Shobit

Mozilla applied Claude Mythos Preview to Firefox and found 271 vulnerabilities. All of them are fixed in Firefox 150, which shipped this week. For context: when Mozilla used Claude Opus 4.6 on Firefox 148, it found 22 bugs. Mythos found 271 in the next version. Mozilla's CTO described the experience as giving the team "vertigo." For a hardened target, one zero-day would have been a red alert in 2025. The more important line from Mozilla: they found no category of vulnerability that humans can find that Mythos cannot. Security has always favored attackers because finding one bug is easier than defending against all of them. AI just changed that equation - for both sides.

Yesterday, Anthropic silently removed Claude Code from the $20/month Pro plan on their pricing page - no announcement, no email, no blog post. Reddit and Hacker News caught it. Within hours the change was reversed. Anthropic then confirmed it was a test affecting 2% of new signups - but the pricing page and docs had been updated publicly as if the change was real, affecting the other 98% of users who saw it. The lesson is not about this specific pricing decision. It is about how to communicate pricing changes. You do not update your public pricing page for a test you haven't announced. Users price their businesses around your numbers. Trust burns fast and rebuilds slowly.

Meta installed software on all US-based employees' computers to capture mouse movements, clicks, and keystrokes. The stated purpose: train AI agents that struggle with basic computer tasks like selecting from dropdown menus and using keyboard shortcuts. The program is called "Model Capability Initiative." The broader initiative, previously called AI for Work, has been rebranded: "Agent Transformation Accelerator." Meta is simultaneously preparing to lay off 10–20% of its workforce. Employees are now generating the training data for the agents that will replace them. And the company named the program so clearly that no one can claim they didn't know what it was for.

Two years ago, AI image generators invented new Mexican dishes. "Enchuita." "Churiros." "Margartas." ChatGPT Images 2.0 just generated a restaurant menu that could go straight to print. The model can now reason before it draws - searching the web, verifying its own outputs, generating up to 8 coherent images from one prompt. It hit the top spot on the Image Arena leaderboard by a 242-point margin within 12 hours of launch. The last hard thing AI image generation had was text. That is no longer hard.

PostHog's session replay started as a hackathon project. So did their data warehouse. Two of their most significant products were built in 24-hour sprints at company offsites, not in roadmap planning meetings. The pattern holds across a lot of great products. Constraints force real decisions. Deadlines eliminate committee thinking. Working code beats slides. The most productive 24 hours a team can spend together is often a hackathon. Not because hackathons are magic. Because they strip away everything except the question: can we actually build this thing?

Sub-one-year contracts for new B2B software deals: 4% in 2023. 13% today. Three-year deals: down from 28% to 23% over the same period. This is not buyers losing confidence in your product. It is buyers being rational in a market where the category leader in 12 months might be completely different from the one today. ICONIQ surveyed 150+ GTM executives for this data. The pattern holds across revenue bands. Large deals. Small deals. Everywhere. The vendors winning are not fighting it. They are making the first 90 days so embedded in the customer's workflow that the renewal is not a negotiation.

SaaS gross margins run 70–85%. AI companies are running 40–60%. Every time a SaaS company ships an AI feature, they are absorbing inference costs into a margin structure built around near-zero marginal cost. Inference alone averages 23% of revenue at scaling-stage AI companies. That is not a temporary problem. It is a structural one. The companies winning this shift are not the ones bolting AI onto SaaS. They are the ones pricing on outcomes from day one - selling into labor budgets instead of software budgets, and charging a price that reflects what they actually cost to deliver. SaaS valued at $100/seat gets repriced at $10,000/outcome saved. The math works. The transition is brutal.

@gregisenberg A whatsapp group?

It feels pretty obvious at this point that someone’s going to make billions building a social app that’s just for friends, no AI slop, no brainrot, calm design, chronological feed and no concept of followers

@Schmett Would love to connect and chat. I'm Building Farcast Always-on AI Growth Agent for Founders - finds your ICP, runs your content engine on 25+ social channels, and sends warm leads to your inbox daily. getfarcast.com

got an idea? want $1M? apply to speedrun. I can refer you with a 10k scout check Amazing team and an incredible opportunity!

