TechLook
Two stories from today capture the AI era's central tension better than any earnings report: the fastest ASIC ever built took nine months because AI helped design the chip that runs AI, and Google just fired its fastest builder for shipping too fast. Meanwhile, a 15-year-old with seven AI agents is out-earning most dev shops, and ByteDance dropped a video model that makes everything else look like it's standing still.
9 Months to a Chip — OpenAI and Broadcom Just Rewrote the ASIC Playbook
OpenAI revealed details on "Jalapeño," its first custom AI chip co-built with Broadcom in nine months — roughly half the typical 18-to-24-month ASIC cycle.
Here's everything you need to know:
- Built with Broadcom in nine months — claimed the fastest ASIC dev cycle in advanced semiconductor history.
- Handles inference for running finished models (not training). Already running GPT-5.3-Codex-Spark at target power.
- OpenAI used its own models to assist in chip design and optimization — AI helped design the hardware it runs on.
- Claims 50% cheaper inference per token than current Nvidia GPUs.
- The company targets 10 GW of compute powered by custom chips by 2029, with Nvidia still anchoring training.
- Confirmed across multiple independent sources.
The most important detail isn't the 50% cost savings or the 10 GW target — it's the nine-month cycle. If a chip newcomer can tape out this fast with AI-assisted design, it resets the timeline expectations for every hardware team in the world. The gap between "idea" and "silicon" just collapsed by half. Founders building any hardware-adjacent product should ask: what happens when chip iteration speed doubles?
Google Fired Its Fastest Engineer for Shipping Too Fast
Justin Poehnelt, a seven-year Google engineer, built a Rust CLI that let AI agents drive Gmail, Calendar, and Drive via structured JSON. It hit #1 on Hacker News. Two days later, Google announced its own official Workspace CLI — then fired Poehnelt.
Here's everything you need to know:
- Poehnelt built the CLI as a side project using internal Google APIs that are available to any engineer.
- It hit #1 on Hacker News, pulled thousands of GitHub stars, and drew tens of thousands of users in days.
- Google legal raised branding concerns about the tool's naming and unauthorized use of Workspace APIs.
- Google announced its own official Workspace CLI for agents two days after Poehnelt's tool went viral.
- OpenClaw creator Peter Steinberger and Vercel CEO Guillermo Rauch publicly recruited him within hours of the firing.
- Poehnelt's CLI did exactly what Google's agent strategy needs: let AI models control Workspace programmatically.
One engineer on a weekend outshipped a funded org with every API and every user — and got fired for it. In the agent era, speed is the defining moat, and Google just punished its fastest person for having it. For every founder reading this: when your best builder ships something customers want before you do, your job is to promote them, not terminate them. The market will remember.
ByteDance Seedance 2.5 — 30 Seconds of Native 4K From One Prompt
ByteDance unveiled Seedance 2.5, capable of producing a single 30-second 4K clip from one prompt with native audio-visual sync and up to 50 multimodal reference inputs.
Here's everything you need to know:
- Generates a single native 30-second 4K clip from one prompt — no stitching, no multi-clip assembly.
- 10-bit color. Audio generated in the same latent space as visuals — sound and motion are synced natively.
- Accepts up to 50 multimodal reference inputs: images, audio, 3D models, style references.
- Includes a 3D white-box preview for roughing out low-fidelity animation before committing to full render.
- Claims 20% better prompt adherence vs. its predecessor.
- Enterprise beta is live now; public launch targeted for early July 2026.
- CapCut's 400 million users sit downstream.
While Google's Veo 3.1 still accepts only three reference images, ByteDance leapfrogs to 50 inputs with native audio-visual sync. The gap between text-to-video and production-ready is closing fast, and it's Beijing, not Mountain View, that's leading the charge. For founders building on generative video: your timeline just compressed. Whatever you thought was six months out might be three.
Gemini 3.5 Flash Ships Computer-Use Built In
Google DeepMind added native computer-use capabilities directly into Gemini 3.5 Flash — the model can now power agents that navigate browsers, apps, and desktops without requiring separate tooling or middleware.
Here's everything you need to know:
- Computer-use is built into the model itself, not a wrapper or separate agent framework.
- The model can navigate browsers, native apps, and desktop interfaces directly.
