MIT Study: GPT-6 Astra Can Directly Rewrite Other Models' Weights Without Retraining

AIMIT shows AI improving AI, no training required

PLUS: Turn one interview into two clips with Riverside

Zach Mink

Good morning, AI enthusiasts, and welcome to our 4,563 new readers. Every few weeks, it seems another leading AI lab says that an autonomous AI researcher is closer to reality. New MIT research just gave a peek at how much of that capability may be sitting inside models already available to the public.

GPT-6 Astra went inside a smaller AI with poor math scores and improved it directly, skipping the training that normally does that work. Once improving a model is as simple as pointing another one at it, AI’s already frantic pace could hit another gear.

In today’s AI rundown:

OpenAI’s Astra rewrites other AI from the inside

OpenAI grades its AI on everyday tasks

Turn one interview into two clips with Riverside

OpenAI joins Abu Dhabi’s AI-native government push

🔬 OpenAI’s Astra rewrites other AI from the inside

Image source: MIT CSAIL

The Rundown: MIT researchers just published research showing that OpenAI’s GPT-6 Astra can improve some AI models by directly rewriting their weights — the numbers that decide how a model behaves — rather than putting them through more training.

The details:

Weights are normally set through training, since even small AI models have far too many to read or adjust by hand.

Alibaba’s small Qwen model scored 7.9% in math until Astra changed just 0.016% of those weights to cut its habit of repeating itself, hitting 48.1% after the fix.

The same approach did little for Google’s smaller Gemma 3, which lacked a single flaw like Qwen’s repetition, barely moving its math score.

Smaller tests also had Astra name the game an AI had learned from its weights alone, and write a simple Breakout player from scratch without training.

The authors say this capability could one day let AI upgrade itself, add or remove skills in other models on request, or check them for hidden problems.

Why it matters: The top AI labs are racing toward self-improving, automated AI researchers, and while this research is just one small example, it may be a sign of where things are headed. Once models are strong enough to alter and improve their weights, we’re in major “takeoff” territory, where AI starts speeding up its own progress. TOGETHER WITH OPTIMIZELY

🔎 Finding AI use cases in your workflowsVisit optimizely.com

The Rundown: When it comes to AI planning, the highest-value use cases are already living inside your workflows — you just have to know where to look. Optimizely’s template helps you find AI use cases that really matter in just a few minutes.

The template gives you:

A 3-step framework to build the foundation of your next AI agent

The exact process that turned a 2-3 hour GA4 reporting task into a 5-minute one

Real use cases mapped to how your team actually works

Download the template for better AI use cases. OPENAI

📋 OpenAI grades its AI on everyday tasks

Image source: OpenAI

The Rundown: OpenAI just introduced LifeVal, a new test grading how well AI gets through everyday errands like booking appointments, buying groceries, and filling out government forms, with tasks chosen based on what people say they want help with.

The details:

The selected task areas come from a poll of 4K people on where they want help, and global time-use surveys on what everyday tasks consume time.

LifeVal has been OpenAI’s internal measurement for the past year, as it trained models to browse the web, operate a computer, and work in connected apps.

OpenAI’s models have gone from an average score of 49.3% (GPT-5.2) to 80.6% (GPT-6 Astra), though just 0.6 points of that gain came after July’s GPT-5.6.

Astra’s “best recorded” scores hit 100% on 10 of 16 tasks OpenAI listed, like replacing a lost engagement ring, with LSAT help scoring the lowest at 88.7%.

Why it matters: We look at benchmark results on everything from coding and math to business tasks, but few have measured how models handle the everyday tasks that a more general audience is asking AI to do. Seeing this eval tested across other models would be an interesting data point for the personal agent meta seen across the industry. AI TRAINING

🎬 Turn one interview into two clips with RiversideVisit The Rundown University

The Rundown: In this guide, you’ll use Riverside’s new agents to turn a licensed interview into two captioned videos. You’ll choose useful moments, check the edits, and prepare post drafts without filming anything new.

