We're having the wrong conversation about AI in cancer research. One side: "Trust the black box, it works" Other side: "Can't trust what we don't understand" Both miss the point. The real question: How do we build AI we can understand, validate, and improve? New post 🧵

🧵 1/5 GitHub just released Agent HQ, which creates a unified workflow where developers can orchestrate any AI agent, anywhere. This isn't just another tool—it's a fundamental shift in how we build with AI.

🤓 Week 1 journey with my new GPU toy: Days 1-3: Built AI gateway, optimized workflows, deployed infrastructure ✅ Day 4: Hit major roadblocks. Training crashes, memory issues 🚫 Week 1: Found the root cause. 3.6x performance breakthrough 🚀 Why share the whole journey? Thread 🧵

🧵 1/4 OpenRouter processes billions of LLM requests monthly and has identified a critical issue: identical model weights don't always produce identical results across providers. Today they're launching exacto endpoints for higher tool-calling accuracy.

Running Claude Code across Mac, Pi, and DGX? Configuration drift is real. I built an intelligent sync system that: ✅ Syncs agents, commands & hooks 🔒 Preserves machine-specific configs ⚡ Resolves conflicts interactively 💾 Auto-backups everything Open source →

🧵 1/4 Breakthrough in AI model fine-tuning: Write a simple CPU script and automatically run it on distributed GPUs. New tool Tinker makes fine-tuning large language models accessible and efficient. 🚀

🧵 Every AI agent demo looks the same: order food, grab a ride, book a ticket. Cute. But here’s the bigger truth: each one of these has a payment attached. We’re not just watching assistants. We’re watching the scaffolding of an economy.

The first major ChatGPT usage study just dropped. 700M users, 18B messages/week. The reality will surprise you: Only 4.2% is programming. 70% isn't work-related. More people use it for tutoring than coding. The AI revolution looks nothing like we expected.

You can now generate high-quality images directly within Claude. By connecting to Hugging Face, the chatbot gains access to powerful, open-source image generation tools, streamlining the creative process from prompt to picture. 🧵 1/4

The open-source AI race is accelerating. Two massive new models, Kimi K2-0905 and Qwen3-Max, have just been released, signaling major advances in AI capability and scale, particularly from China. 🧵 1/4

We're obsessed with "what's next?" in AI while 95% of organizations can't use what exists today. The "what's next" fallacy is killing real progress. Even if AI development stopped tomorrow, most companies would need YEARS to catch up with current capabilities. Stop waiting. Start building. 🧵

95% of generative AI investments have produced zero returns, reports the MIT Media Lab. This isn't an AI failure—it's a strategy failure. Many leaders are funding scattered pilots that don't connect to real business value, repeating past mistakes. 🧵 1/4

AI research this week was absolutely unreal 🤯 Medical AI just hit 72% cancer detection from blood samples with 98% accuracy (Johns Hopkins MIGHT method) Meanwhile, AI-Researcher can now run entire research projects autonomously - lit review to final paper The tools are becoming colleagues 1/3 🧵

The AI revolution is exhilarating, but let's be honest: it can also feel like a second job just trying to keep up! I'm bullish on AI, but the "cognitive tax" is real. My new blog post dives into how we're navigating this overwhelm. 🧵1/4

Is the rapid progress in AI slowing down? The excitement around AI was fueled by a 2020 OpenAI paper on “Scaling Laws,” which predicted that models would improve exponentially as they got bigger. But that trajectory may be changing. 🧵 1/4

The latest major open model, DeepSeek V3.1, arrived with a surprisingly low-key release. While knowledge benchmarks show incremental gains, the true advances are in coding, agentic capabilities, and improved token efficiency. 🧵 1/4

This week in AI showed the split: speed vs. reasoning. Mercury diffusion = 10x faster inference DeepSeek V3.1 = trillion-scale open weights GPT-5 = reasoning leap, but persona backlash The new moat? Speed + empathy + energy. #AI #LLMs @InceptionLabsAI (1/4) 🧵

