AI Furnace Newsroom: Anthropic's "Jobs at Risk" Report Goes Viral, OpenAI’s Pentagon Deal, OpenAI’s Launch of GPT-5.4, Meta’s New AI Shopping Assistant

Welcome to this week’s AI Furnace Newsroom

In today’s insights we cover:

1. Anthropic maps the jobs most exposed to AI

2. OpenAI’s Pentagon deal is now costing it people

3. OpenAI’s launch of GPT-5.4 looks like OpenAI’s push from chat to real work

4. Meta’s new AI shopping assistant wants shopping to happen inside the chat window

Read time: 5 mins

💡 Furnace Insights

Anthropic

Anthropic Maps the Jobs Most Exposed to AI

Another week, another Anthropic report goes viral. Anthropic released a new labor-market study that tries to separate two things people often blur together: what AI can do in theory, and what workers are actually using it for today. Using Claude usage data plus task-level capability analysis, the company found the highest exposure of AI automation in white-collar fields, especially in areas like software, finance, customer support, and office administration, while hands-on jobs like cooking, mechanics, and bartending remain far less affected for now. Anthropic also says there is still a wide gap between model capability and real workplace adoption.  

That makes this feel less like a sudden layoff story and more like a slow labor-market squeeze, where there are less juniors getting hired, there are fewer humans doing routine tasks, and white-collar work starts changing from the bottom-up.

OpenAI

OpenAI’s Pentagon Deal is Now Costing it People

Caitlin Kalinowski, who led hardware efforts at OpenAI, has stepped down after the company’s Pentagon agreement, saying the deployment of OpenAI systems into classified defense environments moved ahead before the guardrails were clear enough. OpenAI has responded by saying the deal still bars domestic surveillance and autonomous weapons, and Sam Altman has since acknowledged that the rollout looked rushed.  

What makes this story bigger than one resignation is where the disagreement is happening. These are no longer outside critics warning about military AI use; the pushback is coming from inside the labs themselves. Once AI moves from a public chatbot into state infrastructure, the real question stops being “how smart is it?” and becomes “who controls where it can be used?”

OpenAI

OpenAI’s Launch of GPT-5.4 Looks Like OpenAI’s Push from Chat to Real Work

OpenAI’s GPT-5.4 release is less about a flashy new personality and more about making the model behave like a work system. The company launched three versions — base, Thinking, and Pro — added up to a 1 million-token context window in Codex and the API, improved tool search so systems only load tool definitions when needed, and positioned the model around computer use, spreadsheets, presentations, and other long-horizon tasks. OpenAI says GPT-5.4 also reduces factual errors compared with GPT-5.2 and performs better on knowledge-work benchmarks.  

Models are increasingly focused on knowledge work and getting better at handling whole projects instead of isolated prompts, which means teams can hand off bigger chunks of work without constantly re-packaging context. AI is moving from “help me write this” to “take this messy pile of material and come back with something usable.” That is a much more practical shift, and it is why every model release now feels less like a chatbot update and more like a challenge to software people already use every day. 

Meta

Meta’s New AI Shopping Assistant Wants Shopping to Happen Inside the Chat Window

Meta is testing an AI shopping feature for some U.S. desktop users that lets people ask for recommendations and get product carousels back, complete with images, prices, brand information, and link-outs to retailers. The system can also use profile signals like gender and location to shape the recommendations, though checkout still happens on the merchant’s site rather than inside Meta itself.  

This is another sign that shopping is shifting away from search bars and storefront browsing toward AI-guided discovery. The big change for brands is that winning online may depend less on buying the best ad slot and more on being the product the model decides to surface. For shoppers, it will probably feel convenient. For merchants, it raises a new, less comfortable question: what makes an AI assistant pick one product over another, and how much of that is relevance versus platform influence? The new-age SEO (or Generative Engine Optimization) wars have just begun. 

📈 AI Venture Deals of the Week

• AMI Labs, LeCun’s startup building AI “world models” for real-world reasoning, raised its $1.03B Seed.

• Legora, a legal AI platform serving some of the largest law firms in the world, raised its $555M Series D at a $5.5B valuation.

• Vertical Compute, a startup building high-density AI memory chiplets, raised its $57M Seed.

• Vivox AI, a RegTech building AI agents for financial-crime compliance, raised its $1.5M Seed.

• Rhoda AI, a platform for robot intelligence in industrial environments, raised its $450M Series A.

• Rowspace, an AI platform turning financial firms’ data into insights, raised its $50M Seed/Series A.

• Escape, an AI platform automating offensive cybersecurity testing, raised its $18M Series A.

⚒️ New AI Product Launches You Don’t Want to Miss 

• Anthropic releases Claude Code Review, an AI tool that automatically reviews pull requests and catches bugs.

• Wispr Flow, an AI voice-to-text writing assistant.

• Autoresearch, an AI agent for automated ML research experiments.

• Hyperbrowser, infrastructure for AI agents to automate web browsing.

• DomainMaps, a tool for visualizing relationships between domains.

• Skyvern, an AI agent that automates browser workflows.

Upcoming Events 📅

• AIF Founder Supper Club @ Private Room - Tuesday, March 24

Interested in meeting the Who’s Who of AI? Join the AI Hot 100 Summit in NYC

The AI Hot 100 Summit is back for its third edition on May 6-7th, 2026 in New York City. The only AI Summit built for real connections where AI Visionaries meet Industry Leaders. Join over 500 AI Executive Leaders, Founders and Investors for a jam packed 2 days of enterprise AI lightning talks, innovation showcases, and curated networking.

Get ready to meet and greet speakers and attendees from OpenAI, NVIDIA, Haleon plc, Legora, Anthropic, Cursor, Eleven Labs, Adobe, Walmart, Insight Partners & more.

🗓️ May 6-7, 2026

📍New York City