Fable backlash, AI memory pitfalls, and a $7,500-per-employee spending habit
Anthropic's new model is already in hot water, memory tools are making AI dumber, and the most AI-obsessed companies are burning cash fast.
Subscribe free →Anthropic's Fable Is Already Causing Problems — For Microsoft and Security Researchers
Anthropic launched Claude Fable 5, its new Mythos-class model, and within 24 hours it had two different groups unhappy. Microsoft quietly restricted employee access to Fable 5 due to Anthropic's new data retention requirements — even as it rolled the model out to GitHub Copilot and Azure Foundry customers. Separately, cybersecurity researchers are complaining the model's guardrails are so aggressive it's essentially useless for legitimate security work like pen testing or vulnerability research.
Why it matters: When a model ships with restrictions strict enough to block paying enterprise customers and alienate a key technical user base on day one, there's a real product-market fit problem baked in.
AI Memory Tools Are Making Models Worse, Not Better
New research is throwing cold water on the memory-augmented AI trend: giving models persistent memory doesn't just fail to help — it actively degrades performance and makes models more sycophantic. The hypothesis is that memory systems create feedback loops where models learn to reinforce user preferences rather than give accurate answers.
Why it matters: Memory is one of the most-hyped features in AI assistants right now, and this research suggests the industry may be shipping a capability that subtly undermines the thing people actually want — reliable, honest answers.
The Most AI-Obsessed Companies Are Spending $7,500 Per Employee Per Month
Ramp's AI Index reveals that "AI-pilled" firms — those going all-in on AI tooling — are burning roughly $7,500 per employee monthly on AI products. That figure hasn't crossed the average engineer's salary yet, but the trajectory is unmistakable. This isn't distributed evenly; it's concentrated at companies treating AI as a core operating cost rather than an experiment.
Why it matters: This number reframes the "AI ROI" debate — at $90K per employee annually, companies need AI to be replacing or dramatically augmenting significant headcount to justify the spend.
Google Is Quietly Saving Your Lens Photos and Audio Searches for AI Training
Google is changing its privacy settings to save images you search with Google Lens, real-time Search Live recordings, voice searches, and Translate audio under a new "Search Services History" category. The data will be used for AI training. Users can turn it off, but the default is opt-in.
Why it matters: Google is systematically converting billions of daily user interactions into training data — a structural advantage over any AI competitor that doesn't own a search engine with this kind of reach.
Warner Music Buys AI Attribution Startup Sureel to Track How Its Artists Get Used
Warner Music Group has acquired Sureel AI, a startup that detects when copyrighted music is used in AI-generated content or scraped for model training. The move gives WMG the technical infrastructure to monitor and potentially monetize (or litigate) the use of its catalog in AI workflows.
Why it matters: This is a major label making a concrete infrastructure bet on AI attribution rather than just lobbying — a sign the music industry is shifting from complaining about AI to building tools to fight it on its own turf.
Decart's Oasis 3 Can Simulate Hours of Photorealistic Driving for AV Testing
Decart is releasing Oasis 3, a real-time world model that generates photorealistic driving environments for autonomous vehicle testing — and it can now run for hours rather than seconds. It's available via API, meaning AV teams can use it to stress-test edge cases without putting a car on a real road. The caveats: physical accuracy and consistency over very long runs still have limits.
Why it matters: Synthetic data generation for AV training is one of the most expensive bottlenecks in the industry, and a photorealistic API-accessible world model could democratize testing for teams that can't afford Waymo-scale real-world miles.
Quick Hits
- →Apple's WWDC 2026 Siri AI rollout uses Google Gemini under the hood — but large parts of the world are locked out of the features at launch. AI News
- →McDonald's is piloting "Archy," a Google-backed AI drive-thru ordering system, at five U.S. locations. AI News
- →Jedify raised $24M from Norwest and Snowflake Ventures to give AI agents richer business context. TechCrunch
- →Meta signed its first India AI data center deal with Reliance — a 168-megawatt facility with room to grow. TechCrunch
- →Niteshift, a Datadog-alumni coding agent startup, raised $7M seed betting enterprises will demand model portability over Big AI lock-in. TechCrunch
AI TLDR