Jakub Švec

🤖 Is the #AI technological leap over?

Content:
1) ↘️ #Yes, if we consider the following
2) ↗️ #No, if we consider the following
3) ⚡ #My guess

topics:
#llm #future #agi #tech #dev #software #code #programming #society #technology #aicannibalism #data #transformer #finance

More training data for LLM does not necessarily mean that it will produce better results. In the future, we may have a problem with AI cannibalism. This can paralyze products that depend on generated code.

🧵12/12

At first glance, AI is a very appealing solution to basically everything, but beneath the surface lies a large number of questions and problems. Digital gambling has already begun, hold on to your hats!

🧵11/12

REALITY👻
Vibe coding is here. We read celebratory articles that people are no longer needed and at the same time, we see many examples of these issues or code decay. Not only software engineering but all societies will face very difficult trials.

🧵10/12

MY ADVICE🦾
Use LLM as an advisor and a teammate, not as an autonomous builder (at least for now). Cooperate with AI. Use it in chat or your IDE—it is a helpful coworker. Use agents with a limited scope. Let AI review your code. But do not let the LLM be the decision-maker in your codebase.

🧵9/12

Source code is read far more often than it is written. Understanding written logic is crucial, does AI really deeply understand your tendencies and reasons? In my experience, far from always. A big issue.

🧵8/12

3) OWNERSHIP⚙️
Who will understand it? People will stop understanding what is actually happening in your production code. This outcome is inevitable. The result is garbage code that nobody understands with potential security issues. Code that is maintainable only by an AI.

🧵7/12

It is fine for an AI provider, when you have redundant code. You will pay output tokens and then input tokens, when you need to repair or refactor the generated code. A snowball effect can occur fast. Vibe coding is a short-term win, but definitely not a long-term one. A big issue.

🧵6/12

Your code might work at the end of the day, but technical debt will grow silently. Have you also heard of AI cannibalism? The LLM quality can degrade over time.

Vibe coding forces you to blindly trust the AI. LLMs can generate millions of lines of code in a week, but what next?

🧵5/12

An AI-generated 50 lines can cause issues. Vibe coding generates hundreds or thousands of lines and you have no idea what happens on each line. A vibe-coded script might be 80% correct, but production isn't about "80%", right?

🧵4/12

2) GARBAGE CODE🚮
Vibe coding can be helpful when starting a project from scratch and you need an MVP fast. However, iterating over existing code can be very tricky. You risk ending up with redundant, "copy-pasted" lines instead of keeping core logic centralized.

🧵3/12

Any LLM can produce vulnerabilities, and some models even generate them by default. Which ones? Who knows...

Human developers can also make mistakes, but on a completely different scale. A big issue.

🧵2/12

1) SECURITY🛡️
Vibe coding is a security risk. There is a long list of dangerous issues that must be properly handled. Also if the LLM is not isolated, your data and sensitive information can leak. There is much more.

Vibe coding? No, I prefer maintainable code. Here’s why:

Content:

1) SECURITY🛡️
2) GARBAGE CODE🚮
3) OWNERSHIP⚙️
MY ADVICE🦾
REALITY👻

Thread:
🧵1/12

Topic:
#ai #llm #vibecoding #dev #software #security #code #society

/image by imagen4

YouTube video by Bruce Springsteen Bruce Springsteen - Streets Of Minneapolis (Official Audio)

I wrote this song on Saturday, recorded it yesterday and released it to you today in response to the state terror being visited on the city of Minneapolis. It’s dedicated to the people of Minneapolis, our innocent immigrant neighbors and in memory of Alex Pretti and Renee Good.

Stay free

Line graph titled "The top five programming languages on GitHub in 2025." It shows the trends of users of Python, TypeScript, JavaScript, Java, and C# from January 2023 to July 2025. Each language is represented by a color, with TypeScript and Python coming in first and second, respectively.

Chart titled "Tracking the top developer populations on GitHub (2020-2025)." It lists 10 countries in order: United States, India, China, Brazil, United Kingdom, Japan, Germany, Indonesia, Russia, and Canada. Each row shows the country's five-year trend, 2020 to 2025 developer count, and five-year compound annual growth rate percentage.

Graph showing the total number of AI-related open source and public projects on GitHub, with total repos increasing from slightly above 500,000 in 2020 to over 4,000,000 after 2024.

Graph titled "Growth in contributors to generative AI projects." It displays the number of distinct contributors from January 2024 to July 2025, which has fluctuated between around 70,000 and over 200,000.

1.12 billion contributions. Over 180 million developers. This year, GitHub saw record growth, surging uses in AI, and significant shifts in programming languages.

What other stories do the numbers tell? Let’s dive into the data. This is Octoverse. 👇
https://gh.io/AAy5upr

O Babišovi stačí vědět jedno: nepřišel by na pomoc napadenému Polsku 🇵🇱
#volby2025 🇨🇿

The faces of those killed in Russia’s strike on Kyiv on August 28 are now known.

Children. Teenagers. Mothers. The elderly. Just Ukrainians. They are gone. Burned alive. Crushed under rubble. Suffocated in smoke. Russia killed them.

How many more?

What Is Python's __init__.py For? – Real Python Learn to declare packages with Python's __init__.py, set package variables, simplify imports, and understand what happens if this module is missing.

🐍📰 What Is Python's __init__.py For? realpython.com/python-init-...

Data Talk #143: Jakub Švec (Carl Data Company) Data Talk · Episode

🚀In this episode, #we discuss not only prompts and #AI models but also on the engineering complexities that necessarily accompany such implementation.
#multimodal #llm #gemini #python 🐍
🔊 open.spotify.com/episode/0RyI...

🧵(2/2)
#AI models are currently a pure extract of human symbolic worlds, and it will be interesting if they ever manage to leave them - into the non-symbolic layer.

/image by symbolic model Imagen 4

🧵(1/2)
Eight years ago, I wrote a thesis at university on the #human ability to cognitively switch between the #symbolic layer (the flow of thoughts) and the #non-symbolic layer (the absence of the flow of thoughts).

Pointers in Python: What's the Point? In this step-by-step tutorial, you'll get a clearer understanding of Python's object model and learn why pointers don't really exist in Python. You'll also cover ways to simulate pointers in Python without the memory-management nightmare.

🐍📰 Pointers in Python: What's the Point?

In this step-by-step tutorial, you'll get a clearer understanding of Python's object model and learn why pointers don't really exist in Python.
#python

🦾 In the long term, however, you primarily need #data #engineering skills, because most of the time, you will be dealing with production optimization and scalability. 🚀 (6/6)

/image by #imagen3

🤖 This certainly does not mean that prompt engineering skills are not needed—they definitely are. Especially in the beginning, when #testing whether a given #AI model can handle the given tasks. (5/6)

This primarily means ensuring #scalability, #reliability, #parallel data processing, data #quality, #validation of output data, and more. (4/6)

⚙️ Even though we have about 30 pages of prompts in Carl Data Company that need to be continuously tested, improved, and developed, the much bigger #challenge is ultimately keeping the production code in optimal condition. (3/6)

📽️ For the past year, we have been working intensively on a repository for #video and #photo analysis. Most of the time, we rely far more on #data engineering skills than on prompt engineering skills. (2/6)

🚀 #What is the main #skill you will need if you want to implement a #multimodal LLM? It is definitely not #prompt engineering... 🧵(1/6)

🤖For several months now, we have been #developing a robust #prompt at Carl Data Company that, thanks to its connection to #Gemini, can analyze more than 40 categories for each #video or #photo.

🧾Current status: More than 30 pages of production prompt

#ai #llm #multimodal #python
#image by #imagen3