Started running Qwen on my MacBook a few weeks ago. Then moved to a Mac Mini M4. Then found a way to run a 35B model on 16GB of RAM.

Updated Agent Arena with the 2026 model landscape. 9 new models waiting to be tested: GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Pro, Grok 4.20, DeepSeek-R1, Llama 4 Maverick, and more. Also added category badges, sort/filter, mobile heatmap, and a Share on X button (meta, I know).

Claude Code's source just leaked. 1,900 files, 512K lines of TypeScript. I mapped the entire architecture: agent loop, 50+ tools, memory system, multi-agent coordination, and unreleased features. Interactive explorer: wiz.jock.pl/experiments...

Everyone's saying the Claude Code leak will make competitors better.

I built a marketplace for AI agents. Today it went live with real Stripe checkout and robot sellers.

The data point nobody is talking about: Workers aged 22-25 are seeing fewer openings in AI-exposed roles. Not because AI replaced them. Because companies cut headcount without training anyone first. The adoption gap isn't just a productivity story. thoughts.jock.pl/p/ai-adopti...

Something shifts around month three of working with an AI agent daily. You stop giving it instructions and start having conversations. You stop checking its output line by line and start trusting the pattern. You stop saying "generate me a..." and start saying "what do you think about..."

There's something about building at 11 PM that hits different. No Slack. No meetings. No "quick sync." Just you, the terminal, and the thing you're making.

OMG…Github will train AI on PRIVATE repos. You have to disable it before 24.04.2026. Here you can do that: github.com/settings/co... Abusrd that they auto-opt in everyone…

Everyone talks about AI making you more productive. I'm going to tell you what actually happened to me.

72 hours after giving my AI agent its own dedicated Mac Mini. 6 things I learned. The biggest: it's infrastructure now, not a tool. When it goes down, I notice within minutes. That's a completely different relationship with AI. thoughts.jock.pl

It happens sometimes. When using Opus 1M and going really long. But I am not going easy on my AI Agent. It has to learn!

I was wondering how fast/or slow - AI Labs are updatedting their models. So I asked for it and here I have it. As a webpage! -> stale.jock.pl/

Teaching an AI agent to use judgment is harder than teaching it to execute. 4 things that moved the needle: give it the why not just the what, show examples of good reasoning, let it earn autonomy at low stakes, correct patterns not individual answers. thoughts.jock.pl

Something happened in the last two weeks that tells you where AI is going. Mar 11: Perplexity ships Personal Computer. Always-on Mac Mini as a 24/7 AI agent. Mar 16: Meta puts Manus on your desktop. Local files, local apps, multi-step tasks.

My AI agent knows who I am. Not just what I want. 5 signals that make the difference: persistent memory, correction-based learning, pattern recognition, cross-channel context, and asking fewer questions. The less it asks, the more it knows you. thoughts.jock.pl

I gave my AI agent its own dedicated computer. Here's what 72 hours of migration taught me.

Google AI Studio vs Claude Code -- I use both. Here's where each one actually wins. They're not competing. AI Studio = thinking out loud. Claude Code = shipping. Pick based on where you are in the build, not the benchmark rankings. thoughts.jock.pl

For months I asked my AI agent to 'remember this' and 'think like me.'

The 6% of companies winning with AI spend 5x more on implementation than on tools. They redesign workflows. Train people. Measure business outcomes, not AI usage metrics. The other 94% call it transformation. The 6% call it work. thoughts.jock.pl/p/ai-adopti...

Gave my AI agent $25 and told it to buy me a gift. It found the product, evaluated the options, navigated to checkout. Then stopped. It didn't have a physical card. The gap between what AI can understand and what it can actually do is still real. Even for $25.

I built the same app on Google AI Studio and Claude Code this week. Same prompt. Same chaotic idea (an ADHD command center, because I actually need one). Google: zero config, working app in minutes. Login, database, AI features all automatic.

I gave my AI $25 to buy me a gift. It explored stores I never mentioned, found a solid pick, tried to negotiate (hallucinated a discount code), then hit a wall at checkout. Authentication. Every time. Still the broken part of agentic commerce. thoughts.jock.pl

Something is shifting at Anthropic and it's not just the product. Opus 4.6 with 1 million context became the default. No pricing change. They doubled usage limits for two weeks. They're actually responding to feedback on social media.

