- unwind ai
- Posts
- Agentic Vibe Coding IDE with Built-in Browser
Agentic Vibe Coding IDE with Built-in Browser
+ Mistral's open-source coding models, Google's Titan + MIRAS for long-term AI memory
Shubham Saboo & Gargi Gupta
December 10, 2025
In partnership with
Today’s top AI Highlights:
& so much more!
Read time: 3 mins
AI Tutorial
Imagine uploading a photo of your outdated kitchen and instantly getting a photorealistic rendering of what it could look like after renovation, complete with budget breakdowns, timelines, and contractor recommendations. That's exactly what we're building today.
In this tutorial, you'll create a sophisticated multi-agent home renovation planner using Google's Agent Development Kit (ADK) and Gemini 2.5 Flash Image (aka Nano Banana).
It analyzes photos of your current space, understands your style preferences from inspiration images, and generates stunning visualizations of your renovated room while keeping your budget in mind.
We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.
Latest Developments
AI models hit a wall when context grows, compressing everything into a fixed state means losing the details that matter.
Google Research's Titans + MIRAS throws out that constraint entirely. It treats memory as a deep learning problem, not a storage problem.
The papers were released before but presented this year at NeurIPS 2025.
They've basically given AI models the ability to learn what to remember while they're actively reading. The system works like your brain does: it notices when something breaks the pattern (a banana peel in a financial report) and immediately flags it as "this matters, store this permanently." Meanwhile, routine stuff that fits expectations gets ignored. The result? Titans beats GPT-4 on extreme long-context tasks despite being a fraction of the size, handles 2M+ token windows like it's nothing, and runs at RNN speed with transformer-level accuracy.
Key Highlights:
Smart Forgetting - Uses momentum to track both current surprise and recent context flow, while adaptive weight decay strategically dumps information that's no longer useful—keeping memory sharp without bloating.
Depth Matters More Than Size - Deeper memory modules with the same total parameters consistently crush shallow ones on perplexity, proving that memory architecture beats raw size when sequences get long.
Beyond Language - Works on DNA sequences and time-series forecasting just as well as text, showing this isn't some language-specific hack—it's a fundamental improvement in how models handle sequential data.
MIRAS Unifies Everything - The framework proves that transformers, RNNs, and state space models are all just different flavors of the same thing: associative memory with different optimization strategies.
The Titans paper and MIRAS paper are both live. If you're working on long-context applications or building RAG systems, this architecture is worth studying.
Turn AI Into Extra Income
You don’t need to be a coder to make AI work for you. Subscribe to Mindstream and get 200+ proven ideas showing how real people are using ChatGPT, Midjourney, and other tools to earn on the side.
From small wins to full-on ventures, this guide helps you turn AI skills into real results, without the overwhelm.
Orchids is the world's first vibe coding IDE, an environment where you describe what you want and the AI handles everything from code to deployment without breaking your flow.
Ranked #1 on App Bench at 76% and #1 on UI Bench at 30.08, outperforming all vibe coding tools, including Claude Code, Replit, Lovable, and Bolt.
The difference is integration: an AI agent, IDE, browser, Supabase, Stripe, and Vercel deployment all in one local tool. The agent sees your screen and hears your voice, so you can show it a website and say "build something like this" and it'll grab the UI elements directly. Setting up auth, databases, or payments happens through single prompts like "add Stripe checkout" - the agent handles the implementation natively.
Deploy with one click, track analytics on live projects, and never leave the IDE from idea to production.
Key Highlights:
AI Agent with Context - The agent sees your screen and hears your voice like a human developer, letting you show problems instead of describing them. Select any UI element on the web through the built-in browser and Orchids will recreate it in your codebase.
Full-Stack Without the Setup - Native Supabase and Stripe integrations handle auth, databases, and payments through simple prompts like "set up payments." No configuration files or API wrestling required.
Local-First, No Lock-In - Runs on your machine with two-way Github sync, import projects from any platform (v0, Bolt, Replit, Lovable), and restore to any previous state with automatic snapshots at each user message.
Complete Development Environment - Deploy to Vercel with one click, track analytics on live projects, and work in an end-to-end IDE with editor tooling and integrated preview, everything from prompt to production in one tool.
Available now for macOS with free download and usage-based pricing.
