Good morning. It's Monday, March 16, and we're covering Meta’s delayed Avocado AI model, the Anthropic-Pentagon clash over military AI, Runway’s real-time video avatars, and more.
New here? Keep up with the future of tech, sign up here. If you liked this email, share it with a friend.
THE WEEKEND RECAP
Top Stories You Might Have Missed
Runway Launches Real-Time Video AI Avatars: New Characters API turns a single image into expressive conversational video agents powered by GWM-1, enabling brands to deploy customizable avatars for support, training, and marketing across apps.
Meta Weighs Layoffs Up to 20%: Reports say Meta may cut as much as one-fifth of its workforce following recent “low performer” firings. Deeper reductions could reshape the company’s costly AI and metaverse investment strategy.
MIT’s GenCAD Turns Photos Into CAD: Researchers unveil an open-source model that converts a single image into fully editable parametric CAD commands for manufacturing, trained on 840K images and outperforming prior systems in design generation.
Spotify Lets Users Edit Taste Profiles: Spotify’s new beta feature lets Premium users review and modify their algorithmic “Taste Profile,” refining recommendations and fixing issues from shared accounts or non-representative listening.
Claude Adds 1M Token Context Window: Anthropic makes the full 1-million-token context generally available for Opus 4.6 and Sonnet 4.6 at standard pricing, supporting up to 600 images/PDF pages and enabling entire codebases or documents in one prompt.
Random Labs Launches Swarm Coding Agent: Y Combinator-backed startup unveils Slate V1, a “swarm-native” AI coding system that orchestrates multiple models in parallel threads, aiming to tackle complex software projects.
NanoClaw Partners With Docker for Agent Sandboxes: The open-source AI agent platform integrates with Docker Sandboxes using MicroVM isolation, letting enterprises run autonomous agents with stronger security boundaries.
POWERED BY HERE.NOW
Free Instant Web Hosting for Agents
here.now is web hosting for agents:
Just tell your agent to publish to here.now
Your site is online instantly, no account needed
Perfect for anything you create with your agent and need a URL for: documents, websites, dashboards, tools, school work, visualizations, presentations, prototypes, games, articles, galleries, portfolios, media files, etc.
Works inside your favorite agents like Claude Code, Codex, Cursor and OpenClaw.
Free, no sign-up required.
VIDEO
Rivian CEO Sits Down With Matthew Berman
R.J. Scaringe shares how a childhood dream became a multibillion-dollar EV company—and why AI, autonomy, and robotics will reshape the next decade.
THE FUTURE LIVE
“We paused hiring to rethink what an AI-native engineer actually looks like.”Vinay Perneti, VP of Engineering @AugmentCode:“We took a step back to think from first principles about what makes someone excellent in this new world.”“Some skills will matter less, and others… pic.twitter.com/hwRXBPaCHS
@ForwardFutureView post on XUBI
Universal Basic Income Is Your Productivity Dividend
For the first time in fifty years, it's worth asking a simple question: where did the money go? Productivity has more than doubled since 1979. Wages for most workers haven't kept pace. A RAND study put a number on the gap — $79 trillion funneled upward since 1975. That's a policy choice, not a market outcome.
Scott Santens, Founder and CEO of ITSA Foundation, argues there's only one fix that actually addresses the root problem: Universal Basic Income. Not a job guarantee. Not targeted benefits. A direct dividend — to every person — from the economy we all built. His case is sharper than you'd expect, and harder to dismiss. → Read the full article here.
MODELS
Meta Delays “Avocado” AI Model Until May After Internal Tests Trail Rivals
Meta has delayed its next foundational AI model, code-named Avocado, from March to at least May 2026 after internal testing showed it trailing leading systems from Google, OpenAI, and Anthropic. The model beats Meta's previous AI and Google's Gemini 2.5 but falls short of Gemini 3.0 on reasoning, coding, and writing benchmarks — a meaningful gap given the competitive stakes.
The delay adds pressure to CEO Mark Zuckerberg's massive AI push, which includes $600 billion in data-center infrastructure and up to $135 billion in planned 2026 spending. It also follows a rough stretch for Meta's AI ambitions: after Llama 4 disappointed, the company invested $14.3 billion in Scale AI, brought on Alexandr Wang as chief AI officer, and formed TBD Lab to build new models including Avocado and video generator Mango. Even at that scale, keeping pace at the frontier is getting harder. → Read the full article here. (Paywall)
DEFENSE
Startup General Trajectory Unveils $250 AI-Guided Drone Interceptor
Defense startup General Trajectory introduced Phoenix, an AI autonomy stack that turns low-cost FPV drones into autonomous interceptors designed to collide with hostile drones. The system runs fully onboard using a $250 NVIDIA Jetson edge computer, allowing the drone to detect, track, and pursue targets without a human pilot or radio link.
Phoenix uses a transformer-based perception model trained on more than 200,000 images and operates at over 30 frames per second with under 33 milliseconds of latency. The company argues the approach could dramatically reduce the cost of counter-drone defense compared with missile systems such as the Patriot PAC‑3, which cost millions per interceptor. → Read the full article here.
RESEARCH
AutoHarness Uses Code Wrappers to Help Smaller LLMs Beat Larger Models
Researchers including Kevin P. Murphy introduced AutoHarness, a method that lets language models automatically generate a code “harness” to prevent illegal or invalid actions when operating as agents. The system uses iterative code refinement with feedback from an environment to synthesize guardrails around a model’s decisions.
In tests across 145 TextArena games, the harness eliminated illegal moves—an issue that caused 78% of losses for Gemini 2.5 Flash in a Kaggle GameArena chess competition. With the generated harness, the smaller Gemini 2.5 Flash outperformed the larger Gemini 2.5 Pro, and in some cases, researchers generated a full code-based policy that beat both Gemini 2.5 Pro and GPT‑5.2 High on several tasks. → Read the full article here.
