Digest #544 | Building with AI? Get Pro and join the community
Hey folks,
yesterday afternoon was completely free so i decided to record a head to head of; Bolt, Replit, Cursor and Windsurf. Gave them all the same initial prompt and let the battle commence. Video and post - see who won 👀 (the video + post are free)
TLDR; inside today’s newsletter
ChatGPT Ghibli-fies the entire Twitter
Google jumps to Gemini 2.5, skips 2.0 Pro
words from today’s sponsor: Salesforce
Bolt vs Replit vs Cursor vs Windsurf — who wins my new test?
short course - vibe coding 101
tools and news you shouldn’t miss
Want to partner with us? Click here.
🔎 What’s Trending
ChatGPT just upgraded its image game. With the new update, GPT-4o, the default model in ChatGPT can create native images (i.e. without relying on its dumb old partner, DallE-3). It’s a single smart model that reads your request and creates an image. Paid users already have access, and it will come to free users soon.
Native image generation finally solves editing by prompts. Give ChatGPT an image and edit it into something else with plain English. You could create really good-looking ads in just minutes (or you can join everyone and turn your pictures into a Studio Ghibli anime). OpenAI has also relaxed the content moderation filters a lot for this new mode—making it more useful.
Google released a new model — Gemini 2.5 Pro. This new experimental model tops most benchmarks, including coding. It’s a thinking model, and you can’t turn its thinking off. The best thing about this is that Gemini Advanced users can use it in the app (along with the developer access via Google AI Studio and APIs). In my testing, it feels smarter than o3 mini high and Sonnet thinking.
You can now create custom modes in Cursor. What’s that? You can define how the AI needs to work and what tools it can access to suit your workflow. Example modes could be “Teach” or “Refactor” mode. Bolt also added a new Discussion Mode that eats fewer tokens and helps you clarify your idea alongside building it.
AI agents will soon be the new UI – easily orchestrating tasks – with apps taking a backseat. Forward-thinking business leaders must prepare for this new reality. Many companies have already discovered the value of AI agents. Dive into this article by Salesforce SVP Product, AI Platform for more.*
*sponsored
👀 Head-to-head battle among 4 AI building tools
I had a random idea to build a web-based drum machine using AI coding tools. So, I thought it would make for a perfect head-to-head comparison.
I pit four different AI coding tools against each other - Replit, Cursor, Bolt, and Windsurf - to see which could best help me create the best site the quickest.
Let's dive into what happened.
The challenge
The prompt was the same across all tools:
I want to build a website where users can play one of those DJ drum kits that they press buttons on, and I want it to go from really basic level, so maybe there's only a few options, a few buttons to press and a few knobs, all the way up to, like, really extreme and make that really crazily hard.
No technical specs. No design requirements. Just a simple concept with difficulty progression that left plenty of room for each AI to interpret and exercise its creative muscles. A perfect little test that should be straightforward enough for any decent AI coding assistant, right?
I set up with all four tools running simultaneously:
Replit (using Agent v2)
Bolt
Windsurf (first time using it)
Cursor (with the latest update and custom modes)
Four contenders enter. Only one leaves with the crown. May the best AI win.
The early results
I fired the prompt across all four tools simultaneously and settled in to watch the AI programming battle unfold.
Bolt
Bolt jumped ahead initially with a basic interface, but I quickly ran into issues. While some pads worked (the kick and snare specifically), others remained silent. The volume controls were particularly problematic - practically inaudible despite being at maximum level. After some unsuccessful troubleshooting attempts, Bolt became the first casualty in my experiment.
🧑⚖️ Verdict: Eliminated early for critical functionality failure.
Windsurf
Windsurf also struggled with similar audio issues. Despite multiple attempts to fix the non-functioning pads, I couldn't get a complete working set. It took longer than Bolt to reach the same basic functionality, and ultimately hit the same roadblocks. Windsurf joined Bolt on the sidelines.
🧑⚖️ Verdict: Eliminated after extended troubleshooting failed to produce working audio.
The finalists: Replit vs. Cursor
With two tools eliminated, it came down to Replit and Cursor.
Read the results in the full post and watch the video for free here →
⚙️ Tools and News
Stop Searching. Start Building: Open-Source AI Workflows that Work - blueprints.mozilla.ai*
xymake - Turn your X threads into markdown text. This takes in any thread and converts all the posts and replies into a single piece of text.
Microsoft is launching two new reasoning-based agents in Copilot, starting from April. Researcher tries to mimic Deep Research, and Analyst is the new kid on the block with o3-mini powering its brain.
Falcon combines deep research and sales. Instead of random topics, it uses agentic search for finding all there is to find about your target account (or any company).
Agent Teams by Agno - Combine multiple specialized agents, each focused on a different aspect of the problem, to work reliably for a single goal.
OpenAI now accepts MCPs in its Agents SDK. Cloudflare is also on the MCP bandwagon, it helps you build and deploy remote MCP servers. I found this directory of top MCP servers, and I appreciate that it doesn’t make MCPs feel overwhelming.
DeepLearning AI and Replit have partnered to create a short course (<2 hrs) to teach us vibe coding 101 with Replit. Save it for the weekend and get your hands dirty.
Amazon is now building personalized shopping and health assistants using GenAI. Btw, did anyone get their hands on the new Alexa Plus? How does it do?
Portl generates viral TikTok hooks for your product, so your content doesn’t get ignored. It claims it as learned this after analyzing 50k+ viral hooks.
In side quests by Big Tech, Google is finding similarities in how our brains and LLMs store information.
Our free newsfeed for more →
*sponsored
That’s it for today. Feel free to hit reply and share your thoughts. 👋
Enjoy this newsletter? Please forward to a friend.
Building with AI? Get Pro, join our Slack and connect with fellow builders.
Want to advertise in this newsletter? Click here.