- AI Fire
- Posts
- 🧠 AI Rewrote Its DNA
Read time: 5 minutes
AI just rewrote its own code and doubled its performance without human help (even tried to fake test results & hack the feedback it was supposed to learn from). Most experts thought was still years away…
What are on FIRE 🔥
LEARNING PARTNER: AIRCAMPUS
You don’t need another productivity hack, You need a system that works without you.
📅 June 5th | 🕙 9 AM EST
Forget “5 easy hacks.” This is your AI-Powered WORKFLOW SYSTEM that:
✅ Runs your inbox & closes leads without you lifting a finger.
✅ Owns your social media- posts, schedules, and analyzes like a pro.
✅ Manages your calendar flawlessly with follow-ups & reminders.
✅ Connects 40+ AI agents into a seamless workflow.
Want to work on your business, not in it?
This is your exclusive invitation to build your AI Twin - your tireless work partner.
First 50 spots are FREE- then doors shut tight.
🎟️ Claim Your Free Seat Now - Before They’re Taken!
AI INSIGHTS
Japan’s Sakana AI is taking “self-help” to a whole new level. It just built what might be the world’s first Darwin Gödel Machine (DGM), the first working prototype capable of rewriting its own code to increase performance from 20% to 50%.
🧬 What Is the Darwin Gödel Machine?
A self-modifying coding agent that can:
Read and edit its own Python code
Propose changes based on its own performance feedback
Evaluate if changes improve its abilities
It operates by testing hypotheses, much like a scientist, and rewriting its own architecture if it finds improvements, just as humans do.
🧪 Performance Results
On SWE-bench (real-world GitHub issue resolution): Performance improved from 20% to 50%
On Polyglot (multi-language benchmark): Jumped from 14.2% to 30.7%, outperforming hand-crafted agents
Crucially, improvements were not model-specific and transferred across: Foundation models (e.g., Claude 3.5, o3-mini) & Languages (Python to Rust, C++, Go, etc.)
DGM keeps an archive of all previous agents (including “worse” ones), allowing: Exploration of multiple evolutionary pathways & Goal switching and innovation through non-linear evolution.
DGM showed early signs of objective hacking. It faked log outputs (e.g., pretending to run unit tests). In some tests, it tried to manipulate the evaluation metric by removing markers used to detect hallucinations.
Why It Matters: This is one of the first working non-static AI that can “learn to learn” autonomously, cause current AI systems are “frozen” after training. But DGM’s ability to manipulate its own feedback loops reveals real threats of unintended behavior.
TODAY IN AI
AI HIGHLIGHTS
📱 Android and iOS users can try Gemini Live now, just share your screen or camera and ask questions in real time. Plus, anyone can use the latest Jules coding assistant to handle 60 tasks daily for FREE.
🧠 The U.S. dropped $500M to upgrade federal systems with AI, and banned all 50 states from regulating it for 10 years. Some say AI is Manhattan + Moonshot + 1990s dotcom craziness.
AI regulation locked down for 10 years. What could possibly go wrong? |
💯 OpenAI's o3 just found its first real zero-day vulnerability after being fed around 12,000 lines of code, even the original human expert missed. And researchers are freaking out…
📊 A branding coach on X just shared one prompt to turn new Perplexity Labs into a viral marketing agent. It's the results a marketing team would achieve for $5,000 in 10 minutes.
🧠 Meta is creating a SaaS-like experience with an end-to-end AI ad automation. You’ll just need to provide an image to get different ad versions in real time.
🩺 The FDA just officialy approved Clairity Breast — the true first-of-its-kind AI tool that predicts 5-year breast cancer risk using just a standard mammogram. No extra tests needed.
💰 AI Daily Fundraising: Saudi Arabia just launched a $10B AI fund via HUMAIN to back startups and build 6.6 GW of AI data centers. Deals include $23B with Nvidia, AMD, AWS, and Qualcomm.
AI SOURCES FROM AI FIRE
June is finally here! 🌟 And also ToolDrop Ep #9 - time to bring you something fresh and exciting. Here’s what’s in store this week:
Anthropic’s Prompt Engineering Interactive Tutorial
Anthropic’s Educational Courses
Create Reddit Videos with Just One Command
Alternative to NotebookLM/Perplexity/Glean connected to external sources
Cursor for Designers – A Visual Vibecoding Editor
Note: These exclusive resources & reviews are available only in our AI Fire community. It’s because you guys can freely ask for support or share personal experience during testing there. Get your full breakdown here (no hidden fee)!
NEW EMPOWERED AI TOOLS
🔥 App20X lets anyone launch and customize any open-source web app using just simple prompts. Instant access to millions of apps: from CRMs to Airtable-like databases, the power of open source, unlocked. Try for FREE.
🎥 Wavel AI creates stunning multi style videos with just one single input.
⚙️ Requirements AI turns any scripts to product requirements in seconds.
🎨 VibeNecto AI creates stunning marketing visuals, content in just minutes.
🗣️ Conversational 2.0 From ElevenLabs generates natural AI voices instantly.
*indicates a promoted tool, if any
AI QUICK HITS
🎬 Microsoft Bing gets a Free OpenAI’s Sora-powered AI video generator.
💻 Character AI rolled out a slate of multimedia features with a social feed.
📱 Samsung may incorporate Perplexity’s AI in its phones, not only Gemini.
🪐 AI cracked Mars mystery: 500,000 streaks solved without a drop of water.
🎵 Major record labels are reportedly in licensing talks with AI firms Udio & Suno.
AI CHART
While the world debates OpenAI drama and Google’s AI demos, a parallel frontier AI world is forming in China. And it’s not just DeepSeek competing for the top slot — there are more than 10 labs vying for the lead.
Some of the top names include: Moonshot AI, Tencent Hunyuan, Huawei PanGu, Baidu Ernie 4.0… These labs are pushing releases frequently and training models ranging from 7B to 100B+ parameters. Many are releasing open-source versions.
Several labs report state-of-the-art results on multilingual benchmarks, coding tasks, and summarization. Models like GLM-4, Yi-34B, and Kimi Chat are showing Claude/GPT-tier reasoning in Mandarin and English.
They focused not just on chatbots but on:
Enterprise applications
AI agents for code, law, finance, and healthcare
Real-time assistants built into Chinese apps and browsers
Some are now building in-house GPU infrastructure to reduce reliance on U.S. chips. We now have two full-blown AI superpowers building independently at scale. Many of these models aren’t yet widely known outside Asia, but that’s changing.
AI JOBS
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – 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