- No need for third-party computer-use tooling or middleware.
- Follows a wave of similar computer-use features from other frontier labs.
- The move positions Gemini 3.5 Flash as an all-in-one agent-building model.
The practical impact for builders: one API call replaces an integration stack. If you're building agent infrastructure that layers computer-use capabilities on top of LLMs, your value-add just shrunk. The models are absorbing the middleware layer, and they're doing it fast.
Anthropic Accuses Alibaba of the Largest Claude Distillation Attack
Anthropic told the US Senate Banking Committee that Alibaba ran roughly 25,000 fake accounts pumping 28.8 million queries in six weeks to copy Claude's outputs for its own model, Qwen.
Here's everything you need to know:
- 25,000 fake accounts generated 28.8 million queries between April 22 and June 5.
- Anthropic publicly named Alibaba in a Senate Banking Committee hearing.
- Alibaba's stock (BABA) dropped to a 16-month low near $99 on the news.
- Anthropic previously flagged DeepSeek for the same tactic.
- Distillation lets a weak model learn from a strong model's answers for the price of API calls.
- The accusation sets a legal precedent for model theft on a global scale.
Distillation is the new export control. If Alibaba's AI edge is borrowed from Claude, the valuation is borrowed too. The real AI arms race isn't just about building better models — it's about defending the outputs of the ones you already have. For founders building API-based AI products: rate limits and IP detection are suddenly product features, not afterthoughts.
Intercept: $500M From AI Giants to Make Colds and Flu Disappear
Stripe, Anthropic, the OpenAI Foundation, and other donors formed Intercept, a $500 million nonprofit funding respiratory-virus prevention tools and cleaner indoor-air technology.
Here's everything you need to know:
- Donors include Stripe, Anthropic, and the OpenAI Foundation.
- Funds development of shots, sprays, and pills that block dozens of respiratory viruses.
- Also funds air-cleaning technology for offices and schools.
- Estimates people lose 15–25 days a year to routine respiratory infections.
- Quantifies global productivity losses at ~$600 billion annually.
- The $500M covers early-stage research; pharma firms and investors are expected to take over costly final-stage development.
The most interesting detail: AI labs are funding the awkward middle of drug development that pharma won't touch — the gap between basic research and late-stage trials. If Intercept de-risks a universal cold vaccine, the ROI for AI is a healthier workforce that loses fewer days. Pragmatic, focused, and entirely self-interested in the best way.
Fable 5's Thaw Shows Signs of Life
Multiple signals point toward a potential return for Anthropic's Fable 5, which has been offline for nearly a month following a U.S. government order.
Here's everything you need to know:
- A Claude Code update log includes references to Fable usage and removal of "purchased separately" language, hinting at usage changes.
- WIRED reports the Trump administration is "happier" in talks with Anthropic, now dealing with co-founder Tom Brown instead of CEO Dario Amodei.
- Legal-tech firm Legion filed the first lawsuit against the government order, calling it "unlawful."
- Reps. Sam Liccardo, Jay Obernolte, and Ted Lieu gave Commerce until June 26 to explain how the public could get Fable back.
- As we noted on June 18, the suspension was sudden and triggered immediate industry pushback.
Where there is smoke, there is fire. The deadline is tomorrow — Commerce has until June 26 to present a return plan. For builders who relied on Fable's capabilities, the path back is visible but not guaranteed. Bet on it returning, but don't bet the company on the timeline.
Figma Goes Full Stack — Config 2026 Turns a Design Tool Into a Product Shipyard
At Figma Config 2026, the company unveiled features that push the platform far beyond wireframes: timeline-based animation, one-click code export, and AI shader fills.
Here's everything you need to know:
- Figma Motion: timeline-based animation layer for interactive prototypes.
- One-click Code Layers: converts designs to production-ready code.
- AI shader fills: generate textures and materials from text or image prompts.
- Integration with Runway's Aleph 2.0 for on-canvas AI video generation.
- The announcements push Figma further from "design tool" toward "executable product workspace."
- Confirmed across multiple sources covering Config 2026.
Figma is trying to own the entire path from idea to deployed UI. The question for founders: do you want your design team exporting code, or do you want them designing and handing off? Figma's bet is that the line is disappearing, and they want to be on both sides of it.
