Anthropic removes Claude Code from Pro plan, now $100/month

One of the three rules I'll cover Thursday: Don't connect an AI agent (eg. OpenClaw or Hermes) to anything you can't undo.👇 Email with a 30-day trash? ✅ Fine. Draft email before sending? ✅ Fine. Git with version control? ✅ Fine. Production database deletes? 🛑 Not fine. Money transfers? 🛑 Not fine. Send unverified emails to customers? 🛑 Not fine. The mistake here isn't trusting AI. The mistake is connecting it to a place where mistakes are permanent. I'll explore why this is not okay, and how to make it okay in depth on Thursday. I'll also explore two more rules + a live demo on Thursday. Skelin Quarterly Meetup #2, Reykjavík, May 21. 👉 https://lnkd.in/dsWyuBeC Come and join me, business leaders and tech people are welcome!

OpenAI’s Codex and Image-2 wins might actually be Anthropic wins. And Anthropic knows it. 48 hours after Image-2 launched, timelines are flooded with pixel-perfect generations (ours too). Codex already has ~4M users -- many joining after hitting Claude Code's new token limits. But Anthropic doesn’t seem bothered. Tightening limits. They’re raising prices. And pushing users away. Why? Because they don't have cheap compute. They’re not trying to win attention. They’re trying to win serious workloads. Images, hobbyists, and free users can live on ChatGPT. Anthropic wants something else: teams that pay for reliability, reasoning, and performance. OpenAI believes AI will be a "winner-takes-all". Anthropic believes otherwise. One company optimizes for scale and distribution. The other for high-value compute customers. Both might win. But one of them will be more refined. Same revenue. But with 1/10th the compute and 1/4th the funding. Who do think is smarter here?!

The siren song to ship faster with AI is hard to resist. My teammate and co-founder 💡 Matthew S Krick had Claude research and generate this graph while frustrated by a GitHub outage. He asked it plot the number of reported service incidents by year. Notice anything? More AI = more incidents

Most LLM apps don’t fail because of the model. They fail because of context. In long-running systems, context becomes a silent bottleneck. The problem:   As sessions grow, prompts accumulate noise.   Token limits force trade-offs.   Relevant signals get diluted by stale or redundant data. I’ve seen systems where responses degrade after ~10–15 interactions, not because of the model, but because context wasn’t managed. Context:   Building conversational features where users expect continuity across sessions, not just single prompts. What actually worked: - Split context into layers: **system / session / user memory**   - Aggressively prune: keep only state that affects the next decision   - Use summaries instead of raw history after N turns   - Store structured memory (key-value) instead of replaying full chats   - Re-inject only relevant slices per request Result:   More stable outputs, lower token usage, and predictable behavior across long sessions. Principle:   Context is not history. Context is a decision boundary. If it doesn’t influence the next output, it doesn’t belong in the prompt. Question:   How are you handling context decay in multi-turn LLM systems without over-engineering memory? 🤔 #AI #SoftwareEngineering #LLM #SystemDesign #PromptEngineering

Anthropic is shifting the agentic AI landscape. The company just reversed its ban on third-party agents like OpenClaw, but there’s a major catch: the end of "compute arbitrage." Developers are calling it a "massive devaluation," with some claiming an effective 25x cut in usage for external tools. Is this a necessary move for stability or a "plug pull" on power users? Read the full story 👉 https://buff.ly/rjH5brN #Claude #Anthropic #AI #LLM #OpenClaw #DeveloperTools

Most of an AI app isn't AI. NewsCue runs 7 stages on every article before it reaches your phone. Only 3 of them call an LLM. The rest is fingerprinting, scoring, ranking — deterministic plumbing that makes the AI stages trustworthy. I knew from day one this needed a pipeline. What I didn't know was what each stage would teach me once it started running in production. Three things I'd tell anyone building an AI content pipeline: Design for multiple cadences. Full pipeline every 4 hours. Breaking news check every hour. Trend bursts on demand. One schedule for everything means you're either too slow for breaking news or burning money on routine runs. Every stage breaks differently. Sports articles leak into tech. Diaspora context returns null because not every story has an NRI angle. RSS feeds go silent at US-night hours. You don't find these until production, and each one needs its own fix. Prompt caching is not optional. Three LLM stages processing hundreds of articles a day adds up fast. Caching cut my LLM costs ~75%. Swipe through the cards for the full breakdown. What stage looks most interesting to you — or most likely to be the source of bugs? Next Thursday: Stage 3 — why categorization is harder than it looks, and how a sports story kept leaking into tech. #ShippingNewsCue #AIBuilder #VibeCoding #SideProject