Traditional change management advice often fails during uncertainty. New research indicates that gradual tactics like pursuing 'small wins' can hinder leaders when turbulence hits, holding them back from bigger opportunities. 🧵 1/4

Health tech developers: Need to build AI that understands complex medical documents without manual uploads? Google's Gemini API just made a major update for you. The URL context tool is now generally available, changing how AI interacts with specific data sources. 🧵 1/4

China built AI that diagnoses diseases as well as doctors. Only 0.7% of hospitals use it. This isn't about AI failing—it's about why breakthrough tech and real adoption are totally different challenges. Thread on the 99.3% problem: 🧵 1/4

My AI assistant was brilliant but clueless about my own work. So I gave it a brain transplant—by connecting it to my entire Obsidian vault. 🧠⚡️ Now I have a personal AI "Chief of Staff" that knows my projects inside and out. Here's how it works. 🧵 1/4

New benchmarks reveal hosting impacts AI model performance more than expected. gpt-oss-120B varies up to 10% on reasoning tasks (GPQA Diamond x16, AIME 2025 x32) across providers. Top: @PascalAI (78.8%), @TogetherAI (76.5%), @FireworksAI (74.3%). Bottom: @Azure (70.6%), @awscloud (70.3%). 🧵 1/4

This week in AI saw major strides in three key areas: more capable AI agents, groundbreaking medical applications, and new techniques for enhanced LLM reasoning. Here's a quick look at the latest research shaping the future of artificial intelligence. 🧵 1/4

While Big Tech races to build more powerful AI models, one 4B parameter model just quietly outperformed Perplexity Pro at 15x lower cost Running entirely on your local hardware with zero data sharing The future of AI might be smaller than you think 🧵 1/4

Looking to build a powerful AI/ML portfolio? A single GitHub repository offers over 500 projects with code, providing a massive launchpad for your career. 🧵 1/4

New blog: Context engineering is the new pipette. GPT-5 and GPT-OSS aren't competitors. They're lab partners. One explores (cloud, 400K tokens, parallel tools). One executes (local, private, hackable). Stop asking "which model?" Start asking "which tool for which job?" 🧵 1/4

🚀 Run AI Run 2.0 is LIVE! This week's newsletter unpacks OpenAI's bold gambit: GPT-OSS to woo open-source devs & GPT-5 to dominate enterprise. Plus, AI automating AI with Google's MLE-STAR. 🧵 1/4

Struggling to integrate AI models into your coding infrastructure? A new GitHub project, claude-code-router, offers a powerful solution to streamline how you interact with Claude Code. 🧵 1/4

Agents Everywhere: From Diffusion Coders 2 CRISPR-GPT bit.ly/4li7U4D Run the Reads is published every Wednesday as part of the "SO"cial series. This week’s recap highlights autonomous ML engineering, diffusion-based coders, and AI-driven genomics breakthroughs #AI #MachineLearning #Research #LLM

OpenAI just changed the AI game by open-sourcing gpt-oss-120b A 120B parameter model that runs on ONE 80GB GPU. Apache 2.0 license. Performance approaching closed-source frontier models. But that's not even the most interesting part 🧵

1/ 🧵 I trained an AI to write cyberpunk fiction about Meta's future. What emerged was a chilling story about Sarah, a marketer whose AI glasses give her superpowers at work... while her colleagues mysteriously vanish. The twist ending made me question everything about AI "empowerment" 👇

Meta has announced its push toward personal superintelligence, claiming it's now within reach as AI begins to improve itself. What could this mean for individual empowerment in a rapidly evolving tech landscape? 🧵 1/4

AI-powered social content generator that transforms URLs into engaging posts for X/Twitter, LinkedIn, and Instagram with modern UX and accessibility features

Staring at 7 ChatGPT models like you're lost in a cereal aisle? 😅 OpenAI's overloaded menu (GPT-4, GPT-4o, etc.) has users paralyzed by choice. How do you pick the RIGHT model for YOUR task? The struggle is real. 🧵 1/4

