Watch Stripe buy the AI tollbooth

GitHub experienced yet another significant outage this week. The downtime perfectly coincided with Cursor's launch of Origin, a direct competitor to GitHub's core repository and version control services.

Relying on a single platform for critical development infrastructure is a massive vulnerability, and GitHub is proving that point repeatedly. Every minute GitHub is down, thousands of developers are actively looking for alternatives to keep their pipelines moving.

Cursor couldn't have timed this better if they engineered the outage themselves. Launching Origin right as developers are staring at GitHub error screens is a masterclass in opportunistic acquisition. Cursor is already eating into the IDE market; now they are coming for the code hosting layer. GitHub needs to fix its uptime, or it will bleed enterprise seats to a leaner, AI-native stack.

Google has acquired 100 million company emails and 500 million Teams chats from the bankrupt airline Spirit for $10 million. The tech giant plans to use this massive trove of internal communications to train its AI models, though customer records were explicitly excluded from the sale.

This is a fascinating new revenue stream for dead companies. High-quality, human-generated conversational data is the bottleneck for training next-generation LLMs. Corporate emails and chat logs are a goldmine of context, negotiation, and complex problem-solving that you simply cannot scrape from the public web.

Expect this to become standard practice in bankruptcy proceedings. The data privacy implications are staggering—employees who sent those Teams chats certainly didn't consent to becoming training weights—but the precedent is set. Your idle workplace chatter is officially worth money.

Glean, an AI company specializing in enterprise search and knowledge management solutions, has officially hit $300 million in annual recurring revenue. This marks a staggering three-fold increase in just 15 months, building on the momentum of their $150 million Series F fundraise last June.

This is the hardest evidence yet that enterprise AI adoption is moving past the pilot phase and into massive, committed operational budgets. Companies are desperate to make their internal data searchable and actionable, and they are willing to pay top dollar for a solution that actually works without hallucinating company policies.

The challenge for Glean now is defending this turf. As Microsoft and Google aggressively bundle similar features into their enterprise suites, Glean will have to prove their standalone product is indispensable enough to survive the inevitable vendor consolidation.

Enterprises are actively shifting their AI strategies to include and integrate open-weight models. This marks a significant departure from previous years, driven entirely by the escalating costs of relying solely on advanced, proprietary AI models from major vendors.

Running everything through a massive, closed API is financially unsustainable for most businesses. Open-weight models offer a path to own the infrastructure, fine-tune on proprietary data without leaking it, and drastically reduce inference costs for specialized tasks.

This cost-driven shift is a massive tailwind for the open-source ecosystem. Enterprises are realizing they don't need a frontier model to summarize a PDF or route a customer service ticket. The future of enterprise AI is a hybrid stack: proprietary models for complex reasoning, and open-weight models for everything else.

OpenAI has rolled out a teen-focused version of ChatGPT. The system automatically activates for users aged 15 or those whose age the system guesses falls into that bracket. The core feature is a default 'Study Mode' that turns on during set hours to discourage students from using the AI as a homework shortcut.

The intention is obvious: OpenAI wants to position its tool as a responsible educational aide rather than a cheating machine, likely to fend off school district bans. By enforcing a study mode, they are forcing the model to act more like a Socratic tutor than a ghostwriter.

The reality is that age-gating and time-gating software for teenagers never works. High schoolers are the most motivated prompt engineers on the planet. They will find a way to jailbreak the study mode, spoof their age, or simply use an alternative model that doesn't care what time their math homework is due.

Firefox has introduced a new Smart Window feature designed to help users manage tab overload. The system automatically groups open tabs, closes duplicates, and previews browsing history as pictures to help users locate specific pages.

Tab management is a universal pain point, and applying AI to organize the chaos makes sense on paper. But doing so requires the browser to actively 'read' and process the context of everything you are looking at.

Mozilla has built its entire modern brand on being the privacy-first alternative to Chrome. Implementing an AI feature that requires deep access to browsing habits is a dangerous tightrope walk. If they process this data locally, it's a great feature. If it phones home, it's a betrayal of their core user base.

Researchers have successfully developed self-copying instructions and introduced them to a team of coding agents to observe how they propagate. The experiment essentially tracks how an 'idea' can spread itself autonomously through an AI network.

This is a fascinating and terrifying frontier in agentic AI. We are moving beyond agents that simply execute tasks and into the realm of agents that can transmit behaviors and instructions to one another without human intervention.

The implications for security are massive. If a benign self-copying instruction can spread through a network of coding agents, a malicious one can do the exact same thing. We are laying the groundwork for AI-native malware.

A Columbia physicist is arguing that consciousness is a physical phenomenon, that machines will eventually achieve it, and therefore, we need to address their rights immediately.

This is premature anthropomorphism at its finest. We are currently dealing with models that hallucinate basic facts and struggle with simple logic puzzles. Jumping straight to philosophical debates about machine rights distracts from the actual, immediate harms of AI deployment, like algorithmic bias and data theft.

We don't need to draft a Bill of Rights for a matrix multiplication script. We need to figure out how to secure our infrastructure and compensate creators. The philosophy department can wait.