TLDR AI 2026-03-12

25 years after the Agile Manifesto, Thoughtworks and Martin Fowler gathered a small group of technologists in the mountains of Utah. They confronted the new inflection point: AI-native software development, and the hard questions it raises:

• Where does the rigor go? If AI writes the code, where does the engineering discipline live on?
• The new "middle loop": The as-yet unnamed practice of supervisory engineering, directing AI agents, and evaluating their output.
• The decoupling of productivity and developer experience: Engineers can now achieve more, and like their jobs less...

Their answers aren't definite, but they're definitely interesting.

Read key findings and strategic insights from the retreat

Replit launched Agent 4, introducing an infinite design canvas and the ability to run parallel AI agents to build backends and frontends simultaneously. The update shifts from pure code generation to a collaborative product suite capable of shipping web apps, databases, and slide decks within a single integrated environment.

Meta has announced four new chips as part of an effort to diversify its hardware sources, reduce reliance on outside chipmakers, and bring down costs. The MTIA 300 is already in production for content ranking and recommendations training. The MTIA 400 has completed lab testing and is moving toward deployment. The MTIA 450 and MTIA 500 chips are scheduled for mass deployment in 2027.

Cursor is in talks with investors for a funding round that would value it at $50 billion, nearly double its valuation last fall. The discussions are preliminary and may not result in a funding deal. One of the fastest-growing startups of all time, Cursor is a central player in tech's vibe coding era. It has attracted financing from top names in venture capital and the tech industry. The company's last financing round, announced in November, boosted its valuation to $29.3 billion.

Personal Computer is a version of Perplexity Computer that handles things on a nearby computer. It handles requests from users and directs multiple AIs to complete tasks. It is effectively a project manager that delegates tasks to other AIs before combining the results and presenting them. Perplexity has specifically named the Mac Mini as a platform that Personal Computer can run on. There is currently a waitlist for access.

Non-technical users can build functional agents by deploying API connectors and project-based instructions to automate high-cognitive tasks. This modular framework utilizes parallel subagents to execute complex workflows, shifting the user role from conversational prompter to systems architect.

OpenAI outlined how prompt injection attacks against agents increasingly resembled social engineering, and argued that defenses should focus not just on detecting malicious inputs but on limiting the impact of successful manipulation.

To accelerate the adoption of AI robotics, Intrinsic, Open Robotics, Google DeepMind, and NVIDIA are running a global competition for developers / roboticists: train an AI model to handle dexterous cable insertion in sim using Gazebo, Isaac Sim, or MuJoCo. Finalists will deploy in a physical workcell at Intrinsic HQ, with top teams splitting a $180K prize pool. Sign-ups close April 17th, so join the AI for Industry Challenge today!

Google's controlled experiments showed that enabling reasoning improved single-hop factual recall in LLMs by acting as a computational buffer and by generating related facts that help retrieve answers, though hallucinated intermediate facts increased final-answer errors.

Autoresearch is a tool that ingests a training script and a metric and then edits code, runs experiments, checks if the metric improved, keeps or discards, and repeats. It allows researchers to optimize models overnight, even without any machine learning experience. Andrej Karpathy used it to squeeze 11% more speed out of his GPT-2 training. Shopify's CEO used the tool to train a 0.8B model overnight that outscored his previous 1.6B model.

NVIDIA just released Nemotron 3 Super, a 120B open hybrid MoE model (12B active) designed for agentic and multi-agent AI systems.

Anthropic has launched the Anthropic Institute to tackle AI's societal challenges. Guided by co-founder Jack Clark and an interdisciplinary team, key hires include Matt Botvinick for AI law, Anton Korinek for economic impacts, and Zoë Hitzig for economic connections. Anthropic has also expanded its Public Policy team to shape global AI governance, led by Sarah Heck with a new office in DC.

Large language models are probably not currently being used in mission-critical ways, but this will change within 20 years, when 99% of the workforce will be AIs. Our future civilization will run on AI labor. We need to think through some extremely important questions about who this future workforce will be accountable and aligned to, and who gets to determine that. The only way to preserve a free society is to make laws and norms saying it is unacceptable for the government to use AI to enforce mass surveillance, censorship, and control. Individual acts of corporate courage will not solve the problem.

Replay is Temporal's practical conference for devs building real systems. Join for hands-on workshops and leave with prod-ready AI agents. Save with code TLDR75

Nvidia and Nebius will work together to design and deploy large AI data centers.

AI code generation relieves the execution bottleneck that previously functioned as a temporary barrier to entry for software companies.

Cursor added more than 30 marketplace plugins from partners, enabling the editor to read, write, and take actions across a wider range of developer tools and services.

The strategic move is aimed at expanding Anthropic's enterprise footprint and potentially challenging Microsoft's newly launched Copilot Cowork.