OpenAI is launching its first European AI data center in Norway—a strategic move accelerating the continent's AI sovereignty efforts. This positions Europe for greater control over critical compute infrastructure amid global competition. 🧵 1/4

The Economist just revealed AI could trigger 20-30% economic growth—but warned "all income eventually accrues to owners of capital." Are you building to own or renting to operate? 🧵 1/4

This week I curated some fascinating stories that reveal AI's biggest paradox: $600B in infrastructure spending while the most impactful breakthroughs come from efficient, smaller systems. The disconnect is stunning. 🧵 1/4

Oracle just committed 2 MILLION chips to OpenAI in a 4.5GW deal while China's $100B+ AI data centers sit largely EMPTY. xAI raises $10B + Pentagon contracts. OpenAI hits $300B valuation. But infrastructure without expertise = expensive real estate. 1/2

53.7 % LiveCodeBench pass@1 – outperforming GPT‑4.1 💻 – Kimi‑K2 is a 1 T‑parameter open LLM with agentic workflows & MuonClip optimization. Thoughts on open‑weight super‑models? #AI #OpenSource bit.ly/44VsQrL

Every Wednesday I publish Run the Reads - my weekly breakdown of the most important AI research papers. I read through dozens of papers and pull out 20 that actually matter. No fluff. Just the breakthroughs, the warnings, and the stuff that changes how we think about AI.

Brain-inspired Hierarchical Reasoning Model nails Sudoku-Extreme, mazes & scores 40.3 % on ARC-AGI with just 27 M params and 1 k examples—no CoT, no pretraining! Paper ➜ bit.ly/4o0z9mD Code ➜ bit.ly/44I2TNz #AI #DeepLearning #Reasoning

Did you know AI can reduce diagnostic errors in primary care by 16%? A new study from Kenya shows how an AI copilot is transforming healthcare delivery in resource-limited settings, offering a safety net for clinicians without overriding their expertise. 🧵 1/4

Wondering what's buzzing in the AI Twitterverse? Catch this week’s roundup: key insights, trending tweets, and must-know discussions distilled into one quick read. Stay sharp, stay informed! #AITwitter #RunTheFeeds bit.ly/4lK8mJQ

Can AI truly understand science? SciArena tests 23 top models on real research questions, revealing surprising gaps in their abilities. Here's what researchers found. 🧵 1/4

ARC-AGI-3 just dropped, revealing a massive gap: humans ace it at 100%, while frontier AI starts at 0%. But agents are already solving levels – is this the benchmark that finally pushes AI toward true generalization? 🧵 1/4

Just published: "Supercharging Code Discovery: My Journey with Roo Code's Free Codebase Indexing" - How I set up professional-grade semantic search using completely free tools (Qdrant Cloud + Google Gemini). No more grep archaeology! bit.ly/3TJSx9Gm #RooCode #SemanticSearch #DeveloperTools

Perplexity just launched Comet, an AI browser that transforms how we interact with the web. Instead of browsing, you're thinking out loud while AI handles the heavy lifting. 1/4 🧵

When OpenAI blocked China's AI access last year, Chinese developers didn't panic—they innovated. Now companies like DeepSeek and Alibaba are releasing open-source AI systems that rival US giants at a fraction of the cost. 1/4 🧵

New Cell paper launches the #VirtualCellChallenge—an open benchmark where AI must predict CRISPRi responses in held-out stem cells. 300 k single-cell profiles, $175 k in prizes, and a live leaderboard. Join the race to pass the cellular Turing test! bit.ly/4lPuCl8 #SingleCell #AI

Two new AI coding tools just launched with opposite philosophies: Amazon Kiro forces structure (specs → design → code) Anthropic Claude Code gives flexible power This isn’t about features. It’s about choosing how AI changes software development. open.substack.com/pub/rundatar...