TLDR AI 2025-08-28

Codex has rolled out updates, including a new extension for editors like Cursor and VSCode, an enhanced CLI for local use, and seamless management of local and cloud tasks. GitHub code reviews can now be driven by Codex, all integrated into existing ChatGPT plans and backed by GPT-5.

The Gen AI ecosystem is starting to stabilize. There were only 11 new names on Andreessen Horowitz's most recent Top 100 Gen AI Consumer Apps list, compared to 17 newcomers in the March ranking. There were significantly more mobile newcomers as app stores have cracked down on ChatGPT copycats, opening up room for more original apps. ChatGPT is still the leading general assistant, but Google, Grok, and Meta are closing the gap. More on the top Gen AI consumer apps is available in the article.

Nvidia, the world's most valuable company, reported $46.7 billion in revenue this quarter, a 56% increase compared to the same period last year. AI-dominated data center business largely fuelled that growth.

The winners in the AI era will be the tools that adapt to users. Large language models shift the focus from designing the solution to defining the problem. They can handle the 'how' aspects of the problem, so you just need to describe what you want in plain language. The future belongs to software that bends without breaking.

Small language models are models ranging from 270M to 32B parameters that run efficiently on CPUs or modest GPUs. They offer immense potential: privacy through local deployment, predictable costs, and full control thanks to open weights. They also present unique challenges that require a shift in how agent architectures are designed. This article reviews lessons learned from hands-on experimentation, debugging, and optimizing of inference pipelines for small language models.

Omni is an internal platform at Cloudflare built for running and managing AI models on its edge nodes. It can spawn and manage multiple models on a single machine and GPU using lightweight isolation. Omni makes it easy and efficient to run many small and/or low-volume models. This article looks at how Cloudflare uses Omni to run more models on every node in its network, improving model availability, minimizing latency, and reducing power consumed by idle GPUs.

Stax is an experimental developer tool designed to streamline the LLM evaluation lifecycle. Evals give developers clear metrics to help understand what's actually better, so developers don't have to waste hours 'vibe testing' every time they try a new model or tweak a prompt. Stax empowers developers to rigorously test their AI stacks to help make data-driven decisions. It allows developers to easily define their own criteria and build a custom autorater.

A new open-source platform is providing dozens of reinforcement learning environments to support “open AGI” and inviting contributors to add more.

PromptLock is a newly discovered proof-of-concept ransomware that uses OpenAI's gpt-oss-20b model to generate code to exfiltrate data, encrypt it, and potentially destroy it. It uses the Ollama API to generate malicious Lua scripts on the fly, which it then executes. The ransomware runs locally, which allows it to avoid detection more easily as it doesn't use any online resources.

OpenAI and Anthropic allowed each other internal API access for joint AI safety testing, aiming to uncover blind spots in their model evaluations.

Slow, rigid databases choke RAG, semantic search, and GenAI. OpenSearch 3.0 delivers open, fast, scalable vector search to power your AI stack—enterprise-ready and license-free.

Explore OpenSearch

Ironwood is the first Google TPU explicitly designed for large-scale AI inference.

Researchers used ChatGPT outages as a natural experiment to measure AI's impact on financial markets, finding significant drops in trading volume and reduced price informativeness when the service went down.

The primary goal of a world model is to simulate all actionable possibilities of the real world for purposeful reasoning and acting.

This developer found that they stopped listening to music when they replaced their old coding process with prompt-crafting.

All AIs working in adversarial environments are vulnerable to prompt injections, and the people developing these technologies are pretending this isn't a problem.