TLDR AI 2026-03-24

See how to move from AI pilots to production with a six-stage roadmap built for strong governance, security, and control.

In this guide, explore how to modernize your data foundation for AI on Amazon Web Services (AWS). Learn how you can advance toward autonomous systems with key capabilities and featured solutions in AWS Marketplace.

Discover how to:

• Assess data and AI readiness and set your next milestones.
• Create a secure data foundationthat supports AI use cases.
• Apply security controls that protect models, data, and outputs.

Download now

OpenAI is currently courting buyout firms to form joint ventures aimed at raising fresh capital and accelerating the adoption of enterprise AI products. It is offering private-equity firms a guaranteed minimum return of 17.5%, significantly higher than typical preferred instruments. It is also offering early access to its newest AI models. TPG and Advent are potential investors.

Claude Code and Claude Cowork can now open files, use browsers, and run dev tools on users' computers. The agent will prioritize connectors to supported services, but it can still execute tasks when connectors aren't available. It will ask for permission before taking action. Claude computer use is initially available to Claude Pro and Claude Max subscribers on macOS.

ChatGPT's new Library feature allows users to store personal files or images on OpenAI's cloud storage. Available to Plus, Pro, and Business subscribers, the feature is rolling out to customers across the world except in the European Economic Area, Switzerland, and the United Kingdom. ChatGPT will automatically save uploaded or created files. Deleting a chat containing a file will not delete the file saved in Library.

The future of AI accelerators will be AI CPUs - traditional CPUs with big caches and systolic arrays, where the caches and systolic arrays take up most of the chip area and power. This book goes over what is required to design a competitive AI chip. The recommendations made in the book will require an excellent software team to realize, but the topics covered should be easily understandable at a hobbyist level.

This post details a problem that would take an expert human about one to three months to solve. The problem was eventually solved by two people using GPT-5.4 Pro. A link to the full transcript of the conversation with GPT-5.4 Pro is available. Other models that have solved the problem include GPT 5.4 (xhigh), Gemini 3.1 Pro, and Claude Opus 4.6 (max).

A professor of physics supervised Claude through a real research calculation from start to finish without ever touching a file to see whether AI could do theoretical physics. This resulted in a technically rigorous, impactful, high-energy theoretical physics paper in two weeks instead of the usual year. While Claude was impressively capable, domain expertise was essential for evaluating its accuracy. AI is still not able to do end-to-end science, but the project proved that it can do frontier science with guidance.

The rate of model releases and the number of new evals required next year could be such that keeping informed will be a challenge without effective AI assistance. These AI-augmented workflows need to be understood now before they become necessary. METR ran a tabletop exercise to learn what workflows emerged, what the bottlenecks are, and how much faster researchers will actually be. This post details the lessons learned from the exercise.

AI‑native development demands a new approach to application security. Signal combines LLM-powered code analysis with 20+ years of human‑vetted security intelligence to autonomously identify, prioritize, and fix vulnerabilities in AI‑generated code. See it in action and request a demo

NVIDIA released a pipeline for quickly fine-tuning embedding models using synthetic data, improving retrieval performance for domain-specific RAG systems.

DynaEdit introduces a training-free method for editing videos with complex motion and interactions using pretrained text-to-video models, addressing issues like misalignment and jitter.

A good prompt can do a lot of the work that fine-tuning does without the cost. Models are getting so good that fine-tuning isn't really required to get good results. The extra work and resources required for fine-tuning aren't worth it anymore for most use cases. It is still worth knowing how to fine-tune, as there are other situations where the skill is useful.

Providing text search indexes to fast models creates a qualitative difference for Agentic workflows. The impact is more pronounced in larger Enterprise repositories because grep latency scales with the size and complexity of the code being worked on. Removing the time spent searching the code base provides meaningful time savings and allows for much more effective iteration, particularly when the Agent is investigating a bug.

OpenAI has hired former top advertising executive at Meta, Dave Dugan, as vice-president of global ad solutions. The company is urgently looking to generate new revenue streams to support its enormous funding requirements. Dugan is known for his close relationships with the world's leading ad companies. Personal relationships are a significant factor that influences where brands ultimately spend their advertising budgets.

OpenAI says that its close ties with Microsoft are a potential risk to its business as Microsoft is responsible for a substantial portion of OpenAI's financing and compute. Other risks include significant capital expenditures, reliance on compute resources, ongoing litigation with xAI, and its unusual structure as a public benefit corporation. OpenAI's operating results will depend on its ability to successfully develop relationships with additional partners. While OpenAI and Microsoft have a tight bond, they're increasingly competing for users in the generative AI market.

The hype is real. Are the results? Learn how online stores are adding AI capabilities to gain a competitive edge, and where they still struggle. Read the Algolia trends report.

This suggests that agentic commerce isn't ready to replace traditional shopping.

A short clip of an iPhone 17 Pro running Qwen3.5-397B-A17B (a 397 billion parameter Mixture-of-Experts model) at 0.6 tokens per second.

DSPy is a framework that promises to solve the biggest challenges in AI engineering, but it's hard to use.

Software companies face two clear paths: accelerate growth with AI products or achieve 40%+ true operating margins.