Wednesday, September 16, 2026
Today I'm watching OpenAI hand its newest brain a briefcase, Salesforce build a rival model just for salespeople, and pollsters confirm nobody wants a data center next door. Let's get into it.
OpenAI's Newest Model Is Built to Do Your Office Job
OpenAI just introduced GPT-6 Astra for work, a version of its most capable model tuned specifically for business tasks. It can reason through complex problems, operate a computer on its own (clicking, typing, navigating software like a person would), and write or design things with better judgment than earlier models. OpenAI is already showcasing companies that trust it with real jobs: Perplexity lets it change software and watch over live systems with far less human checking, and coding startup Cognition uses it to help its AI assistant Devin test its own work before a human reviews it. The pitch is simple: let the AI handle more of the routine work so people can focus on judgment calls.
What this means for you: If your job involves reviewing drafts, checking software, or monitoring systems, expect an AI coworker that needs less babysitting than the one you tried last year.
What this means for your business: Companies that let AI handle more end to end work, with spot checks instead of constant supervision, could cut costs and speed up delivery faster than competitors still treating AI as a first draft machine.
Source: OpenAI Blog
Radar 01
Salesforce Built Its Own AI Brain, and It's Aimed Right at OpenAI
Salesforce unveiled Koa, its first AI reasoning model made specifically for sales, marketing, and customer support work. It runs on Nvidia's open source Nemotron model rather than renting intelligence from OpenAI or Google. Instead of trying to be smart at everything, Koa is trained narrowly to be very good at the tasks Salesforce customers actually pay for, like handling deals and support tickets.
What this means for you: The AI helping you close deals or answer support tickets might soon come baked into the software you already use, not bolted on from a separate chatbot.
What this means for your business: Software companies with deep customer data are starting to build their own specialized AI instead of buying it, which could squeeze margins for general purpose AI labs and give niche players a real edge.
Source: TechCrunch
Radar 02
Voters Really Don't Want a Data Center Near Them
A new poll from The New York Times and Siena University found 61 percent of likely voters oppose building data centers to power AI. This confirms a trend already showing up in local fights from Massachusetts to Philadelphia, where communities worry about noise, power bills, water use, and land already scarred by past industry.
What this means for you: If a data center gets proposed near your town, you're in good company being skeptical, and local officials are increasingly listening to that skepticism.
What this means for your business: Public opposition is becoming a real business risk for anyone building AI infrastructure, meaning permits, timelines, and costs for new data centers could all get harder before they get easier.
Source: The Verge
Radar 03
Meta Lets AI Agents Set Up Your WhatsApp Business Account
Meta launched a new WhatsApp Business tool called an MCP server, which is a standard way for AI coding assistants like Claude, Cursor, Codex, and ChatGPT to plug directly into a business's setup. Instead of a person manually configuring messaging templates, testing, and troubleshooting, an AI agent can now do that legwork automatically.
What this means for you: Setting up a business messaging system that used to take a developer hours could soon take an AI agent minutes, even if you don't have technical staff.
What this means for your business: Small and mid-size businesses get cheaper access to tools that used to require hired developers, while Meta locks more of the AI agent ecosystem into its own platform.
Source: TechCrunch
Try This Today
Pick one recurring task where you or your team check an AI's work line by line, like reviewing drafted emails or code changes. Try letting the AI complete the full task once with only a final spot check, instead of supervising every step, and see how much time you get back.
Quick Hits
- Fyxer, an AI email assistant, uses OpenAI models plus memory and real user feedback to sort inboxes and draft messages that actually sound like you wrote them, part of a growing wave of AI tools built to earn trust with everyday office work. [1]
- A new Mozilla-backed report found that cheap, open source AI models from Chinese developers have nearly caught up to expensive frontier models, meaning paying premium prices for the best AI now only buys about a four month head start at five times the cost. [2]
- A study comparing over 2,500 head to head tests between AI and traditional scientific methods across 27 fields found AI clearly outperforms in some disciplines but not others, a useful reality check against blanket claims that AI beats every research method everywhere. [3]
Enjoyed today's read? Help me get this in front of more people like you.
Start every morning a little sharper.
Free. Takes 5 minutes to read. No spam.
