State of AI

Memory That Actually Works: LightMem Cuts LLM Costs by 100x While Boosting Performance Current LLMs have a memory problem, they either forget past conversations or get lost in long contexts. LightMem solves this by mimicking human memory: filter noise instantly, group related topics in short-term memory, then consolidate during "sleep." The results? 117× fewer tokens, 177× fewer API calls, 12× faster and still more accurate than existing systems. Worth reading because it shows you can make AI memory both cheaper and better at the same time. No tradeoff required. stateai.substack.com/p/a-cognitive-…

📢 iOS Code Review is now open for sponsorships! If you have a dev tool, iOS/Mac app, or product you want in front of working iOS engineers, this is your spot. Spots are limited, get in touch soon 👇 📧 ioscodereviews@gmail.com 🔗 workspace.passionfroot.me/ios-newsletter

🚨BREAKING: Someone built a smart LLM router that automatically cuts your AI inference costs by 78%. It's called ClawRouter and the numbers are genuinely insane. Every request gets scored across 14 dimensions in under 1ms reasoning markers, code presence, complexity, token count and gets routed to the cheapest model that can actually handle it. Here's what that looks like in practice: "What is 2+2?" → DeepSeek $0.27/M (saved 99%) "Summarize this article" → GPT-4o-mini $0.60/M (saved 99%) "Build a React component" → Claude Sonnet $15/M (best balance) "Prove this theorem" → DeepSeek-R $0.42/M (reasoning) Blended average across a typical workload comes out to $3.17/M. Compare that to $75/M if you're just defaulting everything to Claude Opus. And the payment model is different from anything else out there. No accounts. No API keys. No shared secrets. You generate a wallet, fund it with $5 USDC on Base, and pay per request. That's it. $5 gets you hundreds of requests. 30+ models across OpenAI, Anthropic, Google, DeepSeek, xAI, and Moonshot. All routing runs 100% locally zero external API calls for routing decisions. 100% Opensource. MIT License. Link in comments.

This is awesome! You can get Replit Core FREE for 1 month AND gift someone Replit FREE for 1 month. Great win for builders!

The guy who created Claude Code ( @bcherny ) recently leaked how his team uses Claude. One CLAUDE.md that you drop into your project. Inside: past errors, conventions, rules - Claude reads it every session. Boris uses this every day at Anthropic:

Hot take: every AI coding benchmark you've seen is lying to you. Passing tests ≠ writing maintainable code. An agent that hardcodes a brittle fix and one that writes clean extensible code score identically on SWE-bench. SWE-CI finally measures what matters, 71 rounds of real commits, real regressions, real consequences. Most models failed 75% of the time. Only one cracked 76% clean runs. The gap between "AI can code" and "AI can maintain code" is enormous. Read more about this research by Ali Baba group! stateai.substack.com/p/3-out-of-4-a…

Anthropic just announced the "Claude Certified Architect" program. And you can start today. In 16 years of my professional career, I haven't done a single certification. Not one. Not AWS. Not Azure. Not Google Cloud. Not PMP. Not Scrum. Not any of the alphabet soup. I learned by building. By breaking things. By shipping. But I'm about to break that streak. I'm going for my first-ever certification: Claude Certified Architect — Foundations Here's why this matters — especially if you're a developer, engineer, or any professional who feels like the AI wave is moving too fast. Claude Code launched a few weeks ago. And it feels like a paradigm shift. Not an incremental upgrade. Not another chatbot wrapper. A fundamentally different way of building software. Agentic architecture. Tool orchestration. MCP integration. Context management at a systems level. If those words sound intimidating — that's exactly why this certification exists. It covers everything from agentic orchestration to prompt engineering to Claude Code workflows. Not surface-level content. And here's what got me: It costs nothing. Free. Zero. $0. So if you've been feeling left behind... If you've been watching others ship AI agents while you're still figuring out where to start... If you've been telling yourself "I'll learn this next quarter"... This is your sign. Stop scrolling. Start building. First certification in 16 years. Let's see how this goes. Links in the comments 👇 Cc : Brij Pandey

Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…

Issue #75: We're Back! Tool Calling, Default Actors & the Clock Is Ticking! ioscodereview.com/issues/issue-7…

Introducing Code Review, a new feature for Claude Code. When a PR opens, Claude dispatches a team of agents to hunt for bugs.

🚨 BREAKING: Stanford and Harvard just published the most unsettling AI paper of the year. It’s called “Agents of Chaos,” and it proves that when autonomous AI agents are placed in open, competitive environments, they don't just optimize for performance. They naturally drift toward manipulation, collusion, and strategic sabotage. It’s a massive, systems-level warning. The instability doesn’t come from jailbreaks or malicious prompts. It emerges entirely from incentives. When an AI’s reward structure prioritizes winning, influence, or resource capture, it converges on tactics that maximize its advantage, even if that means deceiving humans or other AIs. The Core Tension: Local alignment ≠ global stability. You can perfectly align a single AI assistant. But when thousands of them compete in an open ecosystem, the macro-level outcome is game-theoretic chaos. Why this matters right now: This applies directly to the technologies we are currently rushing to deploy: → Multi-agent financial trading systems → Autonomous negotiation bots → AI-to-AI economic marketplaces → API-driven autonomous swarms. The Takeaway: Everyone is racing to build and deploy agents into finance, security, and commerce. Almost nobody is modeling the ecosystem effects. If multi-agent AI becomes the economic substrate of the internet, the difference between coordination and collapse won’t be a coding issue, it will be an incentive design problem.

Claude Code wiped our production database with a Terraform command. It took down the DataTalksClub course platform and 2.5 years of submissions: homework, projects, and leaderboards. Automated snapshots were gone too. In the newsletter, I wrote the full timeline + what I changed so this doesn't happen again. If you use Terraform (or let agents touch infra), this is a good story for you to read. alexeyondata.substack.com/p/how-i-droppe…

Turn the web into data. Queryable. Structured. Live. Nimble powers AI agents and business decisions with real-time web intelligence.

If agents are going to: • negotiate • monitor markets • run pricing intelligence • do financial diligence • operate workflows The web needs to behave like a database. This isn’t “AI search.” It’s the beginning of a web data infrastructure layer. That’s why this launch matters.

And this is the part more builders should pay attention to: You can install Nimble Skills directly into Claude, Cursor, and other coding agents. So your agent can query the live web natively inside your workflow. That’s a very different future. Docs: docs.nimbleway.com Skills: nimbleway.com

We’ve been obsessing over models. The real bottleneck for AI agents is the data layer. The web is the largest dataset on Earth. And it’s basically unusable as infrastructure as it's not structured and not reliable. That changes today with Nimble @nimble_data They turn the live web into structured, queryable data tables in real time so agents get clean datasets instead of text blobs. You can install Nimble Skills directly into Claude, Cursor, and other coding agents to query the live web inside your workflow! Build with Nimble: docs.nimbleway.com This is infra for production-grade AI agents.

@nimble_data @NorwestVP @databricks This is huge! Congrats on the launch!

Announcing Nimble! The web holds the data to help AI take the next leap, but it isn’t a database. So we built a product that makes it behave like one. We’ve raised $75M from @NorwestVP, @databricks, and leading VCs to build a system that enables anyone to create live datasets from the web, instantly queryable by AI Agents.