If you're into AI, you're going to want to read this-> Scaling laws work for Time Series Foundation Models!!! Ok that doesn't sound that exciting. But it really is. What are Scaling Laws? You know how ChatGPT got dramatically smarter as it got bigger? That's not magic. More data + more compute + bigger model = predictably better results. That's a scaling law. It's basically the reason the entire LLM revolution happened. What's a Time Series model? Any data that changes over time. Weather. Stock prices. Energy consumption. Server CPU. For decades, predicting this stuff required handcrafted models and lots of domain expertise. Then foundation models arrived, one model that generalizes across ALL time series, like an LLM generalizes across ALL text. Promising. But did making them bigger actually make them better? Nobody knew. Until now. Datadog just released Toto 2.0, open-weights models from 4M to 2.5B parameters! And for the first time ever: bigger = consistently better. Every size beats the previous one. Clean curve. No exceptions. #1 on every major benchmark (BOOM, GIFT-Eval, TIME). How to read the charts 👇 X axis = model size. Y axis = CRPS rank (lower = better). Toto 2.0 is the purple line. See how it keeps going down as it gets bigger? That's the scaling law working. No other model family does that. Why should you care beyond Datadog? Because time series is everything: weather forecasting, energy grids, financial markets, healthcare, infrastructure... Every industry that runs on data over time just got a much more powerful foundation to build on. And it's no coincidence this comes from Datadog. With its own AI research lab, models running in production, and the largest time series dataset in the world, Datadog is leading AI research and development for observability. Data advantage + research advantage. And fully open source, Apache 2.0, available now with open-weights on Hugging Face Proud to work at the company that just proved this. 🐶 #AI #TimeSeries #Datadog #Toto #OpenSource #MachineLearning

One of the biggest mistakes teams make while building AI agents? They think only the latest message costs money. It doesn’t. Every single API call resends the entire conversation history again and again. Which means: longer chats = exponentially higher costs slower responses wasted tokens context pollution scaling nightmares in production A lot of teams realize this only after their AI bill suddenly explodes 🚨 Recently while working on production-style Agentic AI systems, I noticed how misunderstood token economics still is — especially around: ✅ conversation memory ✅ multi-agent systems ✅ RAG workflows ✅ context windows ✅ token accumulation So I wrote a deep dive breaking down: how token math actually works why long-running chats become expensive the “conversation tax” and how production systems optimize memory & context intelligently If you are building AI agents, this will save you money and architectural pain later. 📖 Read here: The Hidden Cost of AI Agent Conversations — Token Math Every Team Gets Wrong #AI #AgenticAI #LLM #GenerativeAI #RAG #GoogleCloud #MCP #ADK #ArtificialIntelligence #TechTrapture

In the Agent Age, it is a fundamental mindset shift AND a ways of working shift for Enterprise AI. Great insights from Wade Weirman

If your AI agent is "mostly helpful" but "occasionally unhinged," unit tests won't save you. In the era of agentic AI, correctness is a moving metric. TDD was built for deterministic software. Large language models are not that. I've just written up why __Evaluation-Driven Development (EDD)__ is no longer optional for enterprise AI — and how the emerging pattern of LLM-as-a-Judge, golden-set evaluation, and CI/CD gating (via tools like the Strands Evals SDK on AWS) brings real engineering discipline to non-deterministic systems. Particularly relevant if you're deploying agents into regulated environments — financial services, health, or the public sector — where "it usually works" isn't a strategy. 🔗 https://lnkd.in/gafMkR8x #LLMEvaluation #AIAgents #GenerativeAI #LLM #ResponsibleAI #Strands