This AI Avatar Fooled Nearly Half of Testers

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.

🎥 This AI video avatar convinced nearly half of testers it was human

AI startup Tavus has previewed Griffin, a new “Human Interaction Model” built to handle face-to-face conversations in real time, combining vision, hearing, speech and lifelike video generation in one system.

The Shift:

It Watches While Talking - Griffin continuously sees and listens during conversations, responding to expressions, pauses, gestures and visual details instead of simply waiting for someone to finish speaking.

48% Thought It Was Human - In Tavus’s live face-to-face study, 48% of participants believed Griffin was a real person, compared with a maximum pass rate of around 2% for its previous system.

Video Chat Gets More Natural - Griffin can nod while someone speaks, react through facial expressions and reference objects visible on-screen, with perception and audiovisual generation happening simultaneously.

AI Nears Human Scores - Tavus says Griffin came within 0.09 points of humans on Nvidia’s VideoFDB benchmark for natural video interaction and beat the next-best AI system by more than one point.

Griffin-Lite is currently limited to select early testers, while Tavus works on safety and disclosure before a wider release. The bigger shift is from avatars that simply answer prompts to video agents that follow the full rhythm of a conversation, including what someone says, sees, and communicates without words.

Together with Modash

You See Their Creator Campaigns. Now Hear Their Thinking. 

A creator campaign can look great in your feed and still leave you wondering whether you’d pay to run it again.

Repeating that uncertainty with more creators gets expensive.

On October 21, Modash’s Return on Influence Festival ’26 opens up how established teams run their programs, including the mistakes behind better campaigns.

Here’s what you’ll learn:

• How teams run campaigns day-to-day, including what they changed after mistakes.
• What creators want from brand partners so that you can approach collaborations from their side.
• What it takes to scale from 5 to 5,000 creators before adding more to your own program.

You’ll hear from Maya Shaff, Head of Influencer at ŌURA, and Josh Rangel, Senior Director, Social Strategy at Ogilvy. More to be announced. It’s online and free.

Reserve your spot for Oct 21st and leave with actionable insights to build and scale a creator program that moves the needle. 

Can't attend every session live? Register anyway, and you’ll get the recordings in 24 hours.

✨How to connect Meta ads to ChatGPT and manage campaigns without opening Ads Manager

Meta released an official connector that plugs your ad account directly into ChatGPT. Analyze performance, find what is bleeding budget, and make real changes, all through conversation.

Step 1: Get the official Meta Ads MCP link - Search "Meta Ads MCP ChatGPT" and find the official link on meta.com or developers.facebook.com. Copy the URL ending in /mcp. Official only, no third-party connectors.

Step 2: Enable Developer Mode - ChatGPT Settings > Advanced Settings > Developer Mode. Toggle it on.

Step 3: Connect your ad account - Apps section > Create App. Name it Meta Ads, paste the MCP URL, click Create. Authorize your Facebook account. Live in 5 minutes.

Step 4: Run the audit first - Before touching anything:

"@Meta Ads Analyze all active campaigns. Show spend, results, CPA, CTR, and ROAS. What are the three biggest issues?"

Read the diagnosis. Change nothing yet.

Step 5: Fix with guardrails - Before every change:

"List every change you plan to make, explain why, flag anything involving increased spend, and wait for my approval."

Never increase budgets more than 20% at once. Big jumps reset Meta's learning phase.

Step 6: Weekly report every Monday - "@Meta Ads 7-day report: spend, ROAS, top 3 ads, bottom 3 ads, and next week's recommendations."

Scale winners by 20%. Pause anything running 5+ days above your target CPA.

⚡ AI agents are getting specialized models just to make decisions

Cloudflare and Amazon have both released open-source decision models, a fast-growing category inspired by TypeSafe’s Jev. Instead of generating long answers, these models choose between predefined options and return confidence scores that software agents can act on.

The Shift:

☁️ Cloudflare Launches Clef

Clef and Clef-flash support structured decisions, images and a 64K context window. In Cloudflare’s domain test, Clef completed the workflow in 2.2 seconds versus 4.7 seconds for its fastest general LLM.

Cloudflare is also building an RL fine-tuning platform where companies can use their own data to customize Clef for tasks like support triage, Trust & Safety and bot detection, then redeploy it on Workers AI.

