📝 12/05/25: New Claude Models, AI Agent Teams and AEO vs SEO
📌 This Week’s Top News Highlights: New Claude Models, OpenAI’s Safety Hub, and Google’s AI Startup Push
Updated Claude Models
Anthropic is reportedly working on a new generation of their Sonnet and Opus models designed to think more autonomously and fix their own mistakes, switching from reasoning to tool use. This could take everyday, conversational AI a step closer to true agent-like behaviour.
OpenAI Launches Safety Evaluations Hub
OpenAI has launched a new Safety Evaluations Hub to track and test its models for safety risks, including misinformation, misuse, hallucinations and unintended outputs. It’s a small but significant step towards more accountable AI.
Google’s New AI Investment Initiative
Google has announced a new initiative to back AI-first startups, offering funding, technical support, and early access to cutting-edge tools. If you’re building in this space, it’s worth a closer look (they're taking applications on a rolling basis).
🕵️ Founder, Anonymous
A founder in the clean tech space I spoke to this week raised an interesting point: as the complexity of AI systems grows and AI agents are "employed", we might soon need dedicated teams to manage a fleet of AI agents within the same company.
They even told me that they are currently reorganising their teams around this provocative idea, and that this will likely happen in the next 12-24 months. The notion that human teams will manage the “team dynamics” of AI agents, coordinating their goals, optimising their outputs and troubleshooting conflicts, feels like a spooky glimpse into the future of work.
💡 Thought of the Week
AEO vs SEO - What’s the Difference?
As search engines get more sophisticated, there’s a growing need to optimise not just for human queries (SEO), but also for AI-driven search (AEO, or Answer Engine Optimisation).
This means focusing on:
• FAQs - Clear, concise answers to common questions
• Featured snippets - Short, direct responses that can get pulled to the top of search results
• People Also Ask (PAA) - Question-based headings and concise answers to appear in related search boxes
• Schema markups - Use tools like Google’s Structured Data Markup Helper or plugins like Yoast SEO to add structured data to your pages and help AI better understand your content
Practical tip: Run your site through Google’s Rich Results Test to check if your schema markup is being read correctly and identify pages that could be optimised for featured snippets or PAAs.
To take this a step further, Meta planning director Peter Buckley recently shared this interesting diagram on the direction of search, which he called The Great Search Split. Essentially, SEO isn't enough - you need REO "Recommendation Engine Optimisation" (filling the feed) and AEO "Answer Engine Optimisation" (owning the answer). In the new search world, you're competing for both prediction and preference.
🔗 Sneaky Links
- Klarna workforce shrinks by 40% - Klarna has reportedly shrunk its workforce from ~5,000 in December 2022 to now ~3,000, as a result of its aggressive adoption of AI, and "natural attrition"
- Hidden Cost of AI Productivity - HBR have reported that AI tools boost productivity, but undermine workers’ intrinsic motivation and increase feelings of boredom when they return to other tasks without using AI
- Perplexity’s in-chat eCommerce upgrade - Perplexity is now integrating PayPal for smoother checkout experiences in-app, signalling a shift from search as "soft" intent to hard conversion. OpenAI are also reportedly integrating with Shopify to offer more seamless shopping experiences for users of ChatGPT
- DeepMind’s AGI Safety Framework - DeepMind released a new framework outlining four key risk areas for AGI: misuse, misalignment, unintended outputs, and long-term structural risks. It emphasises proactive risk management, including model training, monitoring, and system-level safeguards
- OpenAI launches new AI coding agent Codex - which is powered by a version of the company’s o3 AI reasoning model optimised for software engineering tasks. They say this model produces “cleaner” code, follows instructions more precisely, and will run tests on its own code until passing results are achieved.