Discussion – Summary (over-simplified) Camp 1: Vibe Coding Over-Enthusiasm ▪ Prompts / Specs are the new Code ▪ Code is the new Assembler („forget, the code even exists“) ▪ == no SWE expertise required anymore! Camp 2: Categorical Rejection ▪ = It’s not intelligent, ▪ It’s only statistics ▪ It’s only better autocomplete ▪ It’s only a stock market bubble ▪ == Ieave me the *** alone! „It’s only…“ Guys // Status quo can stay - nothing has to change
„philosophical“? Questions arise: ▪ What is SWE / knowledge work about? ▪ What is intelligence? ▪ Does Language == Intelligence? ▪ How does Education work? ▪ == how do we build up intelligence? ▪ What defines „Being Human?“ ▪ … We’re not trying to answer them directly, today! ▪ All I’m saying is that…
to get „philosophical“? ▪ One way that worked for me: ▪ Look at past attempts to replace coders with technology ▪ Ask: why are we still here today? ▪ => Hey, now you understood sth. more about SWE! ▪ Then: back to AI! Same old story again? Or is it new? ▪ Spoiler: It’s both (no easy answers today, sorry!)
close to business language ▪ Did its creators aim to replace developers? ▪ Probably not! ▪ But it was part of broader data processing hype ▪ And some say it sparked such phantasies (on management levels) ▪ Early example of: “Make syntax more natural to make programming more accessible“ The Dream of Business Readable Syntax
close to business language ▪ Did its creators aim to replace developers? ▪ Probably not! ▪ But it was part of broader data processing hype ▪ And some say it sparked such phantasies (on management levels) ▪ Early example of: “Make syntax more natural to make programming more accessible“ The Dream of Business Readable Syntax If this was supposed to be “natural language”, people of the 60s had a weird way of talking!
The Dream of Democratizing SWE: Should I mention Access Or BPMN? Or RAD (Rapid Application Development) / Visual building? Model Driven Development? 4GL = 4th Generation Language …? // Great tool for power users! … until it stops being a great tool and starts being a nightmare // I love BPMN! … as a communication tool to talk with domain experts. But I’ve never heard of projects where domain experts create perfect diagrams which the engines can directly execute. And if I would find those domain experts… I’d probably hire them as programmers.
only the tip of the iceberg For example: ▪ Nobody (in their right mind) reads this book to learn all algorithms by heart ▪ We don’t “learn algorithms” ▪ We want to learn algorithmic thinking!
only the tip of the iceberg For example: ▪ When learning how to code, it’s not about remembering syntax. ▪ It’s about training our neural network (between our ears) so that we can creatively solve new problems
SWE really about? Disclaimer: This is my sloppy approach Other people spent more effort: ▪ Alex Thurow (https://onmoderndev.de/) ▪ Don Gray https://mstdn.social/@alexthurow/116279276457252836
the No-Code Trap // Output // Outcome ▪ Most Non-Tech People are missing many “iceberg skills” ▪ Vibe Coding Tools give them Output Generators ▪ But without the bridge required to turn this to desired outcomes ▪ They only get so far – and then bugs take over, they get hacked, etc. ▪ And need experts with all the iceberg skills to clean up the mess ▪ This is the No Code Trap 2.0 ▪ BUT…
Paper, Google => Transformer Architecture 2017 BERT A quantum leap for embeddings / NLP (Natural Language Processing) 2018 “The Year of Metabolism”? Tool quality lags behind the intelligence of the models => Industry must “Metabolize” this new technology 2026 “Year of Agents“ 2025 11.03.2026 / Florian Sommer Nov 2022 ChatGPT 3.5 „Wow“ Moment Mainstream 2021 GitHub Copilot Preview (Codex Model) 2020 OpenAI GPT 3 1. It’s scalable!? 2. Emerging Capabilities: Solving unknown problems without re- training! („in-context learning“) It’s easy to overlook, with all the hype going on…
Paper, Google => Transformer Architecture 2017 BERT A quantum leap for embeddings / NLP (Natural Language Processing) 2018 2020 “The Year of Metabolism”? Tool quality lags behind the intelligence of the models => Industry must “Metabolize” this new technology 2026 “Year of Agents” 2025 11.03.2026 / Florian Sommer Nov 2022 ChatGPT 3.5 „Wow“ Moment Mainstream OpenAI GPT 3 1. It’s scalable!? 2. Emerging Capabilities: Solving unknown problems without re- training! („in-context learning“) 2021 GitHub Copilot Preview (Codex Model) … but that’s where it all really started.
to swallow pill … ▪ … BUT: ▪ The technological breakthrough with LLMs was: “in-context learning” ▪ Those things can “generalize” to some extent ▪ They don’t just “cite” training knowledge ▪ They apply learned patterns probabilistically to new problems we give them in their context window ▪ When this works: We might call this “effectively intelligent” – because the effect of solving a problem is real. This is not “No Code 2.0”! ▪ When it doesn’t: We call it a “hallucination”
// Output // Outcome 1. Only the combination of “world knowledge” and generalization makes it eat more of the iceberg ▪ Which requires big models ▪ = which is expensive, complex, … ▪ => handling & optimizing this requires technical understanding 2. It’s still probabilistic output – no true understanding. ▪ You must check the result and judge, whether it really solves your problem 3. It’s no lifelong learner. Memory Files do not count. The left part of the iceberg becomes more important to us!
and technical understanding // Generating AI Output without even trying to achieve outcomes – some call this “AI Slop” … I’d call it: AI becoming the “Enshittification Catalyst”
siroemk
marcoroth
minodriven
optim
drumath2237
seike460
takai
sylph01
findyinc
nabondance
satoshi256kbyte
addyosmani
aki_iinuma
samanthasiow
bcantrill
baasie
stephaniewalter
minecr
andyhume
mayatellez
watany
marktimemedia