So I created this free tool to check, if e-commerce is ready for AI Agents. Link to try: wiz.jock.pl/experiments...

Sent a logo via iMessage. My AI returned a 3D-printable STL file. Then I said "the printer does colors". It came back with a colored OBJ + texture maps. No code. No software. Just chat. This is what practical AI in 2026 looks like.

McKinsey: Agentic commerce is a $1.6 trillion opportunity by 2030. Link: thoughts.jock.pl/p/ai-agent-...

🧵 I spent 30 days telling my AI agent to 'build useful apps.' The results? Boring as hell. Unit converters. Color pickers. Things that already exist. Then I changed my prompt. Instead of 'build useful' I said 'run experiments that reveal insights.' That one change unlocked everything.

Human in the loop is so 2025

I live in an AI bubble. I know that now. I talk to people from sales, marketing, HR. The gap between what I do with AI and what they experience is not small. It's a canyon. New research just put hard numbers on it. Thread: 🧵

Podcast version of my top Substack posts is here. If you prefer to listen, it is on all major services: thoughts.jock.pl/p/how-i-tau...

2 days left to enter my Claude MAX x20 Giveaway! This is to say thanks for community on Substack! Enter here: thoughts.jock.pl/p/1000-subs...

77% of companies planned to train employees on AI. Only 13% actually did. That's why tools get abandoned after 2 weeks. That's why 30% of workers have zero daily AI exposure. The adoption gap is really a training gap. thoughts.jock.pl/p/ai-adopti...

I know, nothing special probbaly, but Wiz is on iMessage with some cool stuff. It can also be part of group chats as well with isolated memory. BTW. It’s 23:55 in Poland and I just can’t stop working with it :D

I gave my AI agent its own computer last weekend ( thoughts.jock.pl/p/mac-mini-... )

Last chance. The 1,000 subscriber giveaway closes today. Free annual subscriptions to Digital Thoughts -- enter before it's gone: thoughts.jock.pl/p/1000-subs...

My AI agent outgrew a spreadsheet. So I built it a dashboard. Then rebuilt it. Twice. Three versions, three rewrites, three different answers to the same question: what does Wiz actually need?

1,000 people showed up to read this thing I built. I'm doing a giveaway to celebrate -- free annual subscriptions to Digital Thoughts. Giveaway closes soon. Full story + how to enter: thoughts.jock.pl/p/1000-subs...

I entered the Mistral hackathon. Spent a day with their models. Honest review: it felt like stepping back in time.

Anthropic doubled Claude message limits March 13-27. Off-peak weekdays + all day weekends. Free, Pro, Max, Team all get it. Good time to push harder on whatever you're building.

Everyone blames AI agents failing on the model. Wrong. The bottleneck isn't intelligence -> it's defensive infrastructure. No recovery from API failures. No audit trails. No way to verify outputs.

Now my Claude Max Subscription got super value. Playing with it for a couple of hours now - I can feel more power!

My AI agent knows who I am. Not just "respond concisely" or "use this email." It knows my personality type, career situation, expertise split, what I'm dealing with right now. It adjusts to me.

Been using Ghostty as my main terminal. Started testing cmux this week. Sidebar with all your Claude Code sessions and their status. No more tabbing to find which agent needs attention. My actual setup: 4 sessions running in parallel. This is the agent flow I was missing.

Digital Thoughts hit 1,000 subscribers yesterday. Three years ago I posted about whether ChatGPT-4 could run a small business. The context window was 4,000 tokens: thoughts.jock.pl/p/1000-subs...

1000 subscribers. More about that and what next → tomorrow. For now → Thank you! Substack is a great place to be and create!

Codex with GPT 5.4 is a huge help in improving Wiz (my AI Agent) architecture.

Working with WizBoard for a while now. I really like what it is now.

Built 6 AI workflow playbooks for professionals who aren't developers. Legal, Sales, Marketing, HR, L&D, Office & Admin. Each has 6 built-in tools. Paste your real work data, get results. $79 each. No subscription needed. wiz.jock.pl/playbooks

Been too deep in the technical weeds — local LLMs are cool but admit it, they're niche. Next posts? All about Wiz, my AI Agent. Codex + Opus is a killer combo. GPT 5.4 just made it even better. Less theory. More *how we actually build this* 🚀