🤯 AI Can Now Edit Another AI Without Retraining It

Welcome back to AI Researches! 👋

This week, MIT researchers showed that GPT-6 Astra can directly rewrite another AI model’s internal weights, improving its math accuracy from 7.9% to 48.1% without traditional retraining. It's an early look at how AI could eventually improve other AI systems on its own.

Also inside: Google's Genie 4 and Gemini 4 Carbon, Anthropic's IPO filing, OpenAI's latest research, practical AI tutorials, expert predictions, useful tools, and more.

In today’s AI Researches:

• 🧠 MIT researchers show GPT-6 Astra can directly rewrite another AI's weights

• 🌍 Google DeepMind launches Genie 4, an interactive AI world model

• 🚀 Google unveils Gemini 4 Carbon, its strongest AI model yet

• 💰 Anthropic files for its Nasdaq IPO

• 🎬 Learn how to build your own YouTube alternative with Claude Opus 5.5

• 📊 AI Researches case studies

• 📞 AI Tip: Turn sales call notes into follow-up emails with Mercury 3

• 🔮 Wild predictions from Elon Musk and Sam Altman

• 🛠️ Best performing AI tools of the week

• 📰 More AI news and developments

Read time: 5 minutes

NEW

🧠 MIT Researchers Show GPT-6 Astra Can Directly Rewrite Another AI’s Weights

AI Researches: MIT CSAIL researchers showed GPT-6 Astra can read and directly edit neural network weights using natural-language reasoning, improving another AI without gradient-based training.

The Details

• Astra changed just 0.016% of Qwen3.5-0.8B’s weights, raising MATH-500 accuracy from 7.9% to 48.1% and GSM8K from 13.8% to 67.1%.

• It found that Qwen was getting stuck in reasoning loops, edited weights in two layers, and also pushed instruction-following performance from 21.0% to 62.3%.

• In other experiments, Astra repaired a failing car-driving network to reach 100% success and created game-playing neural networks from scratch.

• The researchers call this Weight-Space Programming, but stress that the work is still early and limited by context size and current tooling; they are now building a benchmark to test it more rigorously.

Why this is important

If this approach holds up at larger scales, AI models may eventually be able to diagnose and fix other models by directly rewriting small parts of their weights instead of retraining them from scratch.

AD

10 AI Stocks to Lead the Next Decade

AI isn’t a tech trend – it’s a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.

These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.

Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.

Download Now

Major AI News that Happened in this week

Apple opened Siri AI to all developers, letting any app plug its own content and actions directly into the assistant.

AI data centers became a major U.S. election issue, with 38 local jurisdictions weighing at least 41 proposals targeting new projects.
OpenAI is trying to reshape the debate by framing its facilities as compute campuses tied to jobs, local funding, and grid investment.

Nvidia CEO Jensen Huang was reportedly targeted in a suspected Chinese intelligence operation involving Megaspeed executive Alice Huang.
The allegations remain unproven publicly, and Nvidia says Huang was unaware of any suspected intelligence activity.

Google unveiled Gemini 4 Carbon, its strongest model yet for cybersecurity and long-running coding tasks.
It scored 59 on Artificial Analysis’ Intelligence Index, taking the #1 spot overall at launch.

Google DeepMind introduced Genie 4, a world model that generates interactive 3D environments with native sound and conversational characters.
Users can explore generated worlds in real time and change objects, weather, or other elements through text prompts.

Anthropic filed its S-1 for a proposed Nasdaq IPO under the ticker ANT.
The filing shows $16.3B in revenue for the first half of 2026 and a $974M operating profit in the June quarter.

🎬 Build your own YouTube alternative with Claude Opus 5.5

Learn how to use Claude Code to build a video recommendation website like Andrej Karpathy’s HackerTube, which curated 6,454 YouTube videos from millions of Hacker News comments.

1. Set up Claude Code
Download Claude Desktop, open Claude Code, and select Opus 5.5 with Medium effort.

2. Get the video dataset
Visit HackerTube and download its reusable dataset containing curated videos and recommendation data.

3. Generate the website
Upload the dataset to Claude Code and ask it to build a YouTube-style interface with search, categories, video recommendations, and filters.

