Mojo AI Brief

🔥 THE LEAD

AR glasses plus LLMs can profile and manipulate someone in real time. Your clients are the target.

Someone wears AR glasses, points them at your client, and an LLM builds a social profile in seconds. The attacker gets real-time prompts to manipulate that person. Elder care and high-net-worth clients are obvious targets.

⚡ MUST READ -- changes how you build or sell

FAMA shows how small open-source LLMs spiral into cascade failures when one tool call goes wrong

Smaller open-source models fail in agentic settings not because of one mistake, but because each mistake compounds the next. FAMA adds a meta-agent layer that detects those failure patterns early and corrects before they cascade. That's the gap it's targeting.

TIER 2 arXiv cs.CL

a16z Is Recruiting Founders to Bet Against Workday

a16z argues Workday has stayed dominant by being hard to replace, not by being good. Their pitch: HR and workforce management software is bloated, slow, and ripe for a leaner AI-powered competitor to take over. This is a VC deal-flow post dressed as analysis.

Microsoft's AGI escape clause is officially dead

For years, OpenAI's deal with Microsoft included a clause saying Microsoft's commercial rights evaporate the moment AGI is reached. That clause just quietly disappeared. Microsoft and OpenAI renegotiated it out of existence, which means Microsoft's grip on OpenAI's commercial output is now cleaner and more permanent.

TIER 4 Simon Willison

📡 ON YOUR RADAR -- developing this week

a16z publishes a framework for how abstractions fail. Worth a skim, not a strategy shift.

The piece outlines how layers in complex systems fail when the layer below them bleeds through. It's a solid concept, but it stays high altitude. No new tooling, no spec, nothing that changes how agents are built today.

Marginal Revolution's link roundup floats an idea: models that know what they're good at and bid for work accordingly

The idea is that agents shouldn't just execute work handed to them. They should self-assess and compete for assignments based on fit. It's a research-horizon concept, not something you can wire up in LangGraph today.

RADAR Marginal Revolution

GPT-5.5 is out. Zvi has the system card breakdown.

GPT-5.5 is another incremental OpenAI release. Zvi's system card analysis is the most honest independent take you'll find on what the model actually does versus what OpenAI claims it does. Worth a bookmark, not a stack change.

RADAR Zvi Mowshowitz

🛡️ SECURITY WATCH -- for your SOC brain

AR glasses plus LLMs can profile and manipulate someone in real time. Your clients are the target.

Someone wears AR glasses, points them at your client, and an LLM builds a social profile in seconds. The attacker gets real-time prompts to manipulate that person. Elder care and high-net-worth clients are obvious targets.

SECURITY arXiv cs.AI

🙅 SKIP TODAY

Independent-Component-Based Encoding Models of Brain Activity During Story Comprehension (arXiv cs.CL) — Academic neuroscience paper on fMRI brain encoding. No connection to Joey's stack, his SMB/mid-market customers, or anything he can build or sell.

Introducing talkie: a 13B vintage language model from 1930 (Simon Willison) — A novelty historical model with no practical deployment path. Qwen3 on the Jetson already handles local inference; this adds nothing to that story.

Quoting Matthew Yglesias (Simon Willison) — Political commentary routed through an AI blog. No stack relevance, no customer angle, no signal worth your 13-days-to-Boca attention.

🎙️ THE MOJO TAKE -- the editorial

OpenAI just published an open orchestration spec called Symphony. They also moved their models onto AWS the same week. Pick a lane.

Today: Symphony wants to be the standard for agent coordination, OpenAI lands on AWS Bedrock, and what either move means if you're running LangGraph and Temporal instead of their stack.

Connect the dots. Symphony, Managed Agents on AWS, and the Microsoft-OpenAI lock-in all point the same direction: the big players are packaging decisions your clients used to pay consultants to make. That packaging will kill conversations before Joey gets in the room. The HubSpot pipeline seed and the public launch on May 26 have to answer that before the pre-built narrative sets.

Watch next. Watch whether Symphony gets real adoption in the Jira-native shops OpenAI is targeting. If it does, Joey's LangGraph pitch needs a sharper 'why custom' answer ready. Also watch how AWS prices the Managed Agents tier. That number will show up in client procurement decks fast.

13 days to Boca. 29 days to launch. The pipeline doesn't seed itself. Go harden that MCP and ship something real this week.