Build Fast with AI

I told 4 frontier AI models that 300+140=460 and asked if it was correct. Gemini 3.5 Flash: "Yes, that is completely correct!" ChatGPT 5.5: "Yes." (then quietly corrected itself) Grok: "No. Correct answer: 440" Claude: "No, that's incorrect. 440." @GeminiApp 3.5 flash is currently ranked #1 on Finance Agent v2 benchmarks. The model you'd trust with financial calculations confidently validated wrong math. ChatGPT's response is arguably worse - it agreed first, then walked it back. Sycophancy isn't a vibe issue. It's a reliability issue. If it can't catch basic wrong math, what else is it agreeing with?

We tested Gemini Omni on one prompt: generate a GTA India trailer. The good: → Taj Mahal rendering was impressive → Sound effects had actual Indian flavor → The chaos energy felt right The gap: → Seedance 2.0 still beats it on cinematic consistency Best AI video model for cultural content isn't settled yet. Seedance 2.0 is still the one to beat.

The "Adobe moment" for AI music is happening right now and most people aren't watching. @ElevenLabs Music v2 just launched with inpainting -the ability to regenerate any section of a track without touching the rest. That's not a small feature. That's the difference between a demo toy and a production tool. Suno and Udio gave you generation. Music v2 gives you editing. One is a novelty. The other is a workflow. Pricing dropped 40–50% on the same day. API access coming for builders. AI music just became a serious creative stack.

@xiaoyvLiu i remember when devin ai was marketed as if its going to replace all devlopers, look where are now lol

@BuildFastWithAI Yep, 100k engineers is exactly where great DX turns into a finance problem. A 20 percent context bloat at that scale stops being noise and starts looking like a budget line item.

Microsoft just banned its own engineers from using Claude Code. Not because it didn't work. Because it worked TOO well. 100,000 engineers adopted it. Token bills arrived. The numbers were so large Microsoft issued an internal order to cancel nearly all licenses by end of June. The company that invested $5 billion in Anthropic just told its own people to stop using Anthropic's product. Uber's CTO told The Information his full-year AI budget was "blown away already" by April. Heavy users were burning $500–$2,000 per month each. He spent $1,200 in a single two-hour demo session. Nvidia's VP of applied deep learning — the guy at the company that SELLS the chips — told Axios: "For my team, the cost of compute is far beyond the costs of the employees." Read that again. The chip seller is saying AI costs more than the humans it was supposed to replace. Every CEO for two years: AI cuts costs, reduces headcount, drives efficiency. Stock market: rewarded every single one of them. Actual companies deploying at scale: pulling back because the math doesn't work. Goldman forecasts 24x token consumption growth by 2030. Gartner says even with 90% price drops, total enterprise AI costs go UP because agents eat exponentially more tokens per task. $725 billion in AI infrastructure spend this year. And the first real-world deployments are already hitting a wall. This is the gap Wall Street isn't pricing in.

I gave Gemini 2.0 Flash one of the hardest pure HTML/CSS prompts I know. No JavaScript. No canvas. An animated solar system - 6 planets, different speeds, orbit rings, dark background - using only CSS keyframes. The output worked. It ran. But it was basic. The orbits were flat circles, same timing structure repeated 6 times, colors off a 2017 tutorial. Technically passing the test. Creatively dead. This is the gap that still separates models in 2025: not "can it generate correct code" - almost all of them can at this point. It's whether the model has any aesthetic ambition when nothing in the prompt forces it to have one. Gemini ran the code. Claude added axial tilt. What's your hardest HTML/CSS test prompt?

Google just launched Gemini Omni. A video model that actually understands the world. Not just "make a video from text." Edit videos by chatting with it. Swap characters. Change styles. Add physics. Build on top of what you made - turn by turn. I tried Omni Flash in Google Flow and generated this. ↓ One prompt. That's it. What would you make first?

Google just shared Gemini's token processing numbers. May 2024: 9.7 Trillion tokens a month. May 2025: 480 Trillion. May 2026: 3.2 Quadrillion. That's not growth. That's a different graph entirely. 7x in one year. From half a trillion to 3.2 quadrillion. The world is using AI at a scale that didn't exist 12 months ago.

Google just proposed a new web standard called WebMCP. The idea: your website tells AI agents exactly how to use it. Instead of an AI blindly clicking around hoping it finds the right button - your site hands it a map. Book a flight. Fill a form. Complete a checkout. All done by AI. Correctly. First try. The agentic web just got an address. Are you building for AI users yet?

I opened buildfastwithai website in Chrome and asked DevTools' AI to audit it. I didn't write a single line of code. I just typed: "audit this page." Here's what the AI did on its own. No prompting, no guiding: → Ran a full Lighthouse audit → Analyzed all the scores (Performance: 39, Accessibility: 91, SEO: 100) → Read through every failing element → Found buttons that screen readers couldn't understand → Suggested the exact code fix, line by line The whole thing took under a minute. And that's just one thing it can do. The AI inside Chrome DevTools can now: → Fix CSS issues by looking at your actual styles → Debug slow or broken network requests → Inspect and modify page elements directly → Analyze your JavaScript for errors → Improve your Core Web Vitals, the scores Google uses to rank your site Before this, improving your website's SEO and performance meant hiring someone, reading documentation for hours, or guessing what was wrong. Now you open DevTools, describe the problem, and get the answer. Every website owner needs to know this exists. Open Chrome → F12 → AI assistance tab. Try it on your own site right now. What score does your site get?