The Stack Is Splitting — Box CEO on Why the Middle Layer Wins
Box CEO Aaron Levie argues that AI pricing is diverging between expensive frontier models and cheap capable alternatives, making the routing-and-orchestration layer the most valuable part of the stack.
Here's everything you need to know:
- Frontier models are expensive; capable small models are getting cheaper fast.
- The middle layer can route each workflow to the right model, reducing token waste.
- Domain context, customer evaluations, and agent-ready data can tune weaker models to perform well on specific tasks.
- The applied AI layer — not the model itself — captures the most value.
- Levie's argument reframes the "one model to rule them all" narrative.
Every founder trying to build on top of AI should internalize this. The model is a commodity. The routing, the context, the evaluation pipeline, the agent orchestration — that's the moat. Build there.
Advertising's Agentic Reckoning
Advertising agencies are racing to package AI agents as their next product, but brands can use the same tools to pull outsourced marketing work back in-house.
Here's everything you need to know:
- WPP, Dentsu, and Dept are testing agent systems for campaign management and media buying.
- Hyundai's bidding agents reportedly cut video CPMs by 67%.
- Brands can use the same AI tools agencies are deploying, removing the middleman.
- Agencies face an existential question: if your value was execution speed, and AI makes everyone fast, what's left?
- The tension mirrors what happened to design agencies when Canva launched.
If the agency's value was execution efficiency and AI makes everyone equally fast, the agency's value just evaporated. For founders building any B2B service that depends on speed as a differentiator: AI is coming for that advantage next quarter, not next year.
Europe Finally Admits It's Getting Lapped
Eight researchers published "Europe 2031," a manifesto warning that EU AI efforts are 10X to 100X too small and calling for "the most ambitious political agenda in post-war Europe."
Here's everything you need to know:
- The EU controls just 5% of global compute vs. America's 80%.
- Europe has almost no homegrown frontier models.
- By 2031, the US could own the cognitive frontier, China could lead in robotics, and Europe would be forced to pick a side.
- The authors call for massive compute investment, coalition-building, and positioning Europe to win the physical-AI revolution.
- Confirmed by reporting from multiple European tech outlets.
This manifesto isn't news inside Europe — it's a warning for founders everywhere. If you're building AI products that depend on European compute, talent, or regulatory stability, the window is narrowing. The most honest statement in the entire document: "wither away in isolation."
⚡ Quick Hits
- Krea 2: Open-sourced an undistilled image model and a fast 2K generator — 2K images in ~2 seconds on consumer hardware.
- Mistral OCR 4: Layout-aware document understanding that extracts text with structural awareness of tables, columns, and headings.
- Teen agent factory: A 15-year-old runs seven AI agents via OpenClaw, ships ~10 projects a day, and earns $900–$1,500/day on freelance platforms.
- WoW server, 1,800 bots: A World of Warcraft server with zero human players ran entirely on DeepSeek-powered bots — the dead internet theory, playable.
- Aside AI browser: YC-backed startup's agentic browser identifies and cancels unused subscriptions; launch video hit 1M+ views.
- Mercury Command: AI built into Mercury accounts — say "pay this invoice" and Command executes with user approval.
- Reline: AI meeting notes that work without a bot joining the call.
- Tesana: AI game creation platform — describe any game and it builds it.
- Harold: Extract, validate, and import invoice data into ERPs with AI.
- SnapVee Studio: All-in-one AI content workspace for creators, marketers, and educators.
- Salesforce survey: 70% of organizations see AI agent value within 60 days of deployment; adoption has risen to 66%.
- Qualcomm data-center CPU: Landed Meta as first customer — another chip giant pushing into AI infrastructure.
- GPT-5.5 Instant: OpenAI rolled out an update with better conversational understanding and intent handling.
- Mirendil lab: Ex-Anthropic, OpenAI, Google, and xAI researchers launched a lab building AI that conducts autonomous scientific R&D.
- Human Consent Registry: A public registry where individuals can opt out of having their likeness used in AI training data.
- xAI Grok: The Information reports "well over half" of Grok's traffic comes from NSFW content, according to xAI employees.
Techlook — AI & tech signal for founders and builders.