Super cool Kafka Visualization by Adam Warski and SoftwareMill : softwaremill.com/kafka-visualis…

Outbox pattern. Gunnar Morling 🌍 seems to be burnt by dual writes. For example, if you add two rows into different SQL tables in a single transaction, there are no atomic guarantees after Debezium CDC puts that into Kafka. The Outbox pattern fixes that by adding a new table.

+1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window

Fire all prompt engineers, hire context engineers instead.

Great crowd at AI Tinkers Warsaw today. Started with a surprise presentation from Filip Kozera about Wordware.

Great talk on how to optimize Cursor vibe coding by Grzegorz Kossakowski and Patryk Kabaj at AI Thinkers Warsaw. Tldr: Ask to prepare a plan in a prompt, answer clarifying questions, save the plan to markdown, and then kick off a big job.

Your life is not that complex. I mean in terms of tokens needed for the LLM model. Maciej Cielecki from 10Clouds at AI Thinkers Warsaw. Markdown files as a database is all you need.

👀

Early-stage startups are about a lot of trial and error. Most things do not perform well enough, but you still need energy to execute relentlessly. Though occasionally I discover something extraordinary. Started last Friday, got six customer interviews in the next two days.

Data is to validate the vision, not to create it.

Unless you have 10,000 customers: Approximately no one knows who you are. Approximately no one knows your brand. Approximately no one knows your product. You can change everything, any time. Take advantage of that.

Europe is back with a $200M Series A round for Lovable. Meanwhile, Polish Psyho, as the last natural neural network, beat OpenAI's artificial neural networks in a hardcore algorithm competition.

The last of us.

25+ years of dial-up speed was the constraint. Today, it's a lack of tokens or a provider outage.

Two customer discovery calls done, four scheduled from outreach and intros, plus a few commits powered by Claude Code. All during a business trip abroad today between meetings and a flight. The Zen of an early-stage startup founder.

Claude Code is amazing, but it can lie, it also made a deep architectural flaw that I have to recover myself from.

Unpopular opinion, but this is what I’d do if I started a new company today: •6-year vesting (companies take longer to build) •2-year cliff (most breakups happen in the first 2 years) •Founding engineers/growth members get 5%+ but on the same terms as founders

Podobno obrazki z 😺 i 🐶 przyciągają uwagę⁉️ 🚸Jeżeli tak, nie widziałem ostatnio lepszego wykresu, który pokazywałbym spadek zainteresowania rodzicielstwem 👇 Za XYZ

Will miss you o3, you served me well. Time for GPT-5.

Can GPT-oss:20b on my MacBook outperform GPT-5? Sometimes. Prompt: How many bs in blueberry? GPT-5: “3” 🤦 GPT-oss:20b: “2” ✅ (under 16 GB RAM) Why? GPT-5 is a mixture of models with a smart router. Sometimes the router underestimates a problem and picks the wrong expert.