Quick Bites
Build and manage a fleet of reliable web agents
Amazon's AGI Labs just made Nova Act generally available on AWS to build and manage a fleet of AI agents to automate browser-based tasks. Think AI agents that submit perfectly formatted bug reports across multiple repositories, update data in 87 different tools, or your CRM. The system hits ~90% reliability on browser-based workflows. The team has also added features like a no-code playground and human-in-the-loop oversight, as well as preview capabilities like tool-calling to accelerate builders throughout their journey. You can start testing at nova.amazon.com/act
Claude Opus 4.5 is now available in Claude Code for Pro users
Claude Opus 4.5 is now available in Claude Code for Pro users. Just run claude update and use /model opus to access it. Anthropic recommends saving it for complex tasks since it burns through rate limits faster than Sonnet, though Max plan users get enough headroom to use Opus as their daily driver.
Make Claude fine-tune open-source LLMs end-to-end
Claude can now fine-tune language models end-to-end using a new tool called Hugging Face Skills. You describe what you want in plain English ("Fine-tune Qwen3-0.6B on this dataset"), and the agent handles GPU selection, script generation, job submission, and monitoring. It supports SFT, DPO, and GRPO, scales from 0.5B to 70B parameters, and costs around $0.30-$40 depending on model size. The skill validates datasets before training, integrates real-time metrics through Trackio, and can convert finished models to GGUF for local deployment.
Mistral drops open-source agentic coding models and CLI agent
Mistral released Devstral 2 and Devstral Small 2 under permissive licenses (modified MIT and Apache 2.0), scoring 72.2% and 68.0% on SWE-bench Verified while being 5-8x smaller than competing models. They're shipping these models with Mistral Vibe CLI, an open-source terminal agent that for autonomous agentic coding with codebase exploration, multi-file edits, Git operations, and Agent Communication Protocol for IDE integration. Devstral 2 is free during launch, then $0.40/$2.00 per million tokens.
Turn any website into an API with no code
Browser Use just shipped Skills, a way to turn any website into a callable API by describing what you want in plain text. Write a prompt like "get pricing and reviews from Amazon products," click through the site once to show it the pattern, and you get a production endpoint that returns structured data. It costs $0.01 per execution, handles cookies automatically, and generates parameter schemas you can immediately integrate into your apps.
Salesforce trained a Router that can cut LLM costs by 60%
Premium models deliver strong reasoning but are expensive, but smaller models are economical yet brittle. Salesforce Research dropped xRouter, a 7B routing model to orchestrate 20+ LLMs while optimizing for cost alongside performance. The system uses tool-calling architecture to dynamically select between premium models like o3 and budget options like GPT-4o-mini, achieving up to 60% cost reduction on math, code, and reasoning tasks without sacrificing quality. Model and code are open on Hugging Face under CC BY-NC 4.0.
Tools of the Trade
Sloppylint - Detects AI-specific code patterns that traditional linters miss, like hallucinated imports, cross-language leakage, placeholder functions, and mutable default arguments. It scores code across four axes (noise, lies, soul, structure) and assigns severity levels from critical to low.
mgrep - A CLI semantic search tool that uses natural language queries to search across code, images, PDFs, and text files in git repositories. It indexes your repo continuously, then lets you query with plain English instead of regex patterns. The team claims 53% fewer tokens and 48% faster responses for coding agents compared to standard grep.
SeekDB - An open source AI-native database that unifies vector, text, and structured data in a single MySQL-compatible engine. It supports embedded and standalone deployments with hybrid search capabilities.
PyTogether - A fully browser-based collaborative Python IDE with real-time editing, chat, and visualization. Edit Python code together instantly using Y.js. plus voice chat and live drawing for educational pair programming.
Awesome LLM Apps - A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
(Now accepting GitHub sponsorships)
Hot Takes
Two things can be true.
If you're not amazed by AI, you don't really understand it.
If you're not afraid of AI, you don't really understand it.
Among many weird things about AI is that the people who are experts at making AI are not the experts at using AI. They built a general purpose machine whose capabilities for any particular task are largely unknown.
Lots of value in figuring this out in your field before others.
~ Ethan Mollick
That’s all for today! See you tomorrow with more such AI-filled content.
Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!
PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉
Reply