Editor’s Note
I say this every week, but it’s true. It’s been a crazy week for AI.
Usually I start this bit with a bit of fluff about how AI is changing how we all are and all of that. Instead, I am trying a different tactic. Quite a bit of feedback I received around the newsletter was about making it shorter and “bitesized”. Easier to consume, basically.
So, starting today on these weekly newsletters, I’ll be doing a short bulletin with all the AI news. With a small strategic take on the end of one relevant news story in the bulletin from my own perspective.
Of course, if you wanted me to cover another part of the bulletin, then just reply to this email.
From now on, this newsletter will be going out on Tuesday. The Context Tuesday has such a nice ring to it.
Here’s to more AI.
Edward
What happened last week in AI?
Anthropic accidentally leaked half a million lines of Claude Code source via a sneaky .map file in npm. Because nothing screams "enterprise safety leader" like turning your entire agentic codebase into public domain for every dev with a zip extractor. One junior mistake and boom, rivals get the full blueprint while you’re still pretending it’s secure.
Meta decided it’s done playing nice and is going full closed-source with the “Avocado” successor to Llama. After three years of handing out weights like candy, Zuck woke up and realized giving away the farm while laying off thousands wasn’t the vibe anymore. Open-source era? Officially cooked.
Mistral raised $830M in debt just to stack 13,800 more Nvidia chips for a big Paris data center. European AI summer is apparently sponsored by massive loans and the hope that the grid doesn’t black out before 2027.
Microsoft dropped multi-model Copilot workflows with Critique and Council systems so your Office suite can now argue with itself before giving you bad advice. Because one hallucinating model wasn’t chaotic enough. Now they tagteam.
Bluesky launched Attie, an AI feed builder so you can finally escape the algorithm and build your own perfect echo chamber. Meanwhile Starcloud hit unicorn status dreaming of orbital data centers, because running AI on Earth is so last season when SpaceX is right there.
OpenAI quietly shelved Sora as a standalone product (and lost that sweet Disney bag) while everyone else races to fill the video void. Turns out even the mouse house draws the line at unlimited AI slop lawsuits.
Chinese labs kept dropping omni-modal monsters like Qwen3.5-Omni that can chew through hours of audio-visual data in 113 languages, while Arm dropped its first agentic-specific CPU. The chip arms race is now officially multi-front and sweating.
Stanford had to publish a paper warning people not to blindly follow AI advice, right as companies trim middle management layers like it’s spring cleaning. 41% of workers noticing fewer bosses? Yeah, the robots aren’t just coming for coders anymore.
Shopify started wiring millions of merchants straight into AI agents, and Zoom pushed harder on agentic meeting tools. Your small business is about to get a tireless employee that never calls in sick or asks for equity.
More news? Supply chain attacks hitting Python/AI packages (Karpathy-style close calls included), Anthropic winning big enterprise deals while fighting Pentagon blacklists in court, and the usual quiet leaks hinting at Mythos-level step-changes nobody’s ready to name yet.
And it’s only the end of March.
What action can you take?
Seeing as it was in the news, let’s focus in on Microsoft’s multi-model Copilot Researcher.
What you want to do is weaponize model disagreement as your new internal "red team OS." Instead of using this as a better research tool, deploy a company wide system where every high-stakes decision memo, strategy doc, or product spec must run through a forced Critique/Council loop but with your own custom prompts that deliberately seed adversarial perspectives (e.g., "assume this plan fails like WeWork" or "attack this from a regulator's viewpoint in 2028").
Then, treat persistent model divergences as leading indicators of genuine uncertainty or blind spots. Log them systematically and route the raw disagreement traces to a small "uncertainty arbitrage" team whose job is to turn those gaps into new product features, risk hedges, or even entirely new business lines.
This turns Microsoft's hallucination-mitigation feature into an early warning system for market or regulatory shifts that single model setups miss. In regulated industries, you could even use the audit trail of Council outputs as defensible "multi-perspective diligence" in board or compliance contexts. Something regulators might start demanding in the near future.
Prompt of the Week
I used Google Nano Banana Pro for this image. Attach a selfie or a picture of yourself and use this prompt to get the above “Paparazzi” or candid effect. Change details in the prompt to suit your taste.
Use the uploaded reference image for the male subject with facial identity reproduced exactly 1:1, no alterations, no beautification, no substitutions.
The scene is a vertical 9:16 handheld iPhone photo captured inside the cockpit of a midsize private jet just before takeoff during daytime, framed in an observational, non-cinematic manner. The subject is seated naturally in the pilot seat with a relaxed but grounded posture, left hand resting near the throttle quadrant, not looking at the camera. The phone is held in the right hand and angled slightly downward toward the instrument panel and left wrist, with a subtle handheld shake consistent with a real iPhone capture.
The cockpit must be structurally accurate with all elements proportionally correct: instrument displays aligned naturally, throttle levers positioned correctly, and overhead panel switches scaled realistically, with no distorted gauges or duplicated screens. Through the windshield, a runway is visible with proper perspective lines and believable depth, reinforcing a real pre-takeoff position.
On the wrist is a Patek Philippe Calatrava in yellow gold with a dark leather strap, understated and proportionally accurate, reflecting a quiet old money sensibility rather than modern sport luxury. Clothing should follow the same tone: a crisp white or light blue dress shirt, possibly layered with a fine wool or cashmere piece, with no visible branding and a refined, inherited aesthetic.
Bright daylight enters through the windshield, creating strong highlights across dashboard surfaces and a natural glare on the watch crystal. Exposure should feel slightly imbalanced in a realistic way, with the runway appearing brighter while the cockpit interior remains more shaded.
A slight level of sensor grain should be present. The overall image must feel technical, authentic, and grounded in reality — not stylized or cinematic, but a quiet, credible moment inside aviation.
🏆 Thank you!
Just like I said above, if you have suggestions, maybe want to sponsor the fastest growing newsletter there currently is, please reply to the email.
I’m all ears!
Edward Frank Morris
The Context_
Please note that all images are AI Generated and are for illustration purposes only.