Step-by-step:

Download the guide’s openly licensed interview and save its credit. Then, make an MP4 copy and upload it to a Riverside practice project

Ask Clipper for short passages that make sense on their own. Make sure to watch the surrounding interview before choosing your two clips

Ask Editor for a matching vertical layout, keeping the full picture visible and leaving space for captions and the source label

Check the captions against the speech, then ask Marketer for two short post drafts. Keep the interview’s date and full credit with each one

Finally, export both clips and watch the saved files. Check the framing, sound, words, and ending before handing them off

🔧 The simple fix for apps that stallVisit quickbase.com

The Rundown: Vibe coding makes building apps fast. But it’s also left piles of slick-looking apps that sit unused because they can’t talk to the tools you rely on every day. Pave gives you the speed without the dead end.

With Pave, you can:

Describe your process and get a working custom app in minutes

Link to 350+ systems, including Quickbase, on every plan

Create workflows to run reminders, approvals, and handoffs on their own

🤝 OpenAI joins Abu Dhabi’s AI-native government push

The Rundown: OpenAI just sealed four agreements in Abu Dhabi, bringing its AI to state oil giant ADNOC, the emirate’s government, and state investor Mubadala, while a UAE group is reportedly set to be the top investor in the AI giant’s raise of up to $20B.

The details:

ADNOC’s deal, the largest of the four, aims to cut oil field planning from months to hours by running decades of well data through OpenAI’s models.

Food imports come first in “Machine-Readable Government,” a project turning some Abu Dhabi rules into text AI can read to help streamline permit processes.

Abu Dhabi already has a $3.5B plan for AI across all public services, and a separate OpenAI and Microsoft program aims to train 20K staff and the public.

OpenAI’s models will help Mubadala weigh risk and spot new bets, while a five-year pact with AI firm G42’s Inception42 targets UAE businesses and schools.

Why it matters: Albania gained headlines last year by naming a digital AI “minister,” but Abu Dhabi’s version is more practical — giving AI better access to existing data and rules to improve efficiency. The circular AI dealmaking also continues, with Abu Dhabi acting as both a major deployer of OpenAI’s tech and potentially the biggest investor.

▸ Judd built an AI research team that fact-checks itself

Today’s workflow comes from reader Judd Slivka:

“I use a four-step system built on Grok Bot to get AI-generated research reports with fewer errors. It’s for anyone who relies on AI research and needs to trust the numbers and quotes in it.

I start with a research bot that has specific instructions. They cover the types of sources it should use, separating known facts from speculative claims, and cross-checking claims. The report goes to a skeptic. The skeptic is meant to be contrarian but not a roadblock. It asks for evidence, looks for logical fallacies, and creates a short, ranked list of problems. The problem list goes back to the researcher, who then passes the report to my fact-checker. The fact-checker double-checks every number and every quote or paraphrase against the actual document it was found in.

The result is a report with fewer errors, but also flags on what’s verified and what isn’t. Far and away the most valuable workflow that I’ve built. It costs more in terms of usage, but it’s been working great.”

See the workflow on our Community Workflow Hub →

🛠️ Trending AI Tools

🧪 Hydrogen-1.0 - Deep Principle’s open model for chemistry, materials research

🗣️ Decision Model API - Hume’s AI that answers queries about voice recordings

🎥 Riverside Agents - Team of AI agents to produce, edit, clip, and post videos

🎨 Craft - Canva’s first in-house AI image model

📰 Everything else in AI today

Elon Musk responded to commentary surrounding users seeing significant usage cuts on SpaceXAI’s Grok Bot, saying “The cost of serving Grok Bot was unsustainably high relative to revenue.”

Chipmaker Marvell agreed to a $115B all-stock deal to acquire Lumentum, which makes lasers for AI data centers, marking one of tech’s largest takeovers ever.

Anthropic filed an updated IPO prospectus with new figures like nearly $17B in Q3 revenue and a $567M operating loss, also adding language that upcoming models could “imminently” pass its safety line for AI-driven progress.

Deep Principle introduced Hydrogen-1.0, an open model built on Z.ai’s GLM-5.3 that the startup claims beats GPT-6 Astra on OpenAI’s own scientific research benchmark.

Anthropic’s Addy Osmani confirmed that Claude Code is now instructed not to rebuild specific existing games over IP concerns, with some devs pushing back after refusals hit projects they call legitimate.

🎓 Highlights: News, Guides & Events

AI backlash heads to the polls

Inside the spy plot around Nvidia’s CEO

Zach Mink • 8 minutesAI

Rowan Cheung • 4 minutes

Stay Ahead on AI.