TLDR AI 2026-09-16

Current solutions are complex, but they can't see API calls outside of proxy, subprocesses, or config rewrites made by agents. That's unsustainable in the agentic era.

Ory Agent Security works at the harness layer to give you the ability to govern execution before a gateway call. What's more:

• Setup in seconds with auto-install
• Config your agent security by toggling preferences
• Get instant visibility into permissions and make changes with a few clicks in the permissions modeler

Ory Agent Security is vendor-agnostic and works with 11 major AI coding agent harnesses.

See how it works

Periodic Neon outperforms GPT-6 Astra and Claude Fable 5.1 at lower cost on a highly challenging evaluation for scientific analysis. Neon is now deployed in labs to analyze experiments for better superconductors and magnets. The model establishes a Pareto-optimal cost-performance frontier through midtraining and reinforcement learning on lab data. Periodic Neon surpasses frontier models on FrontierXRD at a lower cost per analysis.

Google released Gemini 3.8 Live and Gemini 3.5 Transcribe for developing real-time voice applications. These tools offer enhanced features for voice recognition, transcription, and live interaction capabilities. They aim to improve accuracy and user experiences in voice-driven applications.

Jev is a System One Model that achieves similar levels of intelligence on System One tasks compared to existing large language models while being two orders of magnitude faster and more efficient. System One Models are a new class of frontier models built to make fast, structured decisions that software can use directly. Jev is optimized for structured outputs and can not hallucinate. It is now available in early access.

This developer has been testing an approach to charging AI agents for using their content. Their page sets a price through the x402 protocol before an agent reads it. If the agent wants access, it pays. While the developer has yet to receive any real payments, testing shows the system works, so it's worth looking into if you own content and aren't big enough to negotiate fair licensing deals.

Standard RL results in disproportionately poor performance on the hardest problems. Simple scalar values may not be sufficient for accurate evaluations of large language models. This paper proposes a solution that reduces the compute spent on easy problems, which reallocates it towards harder problems. The solution appears to show substantial improvements on difficult tasks.

Note-takers transcribe calls and send you a bloated summary. Granola is an AI notepad: you jot down what matters, Granola enriches (no bots). When the meeting's over, your notes come to life: chat, generate emails, prep for the next meeting, or pull takeaways. Use promo code TLDR1MO and get your first month free

Provides an open-source humanoid arm for teleoperation, imitation learning, and physical AI research.

Progress in recursive self-improvement hinges on effective exploration. Accumulated discovery history can serve as a replay simulator over a realized search space. Dream-RSI closes a self-improvement loop at the exploration layer. It continuously collects discovery histories through online exploration, constructs replay simulators from them to refine meta-exploration strategies via dreaming, and redeploys the upgraded policy online. Dream-RSI improves both discovery effectiveness and efficiency in several settings.

Recursive scientific AI can build executable worlds, populate them with agent swarms, and compress thousands of simulated trajectories into human-usable principles. In metamaterials, this revealed how architecture can program failure pathways, improving resilience through load redistribution and controlled collapse.

Odyssey-3 is a foundation world model with an autoregressive diffusion transformer trained to simulate highly diverse scenarios. The model is capable of controlling robots, powering humanoids, driving vehicles, piloting drones, training AIs, and playing video games. It has a learned understanding of physics, dynamics, cause-and-effect, human behaviors, and other concepts that make up reality. The model will enable physical agents that can interface natively with physical and virtual systems.

Meta One is a subscription service that offers enhanced AI and self-expression features across Instagram, Facebook, and WhatsApp. It includes single product plans as well as bundled offerings that combine features for individuals, creators, and businesses, starting at $2.99/month.

The Artificial Intelligence Underwriting Company (AIUC) hopes to bring AI safety to enterprises and companies building AI models and agents. It has built a third-party audit and certification layer for AI agents designed to give enterprises an independent assessment of how safe their AI agents are. Its tests show where companies can trust their agents and where there are concerns, allowing them to be more informed before making decisions. AIUC uses AI agents to run the tests and AI to analyze the data, but humans verify the final audit.

Specialized knowledge search. Parallel thinking. Multiple LLMs. See how Databricks improved Genie's accuracy from 32% to over 90% on an internal benchmarkof real-world data analysis tasks. See how Genie pushes the frontier for data agents →

Relay, an AI workflow tool, shut down due to competition from larger companies integrating similar features.

G5 Labs, an MIT CSAIL spinout, raised $14 million to develop a platform where natural language, structured as intent graphs, serves as the source code.

OpenAI's clear priority is recursive self-improvement, and AI could surpass even Brown's own research intuition within one or two model releases from now.

ALTK-Evolve introduces Consistency Guidelines to address the inconsistency gap, where a model's task success rate drops significantly across repeated runs.