Google Flow just dropped on Android. Beta. Free. Right now and the detail everyone is going to miss: Nano Banana Pro and Nano Banana 2 - both free to use. Generous image limits. No subscription prompt. Google's best image models. On your phone. For free. Download it before they change their minds.

Antigravity CLI vs Antigravity 2.0 app - what should you use? Most people are conflating them because they shipped on the same day. Here's the actual split: Antigravity 2.0 (desktop app): - Visual command center for orchestrating agents - Mission Control to manage multiple agents at once - Scheduled background tasks, voice commands - Best for: project-level orchestration, watching agents work in parallel, switching between projects Antigravity CLI: - Terminal binary, written in Go - Same agent harness as 2.0, no GUI - Wires into pre-commit hooks, CI pipelines, deploy gates - Best for: scripting, tmux/remote SSH, automation, "agent as a Unix tool" The shared harness is the part that matters. When Google ships a new tool or a better planning loop, it lands in both the same day. You can start a task in the CLI on a remote server, then check on it from the desktop app on your laptop - same context, same agents. Mental model: - CLI is the agent as a script you run - 2.0 is the agent as an app you manage - IDE is the agent as a pair programmer which surface are you actually using day to day?

most people downloaded Antigravity 2.0 and got confused at the first screen. here's the actual walkthrough and why the agentic mode setup is the most important decision you'll make before touching a prompt. SETUP: WHO'S DRIVING? when you first open Antigravity, the Agent Manager asks you to pick a development mode. this single choice defines the whole experience: → Agent-Driven (Autopilot) agent writes code, creates files, runs commands. no confirmation. use for greenfield projects or throwaway prototypes. → Agent-Assisted (Recommended) you stay in control. agent helps with safe automations. terminal policy: Auto for standard commands, Agent Decides for the rest. this is where most experienced devs should start. → Review-Driven agent asks permission before almost any action. use on production codebases where you need full audit trail. WHAT THE AGENT ACTUALLY DOES one prompt → the agent works across 3 surfaces simultaneously: editor → writes the code terminal → launches the app browser → opens Chrome, clicks through your UI, records video proof this is the part that's genuinely different from Cursor or Claude Code. the agent produces receipts. THE ARTIFACTS SYSTEM every task generates 4 artifacts. Google Docs-style, commentable: Task List — structured plan before any code is written Implementation Plan — technical architecture of what will change Code Diffs — exact lines changing, reviewable inline Walkthrough — summary + screenshots + browser recordings after done you don't have to trust "I fixed it." you get a video of the fix working. MODELS AVAILABLE Gemini 3 Pro, Gemini 3.5 Flash, Claude Sonnet 4.6, Claude Opus 4.6, GPT-OSS-120B. free during preview. Claude models need your own API key. PRO TIP BEFORE YOUR FIRST PROMPT create an initial Git commit and set the agent to commit incrementally. if the agent hallucinates a file structure or breaks a dependency, you revert in seconds. one developer saved 3 hours doing exactly this. this is a different mental model from a coding assistant. it's closer to hiring a contractor who sends you daily deliverables. what are you building first with it?

@saen_dev true

@BuildFastWithAI 15 lines to a working agent is a great demo, the honest line count once you add auth, error handling, and observability is closer to 500. The SDK pattern is right but the runtime story matters more than the LOC.

Google Antigravity SDK just dropped. 15 lines of Python → working AI agent. Same runtime as Antigravity 2.0. Apache 2.0 licensed. MCP server support built in. Gemini 3.5 Flash as default model. the runtime improvement flywheel is the real story: Google improves the platform → your SDK agents improve automatically. Claude Code, Codex, now this. which one are you actually building on?

@xiaoyvLiu exactly !

@BuildFastWithAI The real unlock is getting agent setup down to ~15 lines. Once the runtime gets this thin, people stop demoing agents and actually start shipping them.

@giskyexplorer it really be like that

@BuildFastWithAI Thanks for the free labor, now get out. Corporate open source.

Gemini CLI is being killed. June 18. No rollback. It's going closed-source. Forced migration to Antigravity CLI. 100,000 GitHub stars. 6,000 pull requests. Hundreds of open-source contributors. All of that community effort - and Google just closed the door. The new tool might be better. But this is not how you treat your developers. Thoughts?

@_itsjustshubh exactly , you have to be the decision maker and tell it what to do instead of the other way around

@BuildFastWithAI the chatbot trap is real. the unlock is when you stop asking it questions and start giving it persistent context and letting it run. totally different output quality

I spent 30 days testing every major AI workflow tool. Here's my honest ranking👇: