Notified

This is insane... An open-source model with 3 billion active parameters just scored 73.4% on SWE-bench Verified. Claude Opus 4.6 scores 75%. The gap is 1.6 points. The cost difference is 10 to 30x. Alibaba dropped Qwen3.6-35B-A35. 35 billion total parameters, 256 experts, but only 8 routed plus 1 shared activate per token. So you're running 3B active parameters at inference time. On a laptop. Simon Willison ran it locally and it drew a better pelican than Claude Opus 4.7. (Yes, the pelican benchmark is real, and it's a surprisingly good vibes test.) But the part nobody's talking about: Thinking Preservation. Current models re-reason from scratch every turn. This model retains its chain-of-thought traces across multi-turn conversations. In agent loops where the model makes 50 to 100 tool calls, that eliminates massive redundant reasoning overhead. 262K context native. Extensible to 1M. Apache 2.0 license. The benchmark race is mostly over. The real race now is cost per intelligence. And 3B active parameters matching frontier performance changes that equation completely.

The most productive engineer on GitHub last week got permanently banned. His crime? He used an AI agent that actually worked. A Korean CTO built a 13-step agent harness. Pointed it at 100+ major open-source repos. In 72 hours: 500+ commits. 130+ pull requests. Some merged by Kubernetes maintainers. Some merged by Hugging Face. Then GitHub banned his account for spam. Read that again. The PRs were good enough for Kubernetes to merge. But GitHub's abuse detection flagged the entire account. The platform that hosts your agents cannot tell the difference between a 10x engineer and a spammer. And it chose spammer. This isn't a one-off glitch. This is the infrastructure layer sending a message: we aren't built for agentic workflows yet. Your coding agent's biggest bottleneck isn't the model. It's whether the platform will let you use it.

Cloudflare proved most MCP implementations are wrong. After deploying MCP at enterprise scale, they found the default pattern (register all tools upfront) breaks the moment you connect more than a few servers. The math: a typical MCP setup with 52 tools burns roughly 9,400 tokens just describing what's available. Before the model does anything useful, 15% of your context window is already gone. Their fix is almost offensively simple. Collapse all tools into 2. Tool 1: Search (discover available tools on demand). Tool 2: Execute (run any tool by name with params). The model writes JavaScript in a sandboxed runtime to discover and chain tools. No upfront registration. No tool list bloat. 52 tools become 2. 9,400 tokens become roughly 600. 94% reduction. And the cost stays flat no matter how many MCP servers you add. The insight that matters: every MCP tutorial teaches "define your tools." Cloudflare's production system proves that approach doesn't scale. The model doesn't need a menu. It needs a search bar.

50,000 developers starred an AI trading repo. TradingAgents hit 50K stars by copying the one thing quant funds actually got right: making AI agents argue before they trade. TradingAgents doesn't use one LLM to trade stocks. it builds an entire simulated trading firm. fundamental analyst. sentiment analyst. technical analyst. bull researcher. bear researcher. risk management team. portfolio manager. here's the part that actually matters: the bull and bear researchers literally debate each other before any trade happens. they argue over market signals, weigh conflicting data, then pass that synthesis to a trader who has to defend the decision to risk management. sound familiar? that's how real prop trading desks work. the result: better cumulative returns, higher Sharpe ratio, and lower max drawdown than single-agent baselines across all tested tickers. everyone's building one-agent trading bots with a fancier prompt. meanwhile this open-source framework (built on LangGraph, works with GPT-5.4, Claude 4.6, Gemini 3.1) proved the edge isn't a smarter model. it's structured disagreement between specialized agents. one genius agent isn't the answer. a team of agents with built-in conflict is. link in comments

you're PAYING your LLM to read GRAMMER it already knows. every "in order to", "therefore", "however" and "the user is asking" in your prompt is a token the model can predict with near-100% confidence. you're literally buying compute for words it would fill in itself. Caveman Compression strips exactly those tokens. keeps facts, numbers, names, constraints. deletes everything the model can reconstruct. the results are hard to argue with: system prompts: 171 tokens down to 72. that's 58% gone. agent reasoning chains: 196 tokens to 102. 48% reduction. RAG knowledge bases: 199 to 118. 41% fewer tokens, 2-3x more context per query. factual preservation benchmark: 13/13 facts survived. 100%. comes in 3 flavors. LLM-based (40-58% reduction, needs API key). MLM-based using RoBERTa (20-30%, free, runs locally). NLP rule-based (15-30%, free, offline, under 100ms, supports 15+ languages). the NLP version is the interesting one for production. no model calls, no latency, no cost. just regex and spaCy ripping out grammar at <100ms per document. everyone's optimizing prompts by rewriting them. nobody's asking which tokens the model doesn't even need to see. link in comments

Day 5 of 7: Anthropic Ate Their Own Cooking And the Results Are Worth Studying 3 teams marketing, design, legal are building things that used to need engineering tickets and sprint planning. None are staffed with developers. GROWTH MARKETING: A Team of One, Operating Like Ten One non-technical marketer built: → Automated Google Ads generation: An agentic workflow that reads ad performance CSVs, identifies underperformers, generates hundreds of variations within character limits. 2 hours of work → 15 minutes. → A Figma plugin: Programmatically generates 100 ad variations by swapping headlines and descriptions. Half a second per batch. 10x creative output. → A Meta Ads MCP server: Queries campaign performance directly in Claude Desktop. No platform switching. Zero engineering support requests. How to replicate: • Survey non-engineering teams for repetitive workflows with API-connected tools • Start simple: CSV input → structured output • Break workflows into specialized sub-agents (headline agent vs. description agent) • Plan in Claude.ai first, then implement in Claude Code PRODUCT DESIGN: "I Am a Developer Now" Two distinct experiences emerged: → Developers: "augmented workflow" faster execution of known tasks → Non-developers: "holy crap, I am a developer" entirely new capabilities Designers now make state management changes engineers did not expect from them. They paste mockup images into Claude Code and get functional prototypes engineers can immediately build on. A GA launch that needed "research preview" messaging removed across the entire codebase coordinated with legal went from a week of back-and-forth to two 30-minute calls. Figma and Claude Code are now open 80% of their working time. How to replicate: • Have an engineer help one designer with initial repo setup • Create a custom memory file: "I am a designer with limited coding experience. Explain changes in detail. Make small incremental changes" • Start with visual polish — spacing, colors, typography • Graduate to image-to-prototype: paste screenshot, ask Claude to build it LEGAL: Building What Nobody Asked Them to Build A lawyer built a communication assistant for a family member with speaking difficulties. In one hour. Predictive text with speech-to-text and voice banks. They also built team coordination tools, workflow automation for legal reviews, and prototype "phone tree" systems. Their process: plan in Claude.ai , build step-by-step in Claude Code, use screenshots to show what they want, share prototypes even when imperfect. That last point is their biggest tip. Sharing rough prototypes inspired other departments to see possibilities. How to replicate: • Run a 90-minute build session with non-engineering teams. One prompt: "What repetitive task do you wish a computer would do?" • Standardize the two-step process: plan conversationally, build incrementally • Create a monthly demo session. Celebrate rough prototypes • The cultural shift matters more than any individual tool