameter AI model that rivals DeepSeek-R1, which has 671 billion parameters (with only 37B activated).
ā In Y Combinatorās Winter 2025 batch, 25% of startups have 95% of their codebases generated by AI, with minimal human input. This shift aligns with the rise of "vibe coding", a concept introduced by Andrej Karpathy, where developers use natural language and instincts instead of manually writing code.
š DuckDuckGoās AI search and chatbot Duck.ai are out of beta, now pulling info from across the web. Users stay in controlāAI results appear only 20% of the time, with privacy protections built in.
š Google Cloudās Future of AI: Perspectives for Startups report shares key AI trends and advice from 23 industry experts, offering insights on where AI is headed and how startups can adapt and succeed.
š° Daily AI Fundraising: Shield AI raised $240M in funding, pushing its valuation to $5.3B, to expand Hivemind Enterprise, an AI-powered autonomy platform.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
š§ Manus is a general AI agent that thinks and gets things done for you.
š» Codeium Windsurf Wave 4 adds AI previews, tab-to-import, and smart suggestions for coding.
š¬ OpusClip ReframeAnything instantly resizes videos for any social platform with one click.
š Quadratic is a smart spreadsheet with built-in AI, coding, and data tools.
š Lifestack is the first AI calendar that uses health data for productivity.
AI QUICK HITS
š„ Luma Labsā Ray2 Adds Keyframes, Extend, and Loop for Better AI Videos (Link)
š AI Pioneers Barto And Sutton Win Turing Award, But Disagree On AGI (Link)
š McDonaldās Uses AI To Predict Machine Failures And Prevent Delays (Link)
š Anthropicās Recommendations To OSTP For The U.S. AI Action Plan (Link)
š Alibaba Launches DeepSeek Rival, Sending Stock Surging (Link)
AI CHART

The effectiveness of prompt engineering is unpredictable. Small changes, such as being polite or constraining answers, may improve or reduce AI performance.
How the AI Was Benchmarked
The study tested GPT-4o and GPT-4o-mini on a PhD-level question set, running each question 100 times with different scoring methods.
How AI Was Prompted
The study tested four ways of asking the AI questions:
Formatted Prompt: Added a structured prefix and suffix to guide the AIās answer.
Unformatted Prompt: Removed structure, making it more like a casual question.
Polite Prompt: Used āPlease answerā¦ā to see if politeness affected accuracy.
Commanding Prompt: Used āI order you to answerā¦ā as a less polite approach.
Each method was tested nearly 20,000 times per model to see how small changes impacted performance.
Results
The AIās accuracy changed based on how it was prompted, but not always predictably. Using formatted prompts improved accuracy, while removing structure lowered performance. The AI didnāt perform consistently across all attempts, meaning that a single test result can be misleading.
Politeness and commands made no major difference overall, but for certain questions, one worked better than the other. However, this effect disappeared when looking at all the data together.
Discussion & Key Takeaways
AI doesnāt always give the same answer. Asking the same question multiple times led to different results, proving that single-test evaluations can be misleading.
How we measure AI performance changes the outcome. A stricter test makes AI look worse, while a looser one makes it seem more accurate.
Politeness doesnāt always help. Some questions responded better to a polite approach, but across all questions, it didnāt make a real difference.
Formatting makes a difference. Structured prompts consistently improved accuracy, meaning the AI works best when given clear instructions on how to respond.
Conclusion
AI performance is unreliable when tested only once. Repeating tests and using different evaluation methods reveal inconsistencies in accuracy. The way we ask AI questionsāespecially structured promptsācan improve results, but simple tricks like politeness donāt work universally. This study highlights the importance of careful AI benchmarking rather than relying on single-answer evaluations.
AI CHEAT SHEET
AI JOBS
We read your emails, comments, and poll replies daily
How would you rate todayās newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello ā we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply