Peter Gfader's brain noise

Dan North disagrees. As a software consultant :) Minute 34: https://youtu.be/6AxuSfSe4BA?si=b4goYEDg34WvFV5Q&t=2044

30 seconds later there is something interesting...

Dan argues he can tell (with in a 6month range) how long the person has been programming by looking at some source code.

At (0:37:19), Ron Minsky references an experiment conducted by Anthropic where they harnessed multiple AI agents together to write a C compiler.

Key takeaways from this story:

The Stunt: It was an impressive demonstration of using agents to build a complex software artifact, even though the resulting compiler had significant limitations—such as a lack of a type system and restricted capabilities (e.g., it could compile the Linux kernel but not many other programs).

Note: A compiler has a lot of tests in place to validate correct behaviour.

The Outcome: The experiment was eventually turned off because the agents reached a point where they stopped making forward progress.

The Lesson: As the agents continued to work, each new change became more likely to break existing functionality. This serves as a cautionary tale about using agentic systems in messy or poorly maintained codebases.

Because agents are trained to "win the game" (i.e., turn a red light green) without necessarily preserving the non-functional quality of the code, they can inadvertently create a state where further progress becomes impossible.

DELEGATE-52 is a new benchmark for evaluating how well LLMs handle delegated work — tasks where the model is left to execute on its own, like vibe coding. The premise: delegation requires trust that the model won't silently break things. To test this, the authors built long workflows across 52 professional domains (coding, crystallography, music notation, etc.) that require in-depth document editing, and ran 19 LLMs through them.

The headline result is unflattering across the board. Frontier models — Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4 — corrupt about 25% of document content by the end of long workflows on average. Smaller models fare worse. Giving the models agentic tool use doesn't help. Three factors make degradation worse: larger documents, longer interactions, and the presence of distractor files in the workspace.

Jim McNerney (Boeing CEO from 2005) coined the phrase that companies “overvalue experience and undervalue leadership.” Burford breaks this down: What is sold as leadership is, in reality, the managerial class’s desire to distance itself from experts so that no one can say what is right anymore. At Boeing, this resulted in hundreds of deaths—more than in the Oklahoma City bombing.

Full article https://docseuss.medium.com/the-biggest-threat-facing-your-team-whether-youre-a-game-developer-or-a-tech-founder-or-a-ceo-is-8cd1ad359508

Successful companies need two things: a product and customers.

Examples like FTX (Bankman-Fried), Theranos (Holmes), and the NFT hype show what happens when the “founder” myth and investor storytelling take the place of real value creation—bankruptcies, lawsuits, prison sentences. NFTs failed because simply no one wanted the product; artificially inflated markets without customers inevitably collapse.

True leadership is an action, not a status.

Whenever the AI needs to discover the rules, controls and goal while playing/working. -> You need human intelligence.

Solving Simple Games :)

- from ARC-AGI-3

The key point in all three games: the focus isn’t on whether a goal is achieved, but on how efficiently—that is, how many actions are required compared to humans. Intelligence is efficiency. ARC Prize The difference from ARC-AGI-1/2 is fundamental: instead of static puzzles, agents must explore, learn, set intermediate goals on their own, and adapt their strategy—without any guidance. ARC Prize Humans enjoy the process. AI systems click around blindly hundreds of thousands of times.

Stand-ups. Story points. Sprint reviews. YOUR BOARD DOESN'T CARE ABOUT YOUR VELOCITY.

NO-THEATER is the elite impact engine that translates your team's real work into the language executives actually act on. No jargon. No ceremony. Just proof of value.

https://notheater.app/

Amazing tips!

Summary:

What are user interviews

How to plan and conduct user interviews sessions

Focus on the content of an interview guide

Tips to write good and effective interview questions

Using my User Interview Cards to plan and conduct interviews more effectively

Facilitating user interview sessions like experts

Tips on how to analyze the data.

According to Cory Doctorow in April 2024 https://pluralistic.net/2024/04/01/human-in-the-loop/

The whole AI bubble is based on convincing investors that one or more of the following is true: I. There are low-stakes, high-value tasks that will recoup the massive costs of AI training and operation; II. There are high-stakes, high-value tasks that can be made cheaper by adding an AI to a human operator; III. Adding more training data to an AI will make it stop hallucinating, so that it can take over high-stakes, high-value tasks without a "human in the loop."

Software engineering was NEVER a good candidate for me.

And now Cal Newport's reasoning as per 2026-02-25: https://www.youtube.com/watch?v=Ijt8lV6b7QY

From Milan https://newsletter.techworld-with-milan.com/p/learn-fundamentals-not-frameworks

If you’ve been around this industry long enough, you’ll have witnessed many a framework come and go. Ever heard of frameworks like GWT, Apache Tapestry, Backbone.js, or Knockout.js? Yes, they were very popular back in the day, and you needed to read a lot of docs to understand them.

Analysis done by Stackoverflow: https://stackoverflow.blog/2018/01/11/brutal-lifecycle-javascript-frameworks/

There’s even a hierarchy of technology longevity: language choices last 10+ years, databases 7–10 years, backend frameworks 3–5 years, and frontend frameworks around 18 months. If you’re debating all of those equally, you’re just wasting time.

The data on the code half-life also confirms this. Erik Bernhardsson has analyzed 26+ open-source projects and calculated a median half-life of 3.33 years. Frontend frameworks such as Angular are at the very short end of this scale, at just 0.32 years, whereas Linux has a figure of 6.6 years. Again, the further up the stack you go, the shorter the lifespan becomes.

% of commits still present in the code base over time (Source: Erik Bernhardsson blog) -> a lot of projects in aggregate (also sampled somewhat arbitrarily):

«Es kann nicht sein, dass ein Digitalprojekt, welches über 10 Jahre lang entwickelt wurde, beim ‹Go Live› wieder einmal mehr scheitert»,

English: “It is unacceptable that a digital project that has been in development for over 10 years should fail once again when it goes live.”

Statement from a responsible Person in the department SECO. https://www.srf.ch/news/schweiz/verspaetete-arbeitslosengelder-kaum-selbstkritik-nach-it-panne-im-seco Minute 2:00

Time pressure

Old programs

Many interfaces

Predecessors (as in people)

What happened?

On January 6, 2026, a new IT system for processing and paying unemployment benefits (ASAL 2.0 -> AsalFutur) was launched at the State Secretariat for Economic Affairs (SECO). This is intended to replace the old software from the 1980s.

Technical problems during the introduction have led to significant delays in the payment of daily allowances. Many unemployed people in Switzerland are therefore affected.

References

https://www.srf.ch/news/schweiz/verspaetete-arbeitslosengelder-kaum-selbstkritik-nach-it-panne-im-seco

Tea, officially Tea Dating Advice, is a dating surveillance mobile phone application that allows women to post personal data about men they are interested in or are currently dating.

🤯🤯🤯 3 major data leaks in July and August 2025. 🤯🤯🤯

while some public discourse on social media around Tea's breach includes criticisms of vibe coding, some security experts said they doubted the platform itself used AI to generate its code.

https://www.businessinsider.com/tea-app-data-breach-cybersecurity-ai-vibe-coding-safety-experts-2025-8

Cause of the breach

The core issue was an exposed Firebase / cloud database with no proper authentication or access controls, effectively leaving backend storage publicly reachable.​

Reporting has highlighted Tea’s reliance on unreviewed AI‑generated “vibe coding” for backend logic, which contributed to insecure configuration and lack of basic security hardening.

https://www.theserverside.com/blog/Coffee-Talk-Java-News-Stories-and-Opinions/How-not-to-spill-the-Tea-A-real-world-app-security-lesson

"Enshittification" 4th phase. Context = Lifecycle of Digital Platforms.

PS: Die ersten 3 Phasen wurden von Cory Doctorow im 2022 erfunden. "It explains why services like Google, Amazon, and Facebook, which once seemed miraculous and customer-centric, now feel cluttered, expensive, and frustrating."

Good to Users

Good to Business Customers (advertisers, marketplace sellers, or publishers)

Good to Shareholders

💡4. Good To Executives + Insiders

they are cannibalizing shareholders themselves via financial engineering and hype cycles.

the 4th phase is about manufacturing value through financial smoke and mirrors

via Ed Zitron https://www.wheresyoured.at/the-enshittifinancial-crisis/

The "Chain of Pain" (The 4th Stage Collapse)

Zitron argues that this cycle is a "subprime AI crisis" because the underlying assets (GPUs and LLMs) don't currently generate enough "real-world" revenue to pay back the $100B+ in debt being amassed.

