ChatGPT's Co-Creator launches a non-text AI
Plus, 📧How to let Claude run your email inbox on autopilot, Salesforce builds its own AI model for CRM work, 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.
⚡ ChatGPT Co-creator built an AI that doesn’t generate text
Former OpenAI co creator Diogo Almeida’s TypeSafe AI has emerged from stealth with Jev, a new “System One Model” designed to make fast, structured decisions inside software instead of generating open-ended responses.
The Shift:
Jev replaces text generation with preset decisions - Developers define the possible outputs in advance, and Jev selects among them with probabilities and confidence scores, which TypeSafe says prevents hallucinated outputs.
Responses arrive in 70 to 500 milliseconds - TypeSafe claims Jev can deliver comparable intelligence on its targeted System One tasks 40x to 200x faster than frontier LLMs.
TypeSafe is pricing Jev at $42 per billion input tokens - That works out to $0.042 per million input tokens, while structured outputs are currently free because TypeSafe says they are too cheap to meter.
The model is built for decisions inside automated workflows - Jev can classify, route, score and extract data, make real-time decisions, or evaluate other AI outputs for issues such as jailbreaks.
TypeSafe is betting that automation does not always need another general-purpose chatbot. By restricting outputs to predefined choices, Jev trades open-ended generation for speed, predictable structure and explicit uncertainty, potentially making AI easier to embed directly into software workflows.
Together with Zenty
Your Store Is Quietly Losing Revenue Right Now
Sending paid traffic to product pages with hidden errors makes every campaign work harder than it should.
Zenyt checks your live store continuously, catching errors your team cannot manually find across thousands of SKUs. Every finding includes visual proof and is ranked by revenue impact, helping your team fix the most expensive problems first.
Instead of reviewing pages one by one, your team gets a prioritized list of fixes backed by screenshots.
Setup takes only your store URL. No code or platform migration.
- shortyLOVE cut manual store checks by 95% and lifted conversion by 35%.
- London Victorian Ring reduced manual store-checking time by 90%.
- Brooklinen uncovered more than $250K in annual opportunity through checks after weekly releases.
That means campaigns land on accurate pages and more paid traffic turns into revenue.
Peak-season updates are moving fast. See what is costing your store before shoppers find it.
📧How to let Claude run your email inbox on autopilot
Stop living in your inbox. Here's how to set Claude up to handle it for you.
Step 1: Connect Gmail to Claude - Open Claude Cowork, connect your Gmail account, and create a new project called Inbox Manager.
Step 2: Set your VIP rules - "I want you to manage my inbox. VIPs are: [list names]. Always flag their emails immediately. Everything from [domain] is low priority. Anything with the word 'urgent' in the subject needs my attention first."
Step 3: Get your morning briefing - "Every morning at 8am, send me a briefing of my inbox: how many unread emails, the 3 most important ones, anything that needs a reply today, and anything I can ignore."
Step 4: Auto-draft replies - "For every email that needs a reply, draft a response in my voice. Do not send anything. Just draft it and label it 'Ready to Review' so I can approve with one click."
Step 5: Set the weekly digest - "Every Friday afternoon, send me a summary of the week's email activity, what came in, what got replied to, what is still waiting, and anything I missed."
Average time saved: 90 minutes per day.
🤖 Salesforce builds its own AI model for crm work
Salesforce has introduced Koa, its first in-house CRM reasoning model for Agentforce, built on NVIDIA’s open Nemotron 3 Super model and specialized for multi-step sales, service and enterprise workflows.
The Shift:
Koa was trained without customer data - Salesforce created synthetic scenarios covering personas, workflows and operational policies across 14+ industries, from generating leads and closing deals to resolving support cases.
Salesforce says Koa makes fewer mistakes on CRM tasks - On its internal CRM Bench, Koa reportedly made 3x fewer errors than leading models, with 11% higher precision when selecting the correct action.
Customer requests stay inside Salesforce infrastructure - Salesforce controls Koa’s model weights and hosts inference itself, so customer data does not need to leave its trust boundary for an external model provider.
