TLDR AI 2026-03-11

👉 Here's how to put it all together

META ACQUIRED MOLTBOOK

Meta has acquired Moltbook, a Reddit‑like network where AI agents built with the OpenClaw framework interact with each other and maintain an always‑on directory of agents.

NVIDIA INVESTS IN MIRA MURATI'S THINKING MACHINES LAB

Nvidia and Mira Murati's Thinking Machines Lab have formed a multiyear partnership in which the startup will deploy at least one gigawatt of cutting-edge chips to train and serve its frontier models. The companies will collaborate to design AI training and serving systems that use Nvidia technology. The deal gives the relatively young AI lab computing power to advance its research, as well as funding to pay its staff. It currently has 120 employees and is competing fiercely for AI researcher talent as valuations in the industry have skyrocketed.

GOOGLE LAUNCHES NEW MULTIMODAL GEMINI EMBEDDING 2 MODEL

Google's Gemini Embedding 2, available via the Gemini API and Vertex AI, unifies text, images, videos, audio, and documents in over 100 languages. The model processes up to 8,192 text tokens, six images, 120-second videos, and six-page PDFs, incorporating Matryoshka Representation Learning for customizable output dimensions. Early access users already utilize it for Retrieval-Augmented Generation and semantic search, with initial feedback showing superior performance against competitors.

CODEX, FILE MY TAXES. MAKE NO MISTAKES

Codex can be used to file personal taxes, and it can even be more accurate than a human accountant. The immediate feedback that Codex gives users helps them understand their situation and the tax code better. Give Codex access to personal details, and it can make better assumptions about tax treatment. Accountants probably don't have to worry - most don't enjoy doing personal tax returns, and their other duties are likely much harder to automate.

OPEN WEIGHTS ISN'T OPEN TRAINING

Open source models offer the proposition of distributing the value created by AI more broadly, enabling more people to build. However, the current ecosystem doesn't make building easy. The stack contains many bugs, and there's debt hidden in every layer. Developers need to stop patching bugs and build a better stack.

THE STATE OF CONSUMER AI. PART 2: ENGAGEMENT AND RETENTION

ChatGPT's engagement lead is wider than its market share lead: DAU:MAU sits at 45% versus Gemini's 22%, and WAU:MAU has climbed from 50% in mid-2023 to 82% today, putting it ahead of Gmail and Spotify and approaching Instagram. Week 4 retention of 66% beats every enterprise app in the dataset and is more than double Perplexity's 24%, while most apps see retention flatten at scale. The rarest signal is ChatGPT's smile curve: one of only three products alongside Gmail and Chrome where retention dips and then recovers, meaning OpenAI's product releases are actively reactivating lapsed users, not just attracting new ones.

open source Python repositories. See what each benchmark measures

QUANTIFYING INFRASTRUCTURE NOISE IN AGENTIC CODING EVALS

Agentic coding benchmarks are commonly used to compare the software engineering capabilities of frontier models. These scores are often treated as precise measurements of relative model capability. However, research shows that infrastructure configuration alone can produce significant differences in scores. Eval developers have begun accounting for this, but current fixes potentially change what benchmarks end up actually measuring.

RCLI

RCLI is an on-device voice AI for macOS that can control apps and perform other actions via voice. Users can choose from a variety of local AIs and perform 38 macOS actions. The tool can be used to ingest local documents for users to query. RCLI is open source under the MIT license.

THE ANATOMY OF AN AGENT HARNESS

An agent is a model with a harness. Harness engineering is the process of turning models into work engines by building systems around them. Models contain the intelligence, and the harness makes that intelligence useful. This post looks at what harnesses are and details the core opponents that agents need.

HOW NVIDIA BUILDS OPEN DATA FOR AI

Every AI training pipeline rests on a data layer that determines how those models behave. This data determines what the models know, how they reason, and what they can safely do. However, much of today's training data remains opaque, fragmented, or siloed across teams. Open data access gives developers a faster and more cost-effective path to building high-quality models. This is why Nvidia releases open datasets along with its open models, tools, and training techniques.

AMAZON WINS COURT ORDER TO BLOCK PERPLEXITY'S AI SHOPPING AGENT

Perplexity's Comet AI browser has been blocked from accessing Amazon's site. Amazon sued Perplexity in November, alleging the startup concealed its AI agents so it could continue to scrape Amazon's website without approval. The Comet browser allows users to ask the assistant to find items on Amazon and make purchases. Amazon says Perplexity's agents pose a risk to customer data and that they create challenges for its advertising business. It has broadly locked down its shopping sites from AI agents and blocked dozens of agents.

NOTES FROM TOKEN TOWN: NEGOTIATING FOR THE FORTUNE 5 MILLION

Frontier labs are continuing to build first-party products on top of their own capabilities. For every token they sell, they can spend it themselves for COGs of under 50%. Your supplier is also your competitor, and they have a permanent cost advantage over you. Applied AI companies need to win on something other than tokens. Businesses need to build products so good that the token itself becomes the commodity input, not the value proposition.

YOUR DATA AGENTS NEED CONTEXT

Data and AI agents struggle without proper context, and messy and disparate enterprise data complicates their ability to answer basic queries.

THE ERA OF “AI AS TEXT” IS OVER. EXECUTION IS THE NEW INTERFACE.

GitHub Copilot SDK enables AI-driven execution directly within applications, moving beyond simple text interactions.

INSTRUCTION HIERARCHY TRAINING FOR SAFER LLMS

OpenAI's IH‑Challenge is a dataset designed to train models to prioritize instructions based on trust level across system prompts, developers, users, and external data.

CLAUDE-GROUND

claude-ground introduces a minimal rule system for Claude Code, providing phase tracking, decision logging, and language-specific best practices to improve coding discipline.

SELECTIVELY REDUCING EVAL AWARENESS AND MURDER IN GEMMA 3 27B VIA STEERING

Google's Gemma 3 models, including the 27B variant, were steered to alter features correlating with evaluation awareness and the intent to murder.