4. Connect YouTube data
Create a YouTube Data API key through Google Cloud and have Claude fetch video titles, thumbnails, and metadata.

5. Test and deploy
Preview the website, fix errors, then ask Claude to deploy and secure it using your hosting server and domain.

Tutorial links: Karpathy’s original build guide

According to LinkedIn estimates, AI-linked roles have created 750,000+ new U.S. jobs since 2023, led by data annotators, data center workers, and AI engineers. These roles also tend to pay much more, with AI-related jobs showing a median salary of about $180,000, compared with roughly $80,000 across all jobs.

OpenAI says its unreleased AI models made new discoveries across math, computer science, and physics, including a proof resolving major geometry conjectures and a result claiming the decades-old Exponential Time Hypothesis is false. The results are being released with proofs and raw outputs so outside researchers can independently check them.

📞 Turn Sales Call Notes into Follow-Up Emails with Mercury 3

1. Open Mercury 3
Go to OpenRouter Chat, select Inception Mercury 3, check the current usage price, and start a fresh conversation.

2. Add your sales notes
Paste your product details and notes from each sales call. Ask Mercury 3 to recommend the right plan and draft a short follow-up email for every lead.

3. Check every recommendation
Verify team limits, required features, pricing, and any promises in the draft. Make sure the model doesn’t treat a requested discount as an approved one.

4. Test changing requirements
Change an important detail, such as increasing a customer from 8 to 18 users, and see whether Mercury correctly updates its recommendation.

5. Compare it with your usual model
Run the same sales notes through another model and compare accuracy, speed, and how much editing each draft needs.

Elon Musk says the AI boom could hit a major hardware wall by the end of 2026, not because of chip shortages, but electricity. He argues chip production is growing much faster than power supply, meaning the industry could soon have more AI chips than it can actually turn on.

Sam Altman says OpenAI wants robotics to help automate the expansion of AI infrastructure. His vision includes robots operating chip fabs, constructing solar panels, and eventually helping data centers build more of themselves with minimal human involvement.

• 🧭 Ariso — Acts like a personal AI assistant across text, email, browser, and desktop, keeping track of commitments and helping coordinate plans with other people.

• 🎬 Clueso — Turns screen recordings, documents, presentations, or a simple prompt into polished product demos, tutorials, and branded videos.

• 🪞 Elmo — Tracks how ChatGPT, Claude, Gemini, Perplexity, and other AI engines mention your brand, with open-source visibility, citation, and competitor analytics.

• 🛎️ Sundae — Gives you a personal AI worker you can message to research, book restaurants, compare options, monitor information, and handle browser-based errands.

• 🎤 Colloqa — Rehearses interviews, salary negotiations, and difficult conversations out loud, pushing back realistically before coaching you on what to improve.

• 🎥 Vidu Q4 — Generates 4K AI video from up to 15 reference images and three voice samples, keeping characters, scenes, style, and voices consistent across the clip.

Hume AI introduced the Audio Decision API, a new model that answers structured questions directly from raw audio.

Anthropic updated its rules to restrict Claude Code from recreating existing copyrighted video games.

Anthropic introduced self-assessment, a system that lets Claude review its own conversations for potential misuse after responding.
On Claude Fable 5, it catches 78% of potentially harmful chats and 94% of high-priority cases at a 1% false-positive rate.

Axiom completed a fully machine-checked version of the Classification of Finite Simple Groups, often described as math’s longest proof.
Its team used an AI agent swarm to rebuild and verify the roughly 15,000-page result in just 24 days.

Anthropic expanded into drug development, using Claude across research, regulatory work, and early-stage therapeutic programs.

OpenAI released GPT-6.1 Terra, a faster and cheaper model focused on computer use, visual reasoning, and efficiency.

Canva launched Craft, its first in-house AI image model, debuting near the top of DesignArena’s leaderboard.

Moonshot AI released Kimi K3.1 Pro with a 1M-token context window and major gains on long-horizon agent tasks.
It scored 67% on Terminal-Bench 4.0 and 73% on AutomationBench-AA, with pricing set at $2 input and $10 output per million tokens.