Koa can power individual agents or entire organizations - Customers can select it for specific agents and sub-agents, make it the organization-wide Agentforce model provider, or use it through Salesforce’s managed model catalogue.
Salesforce is moving beyond connecting general-purpose AI models to CRM data and building intelligence specifically around CRM workflows. Koa gives the company more control over model behavior, infrastructure and data handling while targeting the repetitive multi-step decisions its enterprise agents perform every day.
Together with Playbook Pro
Your Growth Strategy Is Costing You. Here’s the Math
Ad costs are climbing. Most first-time buyers disappear within months, and margins are shrinking fast when discounts drive sales.
We spent 200+ hours on research, distilling what top brands do into a field guide you can read and implement in under 30 minutes.
Inside Rewriting the DTC Rulebook, you’ll learn how to:
📈 Lift retention by 5% to unlock 25–95% profit growth
🔥 Launch emotional campaigns proven to double repeat sales in 8 weeks
💰 Use the Brand Value Canvas to raise prices 15–40% without losing volume
📊 Cut paid-ad dependence by 30% in 60 days with the Owned Media Checklist
🧠 Slash decision cycles from 3 weeks to 3 hours using the Assumption Buster Toolkit
🏆 Tools you Cannot Miss:
🎥 Narrative – Edits uploaded footage through chat, letting you request cuts, motion graphics, music changes, effects, and style matching using natural-language instructions.
🔬 Anthropologic – Analyzes search, social, news, forums, and other web signals to uncover consumer trends, audience segments, cultural conversations, and brand insights.
🎨 Proofrr – Centralizes creative files, annotations, version comparisons, client comments, and approvals so design and marketing teams can manage review cycles in one workspace.
⏰ Fifi – Replaces a standard iPhone alarm with a spoken morning briefing that can read your weather, selected news headlines, and saved voice notes using a chosen character voice.
📊 siift – Organizes business ideas, evidence, decisions, actions, and results on a shared visual canvas while using AI to challenge assumptions and surface risks
🚀 Quick Hits
🏗️ 61% of likely U.S. voters oppose building AI data centers, with environmental and water concerns leading the backlash. Another poll found 7 in 10 Americans oppose local AI data centers.
⚡ Mark Zuckerberg is pushing back on calls to slow AI development, saying companies should individually train models at a safe pace while warning delays could weaken U.S. leadership against foreign rivals.
🏛️ Barack Obama says he’s neither an AI “accelerationist” nor a “doomer,” arguing major AI decisions shouldn’t be left solely to companies and calling for proactive government safety laws and regulations.
💰 The Gates Foundation is committing $1 billion over two years to expand AI access across healthcare, education, and agriculture, including more affordable tools and datasets covering more languages.
⚠️ Common Sense Media gave Perplexity its lowest youth AI safety rating, saying it failed four of five severe-harm tests, including safeguards around pornography and romantic interactions with accounts identified as minors.
🔥 U.S. data centers could consume 18 billion cubic feet of natural gas daily by 2035, more than Germany and Japan combined, potentially adding 1 million metric tons of greenhouse gas emissions daily.
🦾 What’s trending on socials
🧪 AI model takes on materials science - A Menlo Park lab says its Neon model, trained on 1,300 H200s and experimental data, surpassed GPT-6 Astra on its analysis benchmark, targeting superconductors, magnets, and semiconductors.
⚠️ Anthropic’s IPO draws an AI safety warning - A critic is calling for Anthropic’s IPO to be stopped, claiming its backers could use billions in potential proceeds to fund Effective Altruism causes and gain greater political influence.
❤️ Tesla shows how FSD helps john foppe - Tesla highlighted John Foppe, who was born without arms and has driven Teslas for over seven years, saying FSD Supervised helps him commute and visit his daughter 500+ miles away.
😂 People are starting to reject AI content - A viral meme captures a growing online trend: people increasingly labeling and seeking out “no AI” content, as frustration with AI-generated material spreads across online communities.
⚠️ Google AI researcher quits over ai risks - A Google AI researcher has reportedly quit and issued a stark warning about advanced AI, saying the technology could eventually become powerful enough to “kill us all.”
🤳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!