@SpaceX @cursor_ai 60 billion exit for a 4 year old company, real power of AI

SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI. The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will allow us to build the world’s most useful models. Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.

"Designing in code" is like telling an industrial designer to start in the fabrication shop. You waste enormous resources prototyping something nobody has thought through yet. The problem AI tools made worse: now everyone can generate a screen in 30 seconds. So teams skip the harder question of what the screen should do and why it should exist at all. Content is where design actually starts. What does it say? What does it promise? What does clicking it commit the user to? Those questions cannot be answered in a component library. The people building the fastest right now are not the ones who go straight to code. They are the ones who know exactly what they are building before they touch a tool.

An AI coding agent asked to brainstorm new features returned: a notification center, an activity feed, a dashboard, an onboarding wizard. All reasonable. None of them fit the product. The agent had read every file in the codebase. It had no idea the product was built around a conversation model, not a dashboards-and-wizards flow. This is the gap nobody is talking about: the agent sees the architecture. It cannot see the year of design decisions, rejected interaction patterns, and "we never do X" principles that live in your team's head. Context engineering - writing down the product's character, not just its components - is becoming the most load-bearing skill in AI-assisted product work.

GitHub Copilot's weekly operating costs doubled since January. Microsoft is now pausing new sign-ups and moving toward token-based billing. The reason buried in GitHub's own announcement: "it's now common for a handful of requests to incur costs that exceed the plan price." Developers discovered agentic workflows. Long-running, parallelized sessions started consuming compute at a rate the $10/month plan was never designed to support. This is the flat-rate SaaS model meeting the token economy head-on. The all-you-can-eat AI buffet had a price. Microsoft is now asking developers to pay it. Every AI product with flat-rate pricing and usage-based costs is watching this closely.

@Surendar__05 You just predicted their next release

Genuine Question: Why doesn’t Claude have an image or video generation model yet?

Google just added parallel subagents to Gemini CLI. One agent updates your docs, another writes unit tests, a third scans for deprecated libraries - all at the same time. Each subagent runs in an isolated context window. So when one agent goes deep on a 200K-line codebase, the main agent's context stays clean. Google calls it preventing "context rot." That is the right name for exactly the right problem. Every senior engineer knows the feeling: you give an AI 40 files and by message 15 it has forgotten what it read on message 3. Subagents are structural memory, not just a feature.

Jeff Bezos launched Project Prometheus in November 2025 with $6.2 billion. He is now closing a $10 billion second round at a $38 billion valuation backed by JPMorgan and BlackRock. Total raised in under 6 months: $16 billion. The thesis: train AI on real-world experimental data, robotics interactions, and engineering workflows - not just text. Physical AI for manufacturing, aerospace, drug discovery, and logistics. Bezos is also reportedly seeking up to $100 billion to acquire industrial businesses and feed their operational data directly into Prometheus's models. He is not building a chatbot. He is building a data flywheel with a $16 trillion manufacturing market as the target.

Alibaba just shipped its most capable model yet and made it closed-source. Qwen built its entire distribution strategy on free, open models. Qwen went from 1.2% of global open-model usage in late 2024 to roughly 30% by end of 2025 - almost entirely on the back of free access. Now the flagship is proprietary. Open at the bottom. Paid at the top. The same playbook Meta used to get millions of developers to build on Llama is now the platform Alibaba is quietly monetising. Free models build the moat. Closed models collect the rent.

Moonshot AI's Kimi K2.6 just ran a 13-hour autonomous coding session, modified 4,000+ lines of code in an 8-year-old matching engine, and extracted a 185% throughput gain. No human in the loop. No check-ins. Just a task description and a result. It can also coordinate 300 parallel sub-agents on a single problem. The benchmark that matters: 80.2% on SWE-Bench Verified, matching Claude Opus 4.6, with open weights you can download. A model you can run yourself just matched a frontier closed model on real software engineering tasks.

Most solo SaaS founders spend 80% of their time perfecting a product that only 10 people will ever see. The hard part isn't building,it's figuring out who actually gives a damn and how to get in front of them.

@claudeai @grok is this RIP data analysts?

In Cowork, Claude can now build live artifacts: dashboards and trackers connected to your apps and files. Open one any time and it refreshes with current data.