Good evening. Here is what matters in AI today, and how to put it to work.
We see AI's real-world costs crystallizing this week: copyright settlements spark new disputes, and OpenAI reframes how we should think about what models actually do.
~4 min read · last 12 hours
In today's issue
01
Authors push back as publishers claim Anthropic settlement funds
02
Travis Kalanick's Atoms eyes the robotaxi market
03
The universal geometry of embeddings, and how to exploit it
04
Google open-sources Mantis, an AI agent for vulnerability scanning
05
OpenAI publishes "An Alien Mind," exploring how AI reasoning differs from human cognition
Main story
Authors push back as publishers claim Anthropic settlement funds
Authors say publishers and agents are claiming a disproportionate share of payments from Anthropic's copyright settlement, raising questions about who actually benefits when AI companies settle IP disputes.
Why it matters: If your AI product ingests third-party content, watch this case closely: the fight over how settlement money flows will shape how future licensing deals are structured and who holds negotiating power.
What to watch next: Watch whether the court accepts authors' objections to the distribution formula: a ruling in their favor would set a precedent that intermediaries cannot automatically take a cut of AI copyright settlements, which would directly affect how publishers negotiate future training-data licenses.
We see two stories this week that share a common thread: the real-world consequences of deploying AI are landing on humans, whether that means authors fighting over settlement money or a founder betting his legacy on self-driving robots.
The Anthropic settlement fight is a preview of a much larger problem: as AI copyright cases resolve, the question of who captures the value, original creators or intermediaries, will define the economics of licensed training data for years. At the same time, Google's Mantis release and the embeddings geometry paper both signal that the industry is maturing past "can we build it" toward "can we make it reliable and safe in production." OpenAI's "alien mind" framing adds a third pressure: if we misunderstand what models are doing at a conceptual level, we will keep designing human-in-the-loop systems that are miscalibrated from the start.
All the best, the KYFEX team
Quick hits
AI liability and autonomy collide in courts and garages
Travis Kalanick's Atoms eyes the robotaxi market
The Uber founder says Atoms will let him finish "unfinished business," signaling a renewed push into autonomous ride-hailing years after Uber sold its self-driving unit.
Why it matters: Another well-funded entrant in robotaxi raises the competitive pressure on Waymo and the Tesla robotaxi rollout, and signals that the autonomous mobility market is heating up again for investors and fleet operators.
Under-the-hood AI: embeddings, agents, and security tooling
Two research and tooling items this week point at the same underlying shift: teams are moving from building models to understanding and hardening what models already do, whether that means exploiting the shared geometry of embedding spaces or using agents to cut security noise.
The universal geometry of embeddings, and how to exploit it
A new paper argues that embedding spaces across different models share a consistent geometric structure, which could unlock better transfer learning, cross-model retrieval, and alignment techniques.
Why it matters: If your team works with RAG pipelines or multi-model search, this research suggests you may be able to reuse and align embeddings across models more cheaply than previously assumed.
Google open-sources Mantis, an AI agent for vulnerability scanning
Google's Mantis framework uses an AI agent to automate software vulnerability scanning with a focus on reducing the false positives that make traditional scanners exhausting to operate.
Why it matters: False-positive fatigue is one of the biggest reasons security tooling gets ignored in practice, so an agent-driven approach that filters noise before it reaches a human reviewer is worth evaluating in your CI/CD pipeline.
OpenAI publishes "An Alien Mind," exploring how AI reasoning differs from human cognition
OpenAI's essay argues that frontier AI systems reason in ways that are genuinely alien to human cognition, not simply a simulation of it, which has implications for how we interpret model outputs and set expectations.
Why it matters: Engineering and product teams that treat model outputs as human-like reasoning will make systematic errors in system design; this framing is a useful corrective when scoping human-in-the-loop requirements.
Build, debug, and optimize AI agents using DSPy and MLflow.
DeepLearning.AI · Free · 1 hour
Put it to work
Try this today
Audit a RAG pipeline for embedding alignment gaps
I have a retrieval-augmented generation pipeline that uses [Model A] to embed documents and [Model B] to embed queries. Recent research suggests embedding spaces share a universal geometry across models. Help me identify: 1) the top three ways a mismatch between these two embedding spaces could silently degrade retrieval quality, 2) a simple diagnostic test I can run today to detect drift, and 3) one practical fix I can implement without reindexing my entire corpus. Be specific and concise.
Why it helps: The universal embeddings geometry paper suggests cross-model alignment is more tractable than assumed, making this a good moment to pressure-test your own pipeline before that assumption bites you in production.
Before you ship it
The risk
AI vulnerability scanners like Mantis can reduce false positives, but they can also introduce false negatives: an agent that learns to suppress noisy alerts may also suppress real ones that look similar, creating a blind spot that is harder to detect than a noisy scanner.
Do this
Run any AI-assisted security scanner in parallel with your existing tool for at least one full release cycle, comparing flagged and suppressed findings side by side before you retire the legacy scanner or reduce human review.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
