TLDR AI 2026-03-23

OpenAI has set its sights on building a fully automated agent-based AI researcher that will be able to tackle large, complex problems by itself. The company says this research goal will be its North Star for the next few years. It plans to build an autonomous AI intern by September. This intern will be a precursor to a fully automated multi-agent research system to debut in 2028.

An X user reported that Cursor's new Composer 2 model appears to be based on Moonshot AI's Kimi 2.5. Cursor later confirmed that the model started from an open base model and was further trained with reinforcement learning.

OpenAI plans to expand its workforce from 4,000 to 8,000 by the end of the year. The new hires will be across several departments, including product development, engineering, research, and sales. The hiring spree will include specialists for technical ambassadorship, employees who will help businesses better utilize their AI tools. The company is currently in advanced talks with private equity firms to deploy its tools across a firm's portfolio of companies.

Elon Musk announced Terafab, a $20 billion joint venture by Tesla, SpaceX, and xAI to build the world's largest chip manufacturing facility, aiming to generate a terawatt of computing power annually. The facility in Austin, Texas, will produce chips for terrestrial uses, like self-driving cars, and space use, supporting projects such as the proposed orbital data center. Despite Musk's ambitious claims, he has underdelivered on past ventures like Hyperloop and fully autonomous driving.

The tools for building software factories exist right now. Stripe's Minions prove the model works at scale. The role of the software engineer is shifting toward that of a factory biller. Developers can now build self-improving factories that use agents to read user feedback, A/B tests, and production data, then populate the backlog without human curation.

MiniMax M2.7 offers 90% of the quality of Claude Opus 4.6 at just 7% of the cost, excelling in bug detection but less thorough in fixes. It performs better in floating-point calculations and matches Claude in vulnerability detection, though Claude provides more comprehensive solutions. With rates of $0.30/$1.20 per million tokens versus Claude's $5/$25, MiniMax is a cost-effective tool that bridges the gap between open-weight and frontier models.

Token spend doesn't matter if tasks aren't being completed. The total tokens consumed per task varies across models. A company that charges double for tokens could still be the cheaper option if its models return the right answer in fewer turns. Instead of using token spend as a KPI, companies should measure cost per successfully completed task.

Unwrap aggregates all of your customer feedback into a single view, then uses AI + NLP to surface the most actionable insights and deliver them straight to your inbox. Query your feedback in natural language to get a clear view of customer sentiment. Trusted by Perplexity, Oura,Stripe, and Southwest Airlines. Connect with Unwrap to get a free trial, exclusive to TLDR AI subscribers.

This repository contains development skills for AI coding agents. They can be plugged into any AI coding tool for structured, production-quality guidance for frontend, full-stack, Android, iOS, and shader development. Installation instructions for Claude Code, Cursor, Codex, and OpenCode are available.

Flash-MoE is an inference engine that can run Qwen3.5-397B-A17B (a 397 billion parameter Mixture-of-Experts model) on a MacBook Pro with 48GB RAM at 4.4+ tokens/second. It streams the entire 209 GB model from SSD through a custom Metal compute pipeline without Python or frameworks. Details about the architecture are available in the repository.

OpenHarness provides the building blocks for building very capable general-purpose agents in code. These agents can delegate work to other agents that run autonomously without prompting for permission. OpenHarness supports the AGENTS.md spec and connection to Model Context Protocol servers. It integrates with AI SDK 5's data stream protocol, so developers can use it to stream agent sessions directly to useChat-based React UIs.

Nemotron-Cascade 2 is a 30B MoE model with 3B active parameters that achieved Olympiad-level performance while maintaining high efficiency through Cascade RL and multi-domain distillation.

OpenAI CEO Sam Altman went to extreme lengths to secure compute capacity in 2025. The company has signed billions of dollars of infrastructure deals. As the company gears up for a potential IPO later in the year, it is starting to temper expectations and outline a more measured strategy. The company is starting to realize that the market doesn't necessarily appreciate its past approach to growth and spending.

Tencent launched ClawBot to integrate its WeChat platform with the OpenClaw AI agent, advancing in China's AI agent market. ClawBot aims to enhance AI services on WeChat, providing both consumer and business applications and potentially boosting user engagement. The initiative could impact Tencent's revenue streams via increased usage, payments, advertising, and subscriptions, though it faces challenges from competitors like Baidu and Alibaba.

'Tokenmaxxing' is a new status game where AI-obsessed workers spend tokens to prove how productive they are. Some companies have employees compete on internal leaderboards that show how many tokens each worker consumes. There are developers spending thousands of dollars a month trying to automate as much of their work as possible. A single full-time agent can spend 700 million tokens a week. The leaderboards don't show whether these tokenmaxxers are actually producing anything good, or whether it is all productivity theater.

Machine-to-machine auth between services no longer needs a round-trip to validate. Clerk's JWT M2M tokens verify locally — significantly lower latency, zero per-verification cost. Built for services calling each other at scale. See how it works

Open-source AI models are rapidly closing the performance gap with frontier models, potentially making local AI processing more feasible and attractive.

Claude will run scheduled tasks via cloud infrastructure, so users don't need to keep Claude Code running on their local machines.

Perplexity is adding a Market Research section powered by its agentic multi-model Perplexity Computer system.

Companies will increasingly rely on AI agents, necessitating world models to manage them efficiently.

AI startups made up 41% of venture dollars on Carta last year, with Anthropic, OpenAI, and xAI raising substantial funding.