Anthrophic training agents to fix AI

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.

🧠 Claude agents are training AI to behave better

Anthropic tested whether teams of Claude agents could independently research and fix AI safety problems, with the agents outperforming experienced human researchers across several common failure modes.

The Shift:

Claude handled the full research loop - Agent teams searched existing research, designed interventions, trained models, and evaluated results across 10 behaviors including sycophancy, deception, jailbreaks, and reward hacking.

Some behaviors improved dramatically - Anthropic reported improvements ranging from a 26% reduction in sycophancy to a 96% improvement on reward hacking, without reducing the models’ broader capabilities.

Claude outperformed human safety researchers - Six experienced researchers achieved a 20% improvement on deception under the same conditions, while Claude agents averaged an 85% improvement across more than 150 attempts.

Smaller models could improve stronger ones - Sonnet 5 spent 60 hours safety-training a pre-release Opus 4.8 model, using 15,000 times less data than Anthropic’s own training process.

The research suggests AI agents could automate parts of AI safety research itself, rapidly testing and training away unwanted behaviors. Anthropic’s results also show weaker models may help improve stronger systems before they are released.

Together with Tatari

The brands winning BFCM start optimizing early.

Tatari surveyed TV advertisers about their 2026 holiday plans, and brands are moving early.

Nearly 6 in 10 are increasing their BFCM TV budgets, and 54% plan to launch holiday messaging before mid-November

Launching earlier gives brands time to find what converts before TV inventory gets more competitive and expensive.

That’s how Knix approached TV. After testing ahead of peak season, one spot drove 9,000 site visits within five minutes at a 70% lower CPA. Creative testing also doubled response rates and cut CPV by 45%.

The extra runway helped Knix identify what deserved more budget before purchase intent peaked. 

By Black Friday, you don’t want to be still testing. You want to be scaling.

See how Tatari can help you scale TV for BFCM.

P.S. In NYC Oct. 29? Join Tatari at Forward for an afternoon on what’s next in advertising with Reddit CEO Steve Huffman, and marketers from Liquid IV, MANSCAPED, and more.

💼How to research and apply to any job using Claude in one session

Step 1: Drop the job description in - Paste the full JD and say: "Analyze this role. What does the company actually need? What problems are they trying to solve? What will make someone stand out?"

Step 2: Tailor your resume - "Here is my resume. Rewrite it specifically for this role. Match their language, prioritize relevant experience, and remove anything that doesn't serve this application."

Step 3: Write the cover letter - "Write a cover letter that leads with the problem they're trying to solve and positions me as the specific solution. No generic openers. No fluff."

Step 4: Prep for the interview - "Give me the 10 most likely interview questions for this role and a specific answer for each based on my resume. Include one question they'll ask that most candidates fumble."

One session. Fully customized application. Ready to send.

🖥️ Runway wants AI to generate apps as you use them

Runway introduced Solaris, an early-access “Interface World Model” that generates interactive websites and apps visually in real time, rather than relying on traditional code underneath.

The Shift:

Every frame becomes the interface - Solaris combines Runway’s Gen-4.5 video model with an LLM, interpreting clicks and drags before generating the interface’s next frames in real time.

Interfaces can respond visually to actions - Demos show users dragging clothes onto photos, assembling salads with ingredients, and interacting with simulations without predefined coded interactions.

Solaris beat coded alternatives in testing - Runway says testers preferred Solaris over pages coded by Claude Opus 5 in 71% of comparisons for in-scene behavior and 61% for instruction following.

Runway also sees an agent-training opportunity - Constantly changing generated interfaces could train AI agents to adapt to unfamiliar layouts instead of memorizing how specific websites are structured.

Solaris explores a radically different approach to software where interfaces are generated continuously instead of coded beforehand. The technology remains early, with Runway acknowledging problems around text legibility, long-session consistency, incorrect screens, and generation costs.

