The AI and Tech Weekly: June 11, 2026

Mouad Belghiti·2026-06-117 min read

The AI and Tech Weekly: June 11, 2026

Top market news, developer tips, open source picks, and startup ideas from the community

Market News

SpaceX Prices the Largest IPO in History

SpaceX set its offering price at $135 per share today, targeting a $1.8 trillion valuation — the biggest IPO ever, topping Saudi Aramco's 2019 debut. Investor demand hit $250 billion for a deal raising $75 billion, making it more than 4x oversubscribed. Retail investors are getting 30% of shares, an unusually large allocation. Trading starts tomorrow.

The context that matters for tech: Google already agreed to pay SpaceX $920 million per month for compute on its Starlink network. That's over $11 billion per year from a single customer. SpaceX is now both a space company and a critical piece of AI infrastructure, and today's IPO prices that bet into the public market.

Anthropic's Fable 5 Is Frustrating Security Researchers

Anthropic released Fable 5 as the public version of Mythos, its cybersecurity-focused model with a 1 million token context window. The reception from security professionals has been hostile. The guardrails trigger on any prompt containing cybersecurity terminology — including code reviews, blog summaries, and documentation. When flagged, the model falls back to Opus 4.8.

IBM X-Force researcher Valentina Palmiotti said it "rejects any request that could be tangentially cyber related." The pattern appears keyword-based rather than intent-based. Separately, Anthropic requires 30-day data retention for both Fable and Mythos, raising concerns from researchers about sensitive data ending up in training pipelines. The Cyber Verification Program can unlock fewer restrictions, but most professionals say the application friction is too high.

Google Drops AI Subscription to $4.99, Starting a US Price War

Google cut the monthly price of Google AI Plus from $7.99 to $4.99 and doubled included storage from 200GB to 400GB. The plan includes video generation via Omni Flash, the creative studio Google Flow, and NotebookLM. This is not a promotional move. It is a direct signal that the pricing dynamics playing out in India, where both Google and OpenAI launched sub-$5 AI plans over the past year, have crossed into the US market.

Both OpenAI and Anthropic have filed confidentially for IPOs. Their ability to command premium valuations will be tested by exactly this kind of downward price pressure. Anthropic has no budget tier and no localized pricing for India. Goodwater Capital's Chi-Hua Chien put it bluntly: "There will be a period of time when these infrastructure companies are valuable. But over time, you will see them get increasingly commoditized." OpenAI released GPT-5.2 this week as well, with early testers citing strong performance on multi-step business workflows and autonomous coding tasks, adding more pressure to the growing model arms race.

GitButler Raises $17M to Rebuild Version Control for the AI Era

GitHub co-founder Scott Chacon closed a $17M Series A led by a16z for GitButler, betting that Git's 2005 design doesn't fit how developers work alongside AI agents. GitButler runs on top of Git and adds parallel branches, unlimited undo, and clean agent integration with Claude, Cursor, and Codex. It doesn't replace Git. It replaces the context-switching friction that happens between writing and committing code.

The broader claim is that as AI agents become co-contributors on codebases, the coordination layer — how changes are organized, attributed, and reviewed — needs a redesign. This is the infrastructure bet, not the coding assistant bet.

Tips and Tricks from the Community

Go HTML-first before reaching for a framework — A developer replaced a failed React app with a plain HTML form for a utility company. The React build was pulled after 3 days of user complaints. The HTML version doubled users overnight. Before adding a frontend framework to your next internal tool, ask whether a plain HTML form with a server POST would work. It usually will. (mohkohn.co.uk/writing/html-first)
Use MCP for structured agent tool access, not skill prompts — A post titled "I still prefer MCP over skills" got 420 points and 337 comments on HN. The argument: skill-based systems are prompt-level hacks that break under complex workflows, while MCP gives agents typed function signatures and composable tool access. If you're building agent tooling, read the MCP spec before designing your tool interface. (david.coffee/i-still-prefer-mcp-over-skills)
Drop your agent into a marimo notebook for data work — Marimo-pair gives AI agents full access to a running Python notebook session. The agent can read program memory, execute code, and build pipelines in the same kernel you're using. Because marimo is reactive, changing one cell updates all dependent cells automatically. Try assigning Claude Code a data cleaning task and watching it work inside your live notebook. (github.com/marimo-team/marimo-pair)
Add an audit log to every agent before it touches production — The "AI agent runs amok in Fedora" story (413 points, 167 HN comments) sparked a clear consensus: agents that touch shared systems need human-readable logs of every file read, every API call, and every decision. Before your next agent goes near a shared codebase or API, wire up a simple append-only log with timestamps and reasons. It takes 30 minutes and will save hours of debugging.

Open Source This Week

marimo-pair — Reactive Python notebooks as shared environments for humans and AI agents. Agents get full kernel access; you get reproducible code. github.com/marimo-team/marimo-pair
PgDog — Rust-based PostgreSQL connection pooler, load balancer, and sharding proxy. No schema changes needed; already serving 2M queries/sec in production. github.com/pgdogdev/pgdog
Apache Burr — Framework for building reliable, stateful AI agents with built-in observability and action tracking. burr.apache.org
πFS — A filesystem where every file is stored as its position in the digits of pi. Technically real, practically useless, and 756 points on HN worth of community chaos. github.com/philipl/pifs

Startup Ideas for Inspiration

Drawn from YC's Summer 2026 Requests for Startups and other leading accelerators. These are gaps they're actively looking to fund.

Counter-Swarm Defense

Cheap drones now cost less than the missiles used to shoot them down. YC wants startups building the full counter-swarm stack: high-capacity interceptors, sensor fusion software that gives operators a single real-time picture, non-kinetic defenses, and tools that attack the autonomy stack of adversarial drones. The framing from YC partner Tyler Bosmeny: the winning company here will look more like Cloudflare than Raytheon. Software-first, hardware-aware, and deployable fast.

AI-Native Service Companies

YC's thesis: the total global spend on services — insurance, accounting, tax, compliance, healthcare admin — is larger than the total global spend on software. Most of those services are outsourced and process-driven, which makes them structurally easy to replace with AI. YC isn't looking for AI tools to help human teams do these tasks faster. It's looking for companies that replace the outsourced service entirely with a smaller team and AI handling the work. Accounting firms and insurance brokerages are the first targets.

Software Built for Agents, Not Humans

Today's apps were designed for humans: menus, forms, dashboards. Agents interact with them like humans pretending to be robots. YC wants startups building software where the primary user is an AI agent — structured APIs, state machines, and interfaces designed around what agents actually need to do work reliably. One sub-idea from the RFS: dynamic software interfaces that change based on the user's role, context, and current task, rather than showing every feature to everyone all the time.

AI-Personalized Medicine

YC frames this as one of the highest-leverage problems AI can address: treatments optimized for a patient's specific genome, microbiome, lifestyle, and prior response history, rather than population-level averages. The starting point most tractable for a small team is drug response prediction — using a patient's genetic data to predict which medication will work before prescribing it. YC notes that genomic sequencing now costs under $200, which removes the data collection barrier that blocked this space for a decade.

※