Anthropic Beats OpenAI in Enterprise Spend for the First Time — Ramp Has the Receipts

💡 INSIGHT

Anthropic Beats OpenAI in Enterprise Spend for the First Time — Ramp Has the Receipts.

Ramp's corporate card data shows Anthropic overtaking OpenAI in business spending for the first time. This isn't a vibes poll — Ramp processes billions in real corporate transactions, making it one of the most credible signals of where enterprise AI dollars are actually flowing. The shift tracks with Anthropic's aggressive push into coding tools and agentic workflows, which generate sticky, high-volume API usage rather than one-off ChatGPT seats. If your team hasn't benchmarked your API spend breakdown recently, now's the time. (782 likes | 97 RTs) Read more →

Anthropic's Own Engineer Asks Twitter: Where Does Claude Still Lose?

An Anthropic engineer publicly asked users when they reach for other models instead of Claude — and 1,260 likes later, the thread is a goldmine of unfiltered user pain points. This kind of open solicitation from inside a frontier lab is rare and signals the company is actively hunting for capability gaps rather than relying on benchmark scores. The candid responses cover everything from structured output reliability to specific reasoning failures. Worth reading whether you're an Anthropic customer or a competitor. (1,260 likes | 75 RTs) Read more →

Contrarian Take Gains Steam: AI Won't Actually Make Your Processes Faster. A blog post arguing that AI accelerates execution but can't compress the human coordination, decision-making, and alignment that actually bottleneck most processes is gaining massive traction on Hacker News. The core insight: if your slowdown is "waiting for three stakeholders to agree," a faster code generator doesn't help. Audit where your team's real bottlenecks are before throwing AI at them. (471 likes | 335 RTs) Read more →

Gruber: AI Is Infrastructure, Not a Product Category. John Gruber argues that AI should be embedded technology — like electricity or networking — not standalone products. The framing directly challenges the current "AI app" investment thesis and aligns with Apple's approach of weaving intelligence into existing workflows rather than shipping a chatbot. If you're building an AI feature, ask whether it should be a product or a capability. (302 likes | 120 RTs) Read more →

ArXiv Drops the Hammer: One-Year Ban for Fully AI-Generated Papers. The world's largest preprint server will now ban authors for a full year if they submit papers where AI did all the work. This is the first major enforcement mechanism against AI slop in academic publishing — and it sets a precedent other journals will likely follow. Review your lab's AI-assisted writing policy before someone else reviews it for you. Read more →

🔬 RESEARCH

LeCun Sets the Clock: Hierarchical World Models in 12–18 Months.

Yann LeCun just put a concrete timeline on his long-standing world model thesis — within 12–18 months, he expects a general method for training hierarchical world models. If Meta FAIR delivers, it would represent a fundamental architectural shift from next-token prediction to structured understanding of how the world works. LeCun has been beating this drum for years, but a public timeline from Meta's chief AI scientist means there's internal progress worth watching. Mark your calendar for late 2027. (838 likes | 97 RTs) Read more →

Running Local LLMs on Apple Silicon Actually Costs More Than Cloud APIs. A detailed energy cost analysis shows that when you factor in hardware amortization, electricity, and idle time, running models locally on Apple Silicon is more expensive per token than routing through OpenRouter or similar cloud APIs. The "local is free" narrative takes a direct hit — and the math gets worse as cloud API prices keep dropping. Run the numbers for your own setup before committing to local inference. (290 likes | 242 RTs) Read more →

The Open Model Flood: Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 All Drop. Nathan Lambert at Interconnects calls this the most eventful month for open-weight releases in AI history. His comprehensive assessment covers Google's Gemma 4, DeepSeek V4, Moonshot's Kimi K2.6, and more — with CAISI's evaluation showing the gap between open and closed models continuing to narrow. If you're choosing a self-hosted model, this is required reading. Read more →

🔧 TOOL

Anthropic Quietly Ships claude-code-setup: One Plugin to Wire Your Entire Dev Environment. An official Anthropic plugin that scans your project and auto-configures hooks, skills, MCP servers, subagents, and automations for Claude Code. Instead of spending an hour wiring up your dev environment manually, one install turns vanilla Claude Code into a fully-configured AI coding setup. Install via /plugin install claude-code-setup and let it do the scaffolding. (175 likes | 32 RTs) Read more →

DeepSeek-TUI: A Terminal Coding Agent Built for DeepSeek Models. A Claude Code / Codex competitor purpose-built for DeepSeek models, running entirely in the terminal. With 30K+ stars already, it shows the terminal coding agent UX pattern is becoming model-agnostic and commoditized — the interface is no longer the moat. (30,514 likes | 2,561 RTs) Read more →

📝 TECHNIQUE

Mollick: AI Consumer Products Need Pre-Built Skills, Not Open-Ended Chat. Ethan Mollick identifies the core UX gap holding back AI consumer products: users don't know what to ask. His example — ChatGPT for personal finance — shows that guided workflows and pre-built skill templates dramatically outperform an empty chat box for non-expert users. If you're building any AI tool for mainstream users, add structured entry points before polishing your chat interface. (354 likes | 15 RTs) Read more →

🏗️ BUILD

Singapore's Foreign Minister Builds His Own AI Agent on a Raspberry Pi.

