Code that's writing code & more (May 16, 2025)

👋 Hey there,

Today’s AI news is full of big swings and quiet revolutions. From self-coding AI agents and a $3B model play to cryptic teases and model performance drama, the landscape is reshaping fast. Let’s ride the wave.

🔧 Agents & Alpha Drops

🧠 AlphaEvolve: Google's Self-Coding AI Agent

Google dropped AlphaEvolve — a self-coding AI agent that optimizes code, solves math problems, and boosts chip design. It matched or beat top math solutions on 50+ open challenges and accelerates code development by weeks. Think of it as an autonomous hacker for science.

🌊 Windsurf Releases SWE-1 Family After OpenAI Acquisition

Right after OpenAI’s reported $3B acquisition, Windsurf launched SWE-1 — its in-house model suite trained to work across terminals, browsers, and editors. SWE-1 outperforms most open models and nearly rivals Claude 3.7 Sonnet. Major strategic move? Probably.

🌀 Meta Postpones Behemoth Model Launch

Meta’s flagship model "Behemoth" is delayed due to underwhelming performance gains. Internal frustration is building. Meanwhile, OpenAI's silent moves and Manus’s workflow-planning image gen tech are making more noise with less drama.

📈 Trends, Tools & Teasers

📊 Poe Report Shows Fast-Shifting Model Use

Poe’s Spring report is spicy: GPT-4.1 and Gemini 2.5 Pro gained 15% share in weeks. Claude dropped by 10%. ElevenLabs dominates audio with 80% usage, Kling owns 30% of video, and GPT-image-1 is gunning for FLUX’s spot.

🧑‍💻 ChatGPT Becomes the New Internet Gateway

ChatGPT is now the 5th most visited website globally. People are skipping search and going straight to bots — signaling a web where your first (and only) stop might just be an AI.

🧬 OpenAI Teases Mysterious New Tool

Sam Altman called it a “low-key research preview,” but speculation is swirling. Dev-focused, possibly under a new name, and definitely aiming to stir the pot again.

🛠️ Tools of the Day

• Wispr Flow – Talk-to-type productivity boost across 100+ languages
• Jace – AI inbox assistant for replies, meetings, and tasks
• Manus AI – Agentic image generation with workflow planning
• Dynamiq – Fast agent deployment with ready-to-go widgets

💬 Prompt of the Day

⚡ Quick Hits

• FBI warns of deepfake scams mimicking U.S. officials

• Gemini simplifies complex text with near-human fidelity

• xAI’s Grok went rogue — explained as “unauthorized modification”

• LLMs drop to 60% accuracy in multi-turn chats (from 90% single-turn)

• UAE to build largest AI campus outside the U.S.

🧠 TLDR

Google dropped a powerful self-coding agent, Windsurf released a surprise in-house model after a $3B buyout, and Meta’s struggling with its own. ChatGPT is now an internet gateway, Poe data shows Gemini and GPT-4.1 on the rise, and OpenAI is teasing its next big thing. The ecosystem is tilting toward agents and verticalized models with serious horsepower.

Until next week,
David