MIT Taught GPT-6 Astra to Rewire Other AIs
Plus, 🤩 How to turn one phone selfie into a profile photo for every platform with ChatGPT, OpenAI unveils LifeVal to measure how well AI handles everyday tasks, and more!
Hola Decoder😎
If someone forwarded this to you and you want to Decode the power of AI and be limitless, then subscribe now and Join Decode alongside 30k+ code-breakers untangling AI.
🧠 MIT researchers show GPT-6 Astra can improve AI models without retraining them
MIT researchers have demonstrated Weight-Space Programming, a technique that allows advanced AI models to directly understand and modify other neural networks' weights, potentially changing how AI systems are optimized and improved.
The Shift:
- GPT-6 Astra improved Alibaba's Qwen3.5-0.8B math accuracy from 7.9% to 48.1% on MATH-500 by modifying just 0.016% of its weights, without additional training.
- Astra identified and corrected a reasoning-loop problem in Qwen, also tripling its instruction-following performance on IFEval through the same targeted weight changes.
- Astra correctly identified which games neural networks had learned by examining their raw weights alone, without running the networks or receiving gameplay feedback.
- Researchers demonstrated that Astra could create game-playing neural networks from scratch, including models that outperformed networks trained using 100 million reinforcement-learning steps.
- The technique showed limitations across different models, with similar modifications producing little improvement in Google's Gemma 3, suggesting its effectiveness depends on identifiable model weaknesses.
Weight-Space Programming could eventually allow AI systems to diagnose weaknesses, modify capabilities and improve other models without expensive retraining. However, these early experiments demonstrate targeted improvements rather than a general method for reliably upgrading any AI model.
Together with New Engen
Your next winning ad needs a different voice
Your product solves a real problem. But when every ad makes a promise, buyers need a reason to believe yours.
A qualified expert explaining how it works can give them that reason.
Grapevine, now part of New Engen, creates ads delivered through creator and publisher handles with production managed from brief to launch.
You approve the angles and ship product. Grapevine sources creators, handles briefs and payments, then produces and posts the content.
- Madison Reed achieved 20% efficiencies over target CPA and 50% higher LTV.
- Mathnasium cut Meta CPL by 33% in under 30 days.
- Cirkul achieved 15-50% lower TikTok CPA and nearly doubled CTR through creator whitelisting.
Put this approach to work with your first campaign on Grapevine.
🤩How to turn one phone selfie into a profile photo for every platform with ChatGPT
One mediocre selfie in. LinkedIn, X, speaker bio, and dating photos out, still looking exactly like you.
1. Start clean - Open a new chat, upload your sharpest selfie and type /clean-up to strip the messy background and stray people.
2. Lock your face, change the scene - On the result, run one line per platform: “/identity-lock, modern office, soft window light” for LinkedIn “/identity-lock, warm café, candid” for dating
3. Polish - Run /studio-quality for professional lighting, then /upscale on anything you’ll print.
4. Run the face drift check - “Compare this edit to my original photo. List any change to my eyes, nose, jaw, skin, hairline or face shape. If anything changed, redo it with those features identical.”
5. Know the fallback - These codes aren’t official commands. ChatGPT interprets them. If one gets ignored, spell it out: “Keep my face, hair and outfit exactly the same. Only change the background.”
Use a new chat for each photo so edits don’t bleed into the next one.
🤖 OpenAI unveils LifeVal benchmark to measure how well AI handles everyday tasks
OpenAI has introduced LifeVal, a new benchmark measuring how effectively AI models complete practical everyday tasks, reflecting its push to transform ChatGPT from a conversational assistant into a tool that can take action on users' behalf.
The Shift:
- LifeVal evaluates tasks across childcare, healthcare, education, personal finances, shopping, and government services, informed by time-use studies and a survey of 3,995 respondents.
- GPT-6 Astra achieved an 80.6% average score on LifeVal, up from GPT-5.2's 49.3%, demonstrating substantial improvement in OpenAI's evaluations of everyday task performance.
- GPT-6 Astra scored 100% on a home-search task, 98% on flood volunteer coordination and 93.1% on kindergarten enrollment, according to OpenAI's benchmark results.
- GPT-6 is approximately 23% faster than GPT-5.6 on these tasks, although GPT-5.6 delivers similar overall performance at roughly half the estimated API cost.
- OpenAI is improving browser use, computer interaction and app connectors, enabling ChatGPT to move beyond answering questions toward completing multistep tasks across digital services.
LifeVal reflects OpenAI's shift toward measuring AI by its usefulness in everyday life rather than academic benchmarks alone. While the results show progress toward more capable personal assistants, benchmark performance does not guarantee the same reliability in real-world situations.