Stage 4 Enshittification: Management teams are now "cannibalizing shareholders" by spending billions on this loop to keep stock prices high, even if the AI products themselves are losing money.

The "Off-Balance Sheet" Trick: Much like Enron, companies are using complex financial structures to hide how much they are actually spending on AI, making the business look healthier to investors than it actually is.

The "Circular" Revenue: When Microsoft "invests" in OpenAI, and OpenAI spends that money on Microsoft Azure, it looks like Microsoft's "AI Revenue" is growing. In reality, Microsoft is essentially paying itself to create the appearance of growth.

Why this matters to you:

Exposure hours. The number of hours each team member is exposed directly to real users interacting with the team’s designs or the team’s competitor’s designs. There is a direct correlation between this exposure and the improvements we see in the designs that team produces.

People tend to say "No".

Figure via Andrew Ng https://www.deeplearning.ai/the-batch/issue-330/

Examples like:

Ethics - How to use AI tools ethically in software development

Sustainability - How to reduce environmental impact of AI tools

Security - How to use AI tools securely

The department for Environment Food and Rural Affaris has some good guidelines: https://defra.github.io/defra-ai-sdlc/pages/responsible-practices/

Stay accountable

You remain responsible for all outputs, even when AI tools help create them.

Do:

Review all AI-generated outputs before you use it

Document AI tool usage in critical systems

Explain AI-assisted decisions to stakeholders

Don’t:

Accept AI outputs without checking them

Don’t rely on AI when you can’t explain the ‘why’ behind important decisions

Prevent bias

AI tools can copy existing biases, especially from training data.

Do:

Test AI-generated outputs with different scenarios and a diverse set of users (if relevant)

Check outputs for potential bias in how they work or what they say

Consider who might be left out by AI-assisted analysis and solutions

Don’t:

Assume AI outputs are neutral

Put AI-generated algorithms into production without bias testing

Protect data

AI tools may process information in ways that create privacy risks.

Do:

Check what data AI tools collect and how they use it

Follow GDPR requirements when you use AI tools

Use privacy-protecting alternatives for personal data

Turn on privacy settings in all tools

For information on the appropriate privacy settings for your chosen AI tool, see the Defra Tools Guidance

Don’t:

Put sensitive or classified data into public AI tools

Use AI tools without understanding how they handle data

Decision framework

Before you use AI tools, ask:

Does this serve a clear user need?

Can you explain how you’re using AI?

Do you accept responsibility for outputs?

Have you considered potential bias?

Does this protect user data properly?

If you cannot answer “yes” to all questions, change your approach.

Things to consider

Security and reliability

AI-generated outputs may contain security flaws or errors that look correct. Always check AI outputs through testing and reviewed.

Intellectual property

AI tools may generate outputs based on copyrighted material. Make sure AI-assisted outputs follows open source licensing requirements and government IP policies.

Environmental impact

AI tools use computing resources. Consider environmental costs and use AI tools in proportion to the task complexity. See Sustainability for more information.

Skills and development

Use AI tools to enhance rather than replace human expertise.

Use AI to enhance your skills and critical thinking rather than replace them

Public trust

Wilco Burggraaf makes the point that long lived sustainable software needs care and proper design. The real challenge lies not in writing code quickly, but in building stable, maintainable, and sustainable software. AI can do 80% of the work, but the crucial 20%—robustness, security, performance, maintenance—requires human experience, architectural thinking, and discipline.

Because structure is the thing AI still doesn’t grasp. It imitates, but it doesn’t architect. It decorates, but it doesn’t frame the house. And when you’re building something that has to last more than a few days, a few demos, you start seeing the cracks.

scale reveals everything. The shortcuts, the brittle glue code, the missing tests, the code that never should have been merged, but was. Because the prototype looked good in demo day lighting. But nobody asked what happens after. Nobody asked who owns the code, or how it evolves, or how it breaks. And that’s the part AI doesn’t cover.