Justin
@justinhjohnson.com
Executive Director @ AstraZeneca | Nexus of Data, Science, Tech | Global Business Leader | Top Data Science Voice | #datascience #AI #buildinpublic #indiehacker
Blog rundatarun.io
BuildInPublic jandsgroupllc.com
Blog rundatarun.io
BuildInPublic jandsgroupllc.com
🧵 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.
October 30, 2025 at 11:55 AM
🧵 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 🧵
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 🧵
October 28, 2025 at 11:09 AM
🤓 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 🧵
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.
October 23, 2025 at 11:01 AM
🧵 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 →
I built an intelligent sync system that:
✅ Syncs agents, commands & hooks
🔒 Preserves machine-specific configs
⚡ Resolves conflicts interactively
💾 Auto-backups everything
Open source →
October 20, 2025 at 4:10 PM
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 →
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. 🚀
October 3, 2025 at 1:22 PM
🧵 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.
We’re not just watching assistants. We’re watching the scaffolding of an economy.
September 20, 2025 at 11:00 AM
🧵 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.
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.
September 16, 2025 at 11:15 PM
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
September 7, 2025 at 4:00 PM
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
September 6, 2025 at 4:00 PM
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. 🧵
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. 🧵
September 5, 2025 at 2:09 PM
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. 🧵
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
August 29, 2025 at 10:04 PM
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 🧵
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 🧵
August 28, 2025 at 2:45 PM
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 🧵
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
August 27, 2025 at 4:15 PM
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
August 25, 2025 at 4:01 PM
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
August 24, 2025 at 10:07 AM
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
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
August 21, 2025 at 12:27 PM
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
August 20, 2025 at 7:20 PM
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
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
August 17, 2025 at 11:15 AM
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
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
Now I have a personal AI "Chief of Staff" that knows my projects inside and out.
Here's how it works. 🧵 1/4
August 15, 2025 at 11:15 AM
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
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
Top: @PascalAI (78.8%), @TogetherAI (76.5%), @FireworksAI (74.3%).
Bottom: @Azure (70.6%), @awscloud (70.3%).
🧵 1/4
August 14, 2025 at 4:02 PM
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
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
August 14, 2025 at 12:40 AM
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
Running entirely on your local hardware with zero data sharing
The future of AI might be smaller than you think 🧵 1/4
August 13, 2025 at 3:59 PM
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
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
August 11, 2025 at 9:10 PM
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
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
August 11, 2025 at 11:15 AM
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
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