📦 Amazon Builds Its Own Decision Model

AWS released Strands Decider 2B, an open-source model based on Qwen3.5-2B that is small enough to run locally and returns calibrated choices rather than generating text.

AWS sees it handling workflow questions like “what should happen next?” without calling a full LLM every time, potentially giving agents lower latency, lower costs and more predictable outputs.

Decision models are quickly becoming a new building block for AI agents. Instead of using a powerful LLM for every small choice, developers can hand routine decisions to specialized models built to be fast, structured and confidence-scored, then reserve larger models for tasks that actually need deeper reasoning.

Together with Playbook Pro

Marketing That Pays Dividends Month After Month

Most marketing dies the moment the spend stops.

The brands winning now design systems where every customer, partner, and conversation keeps growth compounding long after launch.

Networked Growth Engines is a 26-page guide built on live operator data and real growth experiments. Inside, you’ll learn how to:

🔥 Design referral flywheels with the Referral ROI Calculator, crafted with brands that lifted referral-driven revenue by 40% in their first 60 days.

🤝 Forge cross-brand partnerships using the 50-template Outreach Library and ready-to-sign Partnership MoU to double audience reach in under 90 days.

📈 Activate true communities proven to boost repeat purchase rates by up to 28% within a single quarter.

Your current customers are the growth engine. 

Get the ebook and Build Your First Growth Loop Now!

🏆 Tools you Cannot Miss:

🎬 Clueso MCP – Creates and edits videos through ChatGPT, Claude, or Gemini from an idea, recording, deck, document, or reference video, handling storyboards, scenes, voiceovers, music, and editing. 

🌍 Vitra.ai – Creates, translates, personalizes, reviews, and publishes videos, images, documents, websites, and other content across 100+ languages from one workflow. 

🎙️ Voiskey – Turns rough voice input into polished posts, scripts, emails, notes, and other written content while adapting formatting and tone to where the text will be used.

💬 Chat.sh – Builds an AI-searchable help center on your own website that answers visitor questions from your documentation and links back to the source articles it used. 

📤 Buddy Drop – Turns a folder, ZIP, or website files into a live shareable URL in seconds, with hosting, previews, versioning, and custom-domain support handled automatically.

🚀 Quick Hits

🔍 Google updated its Search guidance to say it is “critical” to manually fact-check and review all AI-generated content before publishing, warning that generative AI can produce inaccurate or hallucinated information.

👁️ Google launched Guided Vision in Gemini Live, using a phone’s camera to provide real-time audio descriptions, read small text, identify objects and answer follow-up questions for blind and low-vision users.

🛰️ Google successfully launched an AI satellite equipped with its TPUs aboard a SpaceX Falcon 9, marking an early Project Suncatcher experiment toward potentially building AI data centers in space.

🔐 OpenAI reportedly parted ways with three safety researchers after they allegedly shared confidential company information with an outside AI safety organization. OpenAI confirmed the departures to The Wall Street Journal.

👗 ChatGPT launched virtual try-on globally, letting users upload photos to see how clothing and accessories could look on them. A new Favorites feature also saves products and try-on images for later.

🥳 What’s trending on socials 

😂 Always read the Y-Axis - A viral chart claiming GPT-6.1 Sol is “crushing” Opus 5.5 is getting mocked because the vertical axis simply says “Model number,” meaning it doesn’t actually compare model performance.

🤯 AI video passes the Turing test - Griffin claims to be the first model to pass a video Turing test, with 48% of people believing it was human during live conversations, versus under 3% for previous systems.

🕷️ Web Crawlers, Literally - A viral animation turns web crawling into an actual spider crawling across a webpage, visually connecting links and references as it moves through the page.

✝️ Grok Sparks Viral Christianity Debate - A viral Grok interaction claims the AI removed every “false god” from an image except Jesus Christ, before responding to users with Christian messages including “Christ is King.”

🔬 Claude Accelerates Scientific Research - Anthropic says a Claude-powered system helped produce 36 manuscripts across 18 scientific fields in three months, tackling roughly 400 candidate problems and completing some research tasks that previously took scientists weeks or years.

🤳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!