Together with Playbook Pro
Your Next 6 Months of Ad Spend Could Disappear in 6 Weeks
CAC is up 60% this year, and 70% of first-time buyers never return. Keep leaning on ads, and the budget you’re planning for the next half year could be gone before spring.
Rewriting the DTC Rulebook is the shortcut top operators are using to escape the trap, a 24-page report distilled from 200+ hours of real brand research and field tests.
Inside, you’ll learn how to:
🔥 Build emotional campaigns that double repeat-purchase rates in 90 days.
💰 Raise prices without losing volume using the Brand Value Canvas.
🧠 Cut 30% of paid-ad dependency with the Owned Media Checklist.
Download today and get 2,000+ viral video hooks already tested to grab attention in under 3 seconds, ready to plug into your next launch.
Stop feeding the ad machine.
Grab the free report now and lock in sustainable growth before your next dollar disappears!
🏆 Tools you Cannot Miss:
🎨 Hera AI Creative Canvas – Generates images, editable designs, and animations on one canvas, allowing creators to build branded visuals and motion graphics without switching between tools.
🎬 Clayzo – Creates motion graphics, product animations, and interactive website visuals through AI coding agents, with exports for videos, GIFs, and web playback.
🌐 Greta.sh – Builds websites from text descriptions and helps grow their search visibility by generating new SEO and AI-search-focused pages for approval.
🖼️ magpie – Collects saved images, videos, and posts from Instagram, TikTok, Pinterest, X, and other platforms into one visual library, with AI tagging and search by meaning or color.
🧠 Xmind 2027 – Turns mind maps into collaborative AI brainstorming spaces where users can research ideas, organize information, and develop plans using their documents and existing maps.
🚀 Quick Hits
🎮 Anthropic updated Claude Code’s system prompt to prevent recreations of existing video games, citing intellectual property concerns. Developers can still build original games, genre-based titles, and projects using their own assets.
⚡ OpenAI announced plans to free up to 1,000 megawatts of electricity during emergencies through its Georgia data center, aiming to support local power grids amid growing opposition to AI infrastructure.
🤖 OpenAI plans to merge ChatGPT’s chat and Work tabs into a unified experience, while keeping Codex separate. The company confirmed development is underway but said the update isn’t rolling out yet.
📊 Anthropic reported $33.24 billion in revenue through September 2026, up 1,200% year-over-year, alongside a $104 billion net loss and $895.4 billion in compute and infrastructure spending commitments.
🛡️ OpenAI is tightening ChatGPT’s safety restrictions around violence, limiting weapons-related information for teens and reducing details about perpetrators, weapons, and acquisition methods when responding to questions about past mass attacks.
🧩 Prompt of the Day
Create Photorealistic Images With Precise Camera and Lens Settings
Image Prompt: Generate an ultra-realistic photograph of [Subject] in [Location], using specific camera hardware, lens settings, and photographic techniques to control perspective, depth of field, distortion, and overall visual aesthetics.
Context
- Subject: [Describe the subject and action]
- Location: [Describe the environment]
- Camera: [GoPro HERO / Fujifilm X-T5 / Disposable Film Camera]
- Lens: [Ultra-wide 16mm / 35mm / 85mm / 100mm Macro]
- Aperture: [f/1.8 / f/2.8 / f/8]
- Camera Angle: [Low-angle / Eye-level / Overhead / POV]
- Focus: [Shallow depth of field / Deep focus / Macro focus]
- Lighting: [Natural daylight / Direct flash / Cinematic lighting]
- Aspect Ratio: [4:5 / 9:16 / 16:9]
Instructions
Photograph the subject using the selected camera's authentic optical and imaging characteristics. For a GoPro, create an immersive action-camera perspective with an ultra-wide field of view, pronounced edge distortion, deep focus, and dynamic framing. For a Fujifilm camera, reproduce natural color science, refined tonal transitions, realistic skin tones, and subtle film-simulation characteristics. For a disposable film camera, incorporate harsh direct flash, visible film grain, imperfect exposure, slight softness, and nostalgic color shifts. Use a wide-angle lens to exaggerate perspective and environmental scale, an 85mm lens for natural background compression, or a macro lens for intricate close-up details. Apply the chosen aperture and focus settings realistically, ensuring consistent depth of field, believable lens distortion, and accurate lighting throughout the composition.
🤳AI Nugget of the Day
Thanks for Decoding with us🥳
Your feedback is the key to our code! Help us elevate your Decode experience by hitting reply and sharing your input on our content and style.
Keep deciphering the AI enigma, and we'll be back with more coded mysteries unraveled just for you!