TLDR AI 2025-09-22

Ditch the IT bottleneck: AI agents are ready to deploy across departments. Think "collaborative playground" with invisible guardrails. Airia lets every department build out their own use cases with security and governance baked-in:

• Use templates and no-code to rapidly prototype, deploy, and manage AI agents that transform workflows across your organization
• Bring your own data and LLM with built-in connectors to dozens ofenterprise applications and LLM tools
• Add security and governance with controls, audit trails, and responsible guardrails to protect sensitive data
• Optimize performance and costs through smart routing and centralized lifecycle management

Plans start at just $49/month. Get a demo

Oracle is in talks with Meta for a multi-year cloud computing deal worth about $20 billion that would see Oracle provide Meta with computing capacity for training and deploying AI models. Oracle unveiled four multi-billion-dollar contracts last week. It expects to sign up several additional multi-billion-dollar customers in the next few months. Booked revenue at its OCI business will exceed half a trillion dollars.

xAI was valued at $150 billion a few weeks ago and $80 billion when it merged with X in March, though Musk later called the report “fake news.”

Grok 4 Fast is comparable to Grok 4 but uses 40% fewer thinking tokens and features the first 2 million token context window.

Chinese models are dominating the open-weight LLM space. Some of the models compete with or even outperform leading proprietary US models on common benchmarks. The Chinese government is promoting open models as part of its AI strategy. Flooding the market with free high-quality AI models helps China build its AI capabilities while making it harder for the US to sustain its lead.

A key difference between pre-training and reinforcement learning (RL) is their information efficiency. Pre-training via next-token-prediction provides models with a token worth of information to learn from for every token the model produces during training, while RL requires a long chain of thousands and even millions of tokens before revealing to the model a single bit of information. RL provides models with much less information to learn from per GPU-hour. This has serious implications for how much the RL paradigm can achieve.

GPT-oss' weights are openly available, but the sources of training data are not clearly described in the model card. This article takes a look at the model's parameters to find out what the model was trained on. There is strong evidence that the models in the GPT-5 and GPT-oss family were trained on phrases from adult websites. There is weak evidence that part of the GPT training corpus was scraped off GitHub.

Between August and September, three overlapping infrastructure bugs caused Claude to produce degraded responses, including random foreign characters, misrouted requests to wrong server configurations, and an XLA compiler issue that dropped high-probability tokens.

Voice assistants used to be glorified answering machines. Today, CallRail's AI Voice Assistant fields calls more fluently than most humans, capturing details that help businesses qualify leads while giving consumers answers faster. Try it yourself in a live call or 14 day free trial, or start by reading the blog.

Large language models compress massive amounts of training data into their parameters. This process can be reversed: structured data sets can be systematically extracted from trained models. This post discusses a decompression technique that can generate substantial training data sets from open-source models. The key is treating inference as a knowledge extraction tool rather than just a generation mechanism.

OpenAI and Apollo Research evaluated frontier models for hidden misalignment behaviors and developed early-stage methods to reduce deceptive tendencies.

Nvidia paid $900 million to hire Enfabrica's CEO and other employees and to license the company's technology. Enfabrica's Technology can connect more than 100,000 GPUs together. It could help Nvidia offer integrated systems around its chips so clusters can effectively serve as a single computer. Nvidia had previously invested in Enfabrica as part of a $125 million Series B round in 2023.

Emad Mostaque, former hedge fund manager and creator of Stable Diffusion, recently appeared in an interview where he discussed how Capitalism is about to become obsolete because of AI. He also talked about how AI will fundamentally transform the economy, replacing workers and making traditional economic structures obsolete - within 1,000 days. He advises people to build network capital and master AI now because human cognitive labor will become economically worthless soon.

Turn your ideas into AI-powered apps - starting from a Figma, a prompt, or a vague idea of what you want to achieve. Get 1M free tokens to power your first build.

Perplexity's AI browser, Comet, now includes built-in support for 1Password credential management.

Scaleway now powers serverless inference on Hugging Face, offering fast, European-hosted APIs with support for structured outputs and multimodal features.

Stability AI has integrated its image editing services into Amazon Bedrock, enabling enterprise users to access advanced features like inpainting, style transfer, and object recoloring via API.