The AI and Tech Weekly: June 9, 2026

Mouad Belghiti·2026-06-097 min read

The AI and Tech Weekly: June 9, 2026

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

Market News

Apple Rebuilds Siri on Google Gemini

Apple walked onto the WWDC 2026 stage and announced that the new Siri runs on Google Gemini models. That sentence would have been impossible two years ago. Apple builds its own chips, operating systems, and programming languages. Shipping a flagship product powered by a competitor's model is a sharp departure from its default playbook.

The new Siri handles multi-step tasks — researching concert tickets, booking reservations, reading context across apps. Apple also shipped a Core AI Framework giving third-party developers direct API access to on-device models. Hacker News hit 535 points and 480 comments. The practical question the community is debating: does this mean Apple is behind in model development, or is this a deliberate infrastructure choice while Apple builds its own foundation models in the background?

OpenAI Files a Confidential S-1

OpenAI submitted its IPO paperwork to the SEC this week. The filing is confidential, so no revenue numbers are public yet. The HN thread hit 334 points and 263 comments, with the main debate focused on whether the company's current valuation holds once compute costs and revenue are visible. OpenAI previously raised at a $300 billion valuation. The filing also came alongside reporting that the Trump administration is in talks to take an equity stake, adding a political dimension nobody expected.

Xiaomi Ships a 1-Trillion Parameter Model at 1,000 Tokens Per Second

Xiaomi published a model called MiMo-v2.5-Pro-UltraSpeed: a sparse Mixture-of-Experts architecture with 1 trillion total parameters and 42 billion active during inference. It runs on 8xA100 GPUs at over 1,000 tokens per second. The cost is roughly 3x standard MiMo pricing, which is still cheap compared to Anthropic or OpenAI rates. The HN thread hit 546 points and 395 comments. Separately, Cerebras is trialing Kimi K2.6 at 3,000 tokens per second. Speed is now a real competitive axis, not just a benchmark number.

AI Progress Is Slowing — or Is It?

A piece on wheresyoured.at arguing that AI capability growth is plateauing drew 504 points and 525 comments on HN. The argument: the gap between model generations is narrowing, benchmarks are saturating, and companies are now shipping UX improvements rather than capability jumps. A separate LessWrong post made the same case more formally. The community pushed back hard on some points but agreed on one: the field is running out of benchmarks that tell us anything meaningful about real-world performance.

Tips and Tricks from the Community

Run AI agents inside reactive Python notebooks — Marimo-pair gives agents a live, stateful Python environment. The agent can run cells, read outputs, and iterate on failures inside the same session you're working in. This is a cleaner setup for data-heavy tasks than a script runner. Clone the repo and try routing a data cleaning task through it. (github.com/marimo-team/marimo-pair)
Use MCP over skills for any tool integration you want to last — The "I still prefer MCP over skills" thread (420 points, 337 comments) produced a practical conclusion: MCP gives you a model-agnostic protocol with typed tool definitions. Skills tied to specific model APIs diverge as models update. If you're building tooling for a team, anchor it to MCP now. (david.coffee/i-still-prefer-mcp-over-skills)
Build a small personal tool this week — The Ask HN thread "What tools have you made for yourself since AI?" (260 points, 430 comments) shows what developers are actually building: local recipe databases, personal finance dashboards, email triage scripts, terminal shortcuts. The pattern: narrow scope, specific daily friction, under 200 lines of code. Start with a task you do manually at least three times a week. (news.ycombinator.com/item?id=48449187)
Set Kagi URL redirect rules to cut search noise — Kagi published a tips post (161 points) on customizing search results with redirect rules. You can route queries for specific topics through filtered indexes, strip tracking parameters, or skip site landing pages. Ten minutes of setup pays off every day. (blog.kagi.com/tips/redirects)

Open Source This Week

Gitdot — A self-hostable GitHub alternative written in Rust. Familiar PR and repository model, fast, minimal hardware requirements. gitdot.io
marimo-pair — Reactive Python notebooks as live environments for AI agents. Agents get full kernel access and a real feedback loop. github.com/marimo-team/marimo-pair
Performative-UI — A React component library cataloging common UI dark patterns. Useful for auditing your own product's persuasive design choices. vorpus.github.io/performativeUI
Apple Core AI — Apple's new developer framework for on-device model inference. Not open source, but a significant API surface worth bookmarking for iOS developers. developer.apple.com/documentation/coreai

Startup Ideas for Inspiration

Drawn from YC's Requests for Startups and other leading accelerators

AI for Low-Pesticide Agriculture

Global food demand keeps rising while arable land and water availability stay flat. The dominant solution for the past 60 years has been chemical inputs. Pesticides and fertilizers produce yield but degrade soil over time, create regulatory risk, and are facing pushback from both consumers and regulators. YC wants startups combining AI, robotics, and biology to reduce pesticide use without cutting yield.

The most tractable near-term approach is precision application: using computer vision to identify which plants are stressed or which areas of a field have pest pressure, then applying treatments only where needed. This can reduce input costs by 30-60% without changing crop output. The first customers are large commodity farms that already have enough data to train a model.

The AI Operating System for Companies

Most companies run on a patchwork of SaaS tools that do not talk to each other. Data moves between them via Zapier, CSVs, and manual processes. An AI agent that can span all of these tools and act on behalf of a team is more valuable than any single tool improvement, but it requires an orchestration layer that does not exist yet.

YC wants startups building this layer: a system that has write access to your CRM, your email, your project management tool, and your billing software, and can execute multi-step workflows across all of them on behalf of a human. The first version does not need to cover every tool. It needs to cover the 3-5 tools a specific job function lives in every day.

Hardware Supply Chain Visibility

Consumer electronics and automotive manufacturers manage their component supply chains the same way companies managed inventory in 1995: with ERP systems, spreadsheets, and personal relationships with account managers. When a component goes on allocation, the first signal is often a missed delivery.

YC wants startups building real-time supply chain intelligence for hardware companies: lead time tracking across distributors, substitution matching when a part is unavailable, and early warning when a supplier is accumulating risk. The customers exist and they're spending money on this problem already, just poorly.

Inference Chips for Agent Workflows

Current AI accelerators were designed for training large models and serving single-turn queries. Agentic workloads are different: an agent calls a tool, waits for a result, resumes with the result in context, and does this dozens of times per task. The memory access patterns and latency requirements are structurally different from batch inference.

YC is funding chip startups that optimize specifically for this pattern. Lower latency on short context resumptions, efficient state management between tool calls, and cheaper per-token cost at low batch sizes are the key metrics. Any company running agents at production scale is a potential customer, and none of them have hardware designed for what they're actually doing.

※