Together with Playbook Pro

Your Next 1,000 Customers Are Already Here

Pouring more money into ads won’t fix flat growth. The real leverage is hiding inside the customers you already have, and every day you wait, that momentum leaks away.

Networked Growth Engines is the 26-page field guide built from 200+ hours of research and real brand wins. It gives you ready-to-use systems you can plug in this week, no giant team or budget required. With it, you can:

🔥 Spark self-propelling communities using the Community Building Playbook, exactly how Nyfter built a 500,000-member Discord that drives product innovation.

📈 Design referral programs that fund themselves with the Referral ROI Calculator, the same loop that powered 34% of Blue Apron’s customer acquisition and delivered a 17.3× ROI for Latico Leathers.

🤝 Form trust-based partnerships fast using the Partnership MoU and 50-template Outreach Library to unlock new audiences without paid ads.

Your current customers are the growth engine. 

Download the guide today and start compounding growth before another week of ad spend vanishes.

🏆 Tools you Cannot Miss:

🎥 Video Agent by Fotor – Creates and edits videos through conversational instructions, allowing users to generate scenes and make targeted visual changes without traditional timeline editing.

🗣️ Orato – Analyzes recorded speeches and provides feedback on pacing, filler words, clarity, structure, and delivery for presentation and public-speaking practice.

🎙️ Radar by Particle – Searches and surfaces information from podcast conversations, making specific topics and discussions easier to find across episodes.

📱 FrameOS – Captures mobile device screens from a Mac to simplify recording app walkthroughs, demonstrations, and other mobile content.

🧑‍💻 WebTerm Learn – Teaches terminal and Git skills through interactive browser-based lessons where users practice commands in a safe environment.

🚀 Quick Hits

🇪🇺 ChatGPT will face tougher EU regulation after being designated a Very Large Online Search Engine under the Digital Services Act, requiring greater oversight of risks involving minors, mental health and illegal content. OpenAI has until the end of December 2026 to comply. 

🤖 Instagram is cracking down on AI accounts pretending to be human by introducing clearer “AI-generated profile” labels and limiting unlabeled accounts from being recommended to non-followers in Reels and Explore. 

🎬 Google is reportedly seeking licensing deals with Disney, Universal and other Hollywood studios to bring their characters and films into its AI tools, while facing concerns from creators and unions over copyright, jobs and artistic control.

⚖️ Apple says it found “shocking evidence” that a former employee used confidential Apple data while working at OpenAI, including an internal circuit schematic, and allegedly worked with an OpenAI colleague to destroy evidence after learning of Apple’s investigation. 

💰 Nvidia is investing $3.5 billion in MediaTek, giving the chipmaker access to its NVLink Fusion technology to build custom AI chips that work directly within Nvidia-based data centers as hyperscalers increasingly develop their own silicon.

👨‍💻Prompt of the day

Create a Cozy Romantic Couple Portrait

GPT Image 2 Prompt: Create an ultra-realistic warm lifestyle photograph of [INSERT COUPLE / UPLOAD REFERENCE IMAGE] sharing a quiet, affectionate moment together in a cozy sunlit room.

Context

• Subjects: [Upload couple reference image]
• Setting: Cozy home study
• Mood: Warm, intimate, comfortable, romantic
• Wardrobe: Oversized neutral knitwear
• Aspect ratio: Vertical 9:16

Instructions

Preserve both subjects’ exact facial identities, natural proportions, skin tones, hairstyles, and recognizable features from the reference image. Show the couple sitting opposite each other at a wooden study table, leaning slightly forward with one hand resting near their chin as they exchange soft, affectionate smiles. Dress the man in a loose dark sweater and the woman in an oversized beige knit sweater with relaxed, naturally styled hair. Fill the room with warm afternoon sunlight streaming through a window, creating large geometric shadows across the cream wall and table. Add subtle lived-in details such as postcards and photos on the wall, an open book, notebooks, pens, stacked books, and a ceramic mug. Use soft golden tones, natural shadows, subtle film grain, shallow depth of field, and imperfect candid framing.

🤳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!