Singapore's foreign minister built NanoClaw, an AI agent running on a Raspberry Pi that he chats with daily via Telegram — his reason: "learn by doing." A sitting cabinet minister soldering together an AI project and publishing it is the most concrete signal yet that AI literacy is becoming a leadership requirement, not a tech hobby. The bar for "understands AI" just moved from "uses ChatGPT" to "can build an agent." (1,087 likes | 140 RTs) Read more →

narrator-ai-cli-skill: One Prompt Generates Full Movie Narration Videos. Open-source CLI skill that plugs into Claude Code or OpenClaw and auto-generates movie commentary videos — script, scene matching, 63 voice options, 90+ templates, and background music. Built for the Chinese short-video content farm use case, but the architecture is a clean example of multi-modal AI pipeline design. (433 likes | 127 RTs) Read more →

Bindu: Open-Source Identity and Payments Infrastructure for AI Agents. The "how do agents pay each other" problem has been blocking autonomous multi-agent systems from going production. Bindu ships open-source identity, communication, and payments rails specifically for agent-to-agent interactions. If you're building multi-agent architectures that need to transact, this is the infrastructure layer to evaluate. (6,094 likes | 377 RTs) Read more →

Can Claude Actually Make You Money on Open-Source Bounties? One Dev Tried. An honest writeup of using Claude to hunt and solve open-source bounties on Algora for profit. The results are mixed — Claude can ship solutions fast, but finding bounties worth the effort and navigating maintainer expectations is still a human bottleneck. Read it before you automate your side hustle. (31 likes | 10 RTs) Read more →

🎓 MODEL LITERACY

Model Routing: Today's agent swarms piece shows teams assigning Opus for UI work, GPT-5.5 for backend logic, and Gemini for vision tasks. This pattern — model routing — dispatches tasks to specialized models based on capability, latency, and cost rather than sending everything to one frontier model. The tradeoffs are real: routing adds orchestration complexity and latency, but can cut costs 5-10x on tasks where a smaller model performs equally well. As open models close the gap on specific benchmarks, routing becomes less about "best model" and more about "right model for this subtask." If you're building compound AI systems, understanding when to route vs. when to use a single model is becoming a core architectural decision.

⚡ QUICK LINKS

The AI Haves and Have-Nots: The gap between companies capturing AI value and those just spending on it is widening. Link
Key LLM Reasoning Researcher Departs Meta FAIR: KempeLab leaves after two years leading reasoning work — watch for their next move. (248 likes | 4 RTs) Link
Clara Health Raises $660M: AI-powered primary care with licensed providers, HIPAA-compliant, can prescribe. (71 likes | 3 RTs) Link
Mollick Tests GPT-5.5 Pro on Academic Humor: Frontier models are crossing into creative territory previously considered AI-hard. (240 likes | 14 RTs) Link
Agent Swarms: Multi-Model Orchestration: Opus for frontend, GPT-5.5 for backend, Gemini for vision — the model-routing pattern goes mainstream. (280 likes | 15 RTs) Link

🎯 PICK OF THE DAY

The Ramp data doesn't just show Anthropic winning — it shows corporate AI crossing from experiment to infrastructure. Ramp's transaction data revealing Anthropic overtaking OpenAI in enterprise spend is striking not just for the leaderboard flip, but for what it reveals about how companies are actually adopting AI. The shift from "everyone gets a ChatGPT seat" to "our engineering team's API bill is a line item" signals that corporate AI has silently crossed from experimentation budgets to production infrastructure. And that transition favors the vendor who ships the best coding tools, not the best chatbot. Anthropic's Claude Code, agent teams, and deep IDE integrations generate recurring, high-volume API usage that sticks — it's the AWS playbook of making yourself indispensable to the build process. OpenAI still dominates consumer mindshare, but enterprise dollars follow developer workflows, and right now those workflows are increasingly Claude-shaped. Read more →

Until next time ✌️