Anthropic Signs Google and Broadcom TPU Capacity Deal
Today's top 23 insights for PM Builders, ranked by relevance from X, Blogs, LinkedIn, and YouTube.
Anthropic Signs Google and Broadcom TPU Capacity Deal
#1 š
Anthropic signed deals with Google and Broadcom to secure multiple gigawatts of next-generation TPU capacityācoming online in 2027āto train and serve its frontier Claude models.
Also covered by: @Lenny Rachitsky
#2 š Simon Willison
Google AI Edge Gallery - Google's official iPhone app for running Gemma 4 models (E2B, E4B and some Gemma 3 family) works very well locally, with the E2B model a 2.54GB download; it supports image Q&A, short audio transcription and an interactive "skills" demo but conversations are ephemeral and the app lacks permanent logs.
#3 š
Philipp Schmid demos Gemma 4 E2B on an iPhone 17 Pro Max via the AI Edge Gallery, spotlighting its skills-driven Wikipedia querying feature and sharing the app link.
#4 š
OpenAI launched the Safety Fellowship, awarding grants, expert mentorship, and community support to independent researchers focused on AI safety and alignment, aiming to cultivate the next generation of talent.
#5 in
Colin Matthews highlights Anthropicās new programmatic tool calling mode, where the model emits a Python script to batch tool calls and only the final output enters the context window.
#6 š
Harrison Chase highlights LangChainās new community middleware page, showcasing agent middleware as a powerful way to tailor agent harnesses to specific use cases. Heās inviting developers to share what theyāre building with these middleware integrations.
#7 in
Abdul Wahab built and shipped a developer platform solo in under two hours using Claude Code (with 10 parallel PRD-editing agents) and Nimbalyst for live UI renderingāno designers, engineers or tickets required.
#8 š
clem š¤ is open-sourcing their agent traces from Hermes, OpenCode, and Claude via Traces.com to kickstart a crowdsourced dataset for open-source agent models, and urges other builders to share theirs too.
#9 ā¶ļø
Claude Code + 15 repos: how a non-engineer answers every customer question | Al Chen
How I AI Podcast
Al Chen uses Claude Codeās VS Code integration to query a local workspace of 15 Galileo microservice repos and Confluence pagesākept current via a 16-line āpull allā scriptāto produce tailored Kubernetes deployment guides for each enterprise customer.
- The VS Code workspace contains 15 separate Galileo repositories (UI, API, auth, repo, etc.), each directory mapping to one backend service in the platformās architecture.
- The āpull allā script is a 16-line shell script that runs git pull origin main on every directory under the root Galileo folder, ensuring all repos are updated daily.
- The custom Cloud Code command āDPLā first retrieves customer-specific āquirksā from Confluence (secrets stores, namespaces, sidecar settings) then traverses local code to output a step-by-step Kubernetes deployment process (e.g., deploying the āwizardā image with Google Secrets Manager and no CRDs).
#10 š
Santiago emphasizes that context is key for reliable agent performance and that Internet access can make them 100Ć more powerful. He warns against using Google wrappers that return HTML blobs, urging developers to feed agents structured data to avoid token waste.
#11 š Doug Turnbull
Is grep all you need for RAG? - Doug argues that with enough engineering effort you can build a RAG-style search system using only grep, but cautions that this approach is difficult and not for the faint of heart.
#12 in
Udi Menkes introduces DESIGN.md, a plain-text markdown file capturing complete design systems (colors, typography, spacing, components) from 55 production sites like Linear, Stripe, and Airbnb.
#13 in
Gohar Ghukasyan š Starting April 14, HubSpotās Breeze AI shifts from $1 per conversation to $0.50 per resolved conversation, so you only pay when the AI actually fixes the issue. This outcome-based pricing aligns incentives and makes ROI crystal clear.
#14 ā¶ļø
Cursor ditches VS Code, but not everyone is happy...
Fireship
Cursor 3.0, rewritten in Rust and TypeScript and powered by its in-house Composer 2 model (based on Moonshotās Kimmy K2), replaces the VS Code fork with an AI-agent orchestration interface across local repos, remote SSH sessions, and the cloud.
- Composer 2 outscored Claude Opus 4.6 on āTrust Me Broā benchmarks for intelligence, speed, and cost, but its metadata model ID revealed it is Moonshotās Kimmy K2 retrained with reinforcement learning.
- Cursor 3.0 dropped the VS Code core, offering āplan modeā for parallel agent execution with visual cues (yellow dots for permission requests, blue for completed tasks) and full dev-tool integration.
- In a Horse Tinder demo, Cursorās agent swarm produced 13,000 lines of code within minutesāincluding architecture planning, landing page generation, live SVG Clydesdale rendering in a built-in browser, and automated CSS fixes via design mode.
#15 š
DeepLearning.AI is partnering with ReductoAI at AI Dev 26 to showcase their platform that turns complex, unstructured documents into structured, LLM-ready data with industry-leading accuracy. Join their speaker session and visit their booth in San Francisco on April 28ā29.
#16 š
Brian Balfour argues that while AI can boost smaller teamsāas Gautam Gupta noted raising the Series A barāit overlooks the second-order competition effects. Once everyone has access to the same AI, the performance bar simply moves higher.
#17 š
claire vo š¤ notes that AI-driven engineering has stretched PM:engineer ratios to roughly 1:20, so @AnthropicAI is hiring more PMs. She highlights @chatprd as a way to turn engineers into de facto PMs āwithout the meetingsā and calls for similar AI tools to boost PM leverage.
#18 ā¶ļø
How I use iMessage and AI to run my life
Greg Isenberg
Demonstrates the two-minute setup and daily use of Lindy Assistant, an AI executive assistant in iMessage that proactively triages emails, manages calendar events, and conducts research by ingesting your inbox, calendar, Slack, and Google Drive for $49/month.
- Initial setup takes under two minutes: supply your phone number and connect your Google account, after which Lindy ingests email, calendar, Slack and Google Drive into its memory.
- In the morning briefing demo, Lindy triaged 63 overnight emails and drafted 4 replies, reported San Franciscoās temperature as 62Ā°F, and detected that Gary Danko is closed on Tuesdays to reschedule the invite to Comic Seafood (2 minutes away).
- Pricing starts at $49/monthāsufficient for over 90% of usersāwhile heavy āwhaleā usage (e.g., polling for new restaurant reservations every 15 minutes) may require an upgrade.
#19 in
Peter Yang shares that OpenAI skips medium-term roadmaps, focusing instead on concrete eight-week sprint goals and a year-long vision, using short-term bets to drive their long-term AI improvements.
#20 š
Peter Yang says OpenAIās fastest-growing Codex uses only 8-week sprints or directional long-term planningāno 6ā12 month roadmaps. After momentum in its IDE extension and CLI, the team launched a minimalist app to manage multiple AI agents based on power-user workflows.
#21 š
Anthropicās run-rate revenue has jumped to $30 billion from $9 billion at the end of 2025 as demand for Claude accelerates. Its new Google-Broadcom compute partnership ensures it has the infrastructure to keep pace.
Also covered by: @Lenny Rachitsky
#22 š
Lenny Rachitsky shares a YouTube interview with their head of growth, who breaks down proven user-acquisition playbooks, key metrics, and scaling tactics.
#23 š
Santiago: Linkup built a web index from scratch that extracts, timestamps, and sources individual facts for semantic retrieval. Instead of returning full pages, it delivers the exact information AI agents need, essentially enabling RAG across the entire web.