Geek Coder
@geekcoder.bsky.social
My name is Manfred. I'm a software craftsman and keen learner, covering software engineering topics. While I use AI-based tools, I'm not crazy enough to bet my career on them. They are useful but still don't replace thinking and need to be supervised.
Node.js version 24.11.0 is now LTS, see nodejs.org/en/download #nodejs
October 30, 2025 at 11:57 AM
Node.js version 24.11.0 is now LTS, see nodejs.org/en/download #nodejs
"Create an an image of an analog watch showing the time 1:25 pm". This is what ChatGPT 5 created. #FunWithAI #LLM #AI
August 13, 2025 at 1:53 PM
"Create an an image of an analog watch showing the time 1:25 pm". This is what ChatGPT 5 created. #FunWithAI #LLM #AI
Impressive PhD-level math skills demonstrated by ChatGPT 5. Have a look and enjoy!!
Honestly: I think the statement that ChatGPT 5 is at "PhD-level" is an insult to all people with a PhD.
#FunWithAI #ChatGPT #LLM
Honestly: I think the statement that ChatGPT 5 is at "PhD-level" is an insult to all people with a PhD.
#FunWithAI #ChatGPT #LLM
August 13, 2025 at 2:09 AM
Impressive PhD-level math skills demonstrated by ChatGPT 5. Have a look and enjoy!!
Honestly: I think the statement that ChatGPT 5 is at "PhD-level" is an insult to all people with a PhD.
#FunWithAI #ChatGPT #LLM
Honestly: I think the statement that ChatGPT 5 is at "PhD-level" is an insult to all people with a PhD.
#FunWithAI #ChatGPT #LLM
Github Copilot getting bored and starts making random suggestions that make no sense, see the last line in this screenshot.
I think Copilot has mood swings as well. On some days it works quite well when you spoon feed it. On other days, e.g. today, the error rate is closer to 80%.
#FunWithAI #LLM
I think Copilot has mood swings as well. On some days it works quite well when you spoon feed it. On other days, e.g. today, the error rate is closer to 80%.
#FunWithAI #LLM
August 10, 2025 at 7:55 AM
Github Copilot getting bored and starts making random suggestions that make no sense, see the last line in this screenshot.
I think Copilot has mood swings as well. On some days it works quite well when you spoon feed it. On other days, e.g. today, the error rate is closer to 80%.
#FunWithAI #LLM
I think Copilot has mood swings as well. On some days it works quite well when you spoon feed it. On other days, e.g. today, the error rate is closer to 80%.
#FunWithAI #LLM
This is the "excellent" code quality produced by an AI agent. A human coder would have expressed this in a single switch()-statement which would also be easier to read and understand.
Looking at this, I'm wondering how bad is the training data really?
#FunWithAI #llm #Agents
Looking at this, I'm wondering how bad is the training data really?
#FunWithAI #llm #Agents
July 31, 2025 at 10:00 AM
This is the "excellent" code quality produced by an AI agent. A human coder would have expressed this in a single switch()-statement which would also be easier to read and understand.
Looking at this, I'm wondering how bad is the training data really?
#FunWithAI #llm #Agents
Looking at this, I'm wondering how bad is the training data really?
#FunWithAI #llm #Agents
An example of a graph that does show the actual quotes at the time. However, the graph isn't really useful this way. (Source: finance.yahoo.com, as of 25 Jul 2025 at end of after-hours trading). One option to improve it would be an optional filter to catch spikes, up and down. #BugCollection
July 26, 2025 at 1:33 AM
An example of a graph that does show the actual quotes at the time. However, the graph isn't really useful this way. (Source: finance.yahoo.com, as of 25 Jul 2025 at end of after-hours trading). One option to improve it would be an optional filter to catch spikes, up and down. #BugCollection
What a useful message! 😂
(Source: Left feedback on a VS Code Documentation page.)
(Source: Left feedback on a VS Code Documentation page.)
July 25, 2025 at 9:06 PM
What a useful message! 😂
(Source: Left feedback on a VS Code Documentation page.)
(Source: Left feedback on a VS Code Documentation page.)
I just removed 5 empty lines in a file to improve the code layout. Zero functional changes. Commit message suggestion by Copilot: "refactor AccountCommandService to use saveAccount method for persisting domain events and index accounts with SearchServiceAdapter" - #TheRealityOfAI
May 4, 2025 at 6:00 AM
I just removed 5 empty lines in a file to improve the code layout. Zero functional changes. Commit message suggestion by Copilot: "refactor AccountCommandService to use saveAccount method for persisting domain events and index accounts with SearchServiceAdapter" - #TheRealityOfAI
Tab programming at its finest. AI in a death spiral ..... I'm observing this multiple times a week. Sad! #TheRealityOfAI
April 27, 2025 at 9:17 AM
Tab programming at its finest. AI in a death spiral ..... I'm observing this multiple times a week. Sad! #TheRealityOfAI
How stupid: suggesting a method that is already there, even if it's the minimum. The AI should be aware there is a method of that name already and at best suggest changing it's implementation. #TheRealityOfAI
April 27, 2025 at 9:00 AM
How stupid: suggesting a method that is already there, even if it's the minimum. The AI should be aware there is a method of that name already and at best suggest changing it's implementation. #TheRealityOfAI
AI getting into the way with TDD: AI writes a test for the existing (minimal!) implementation which just throws an exception. This test would cast this in concrete instead of driving the correct implementation. This demonstrates that the AI doesn't "understand" the concept of TDD. #TheRealityOfAI
April 27, 2025 at 5:19 AM
AI getting into the way with TDD: AI writes a test for the existing (minimal!) implementation which just throws an exception. This test would cast this in concrete instead of driving the correct implementation. This demonstrates that the AI doesn't "understand" the concept of TDD. #TheRealityOfAI
The "loop of death" increasingly becomes a common theme. Typically observed on interfaces, it also struggles to generate correct code for a constructor, despite that constructor being in the code based and imported. It just keeps making stuff up. (Copilot / GPT-4o / Agent Mode). #TheRealityOfAI
April 20, 2025 at 11:05 PM
The "loop of death" increasingly becomes a common theme. Typically observed on interfaces, it also struggles to generate correct code for a constructor, despite that constructor being in the code based and imported. It just keeps making stuff up. (Copilot / GPT-4o / Agent Mode). #TheRealityOfAI
Only one way to make a worse suggestion: A function that generates purely random tokens. Here it suggests a non-existing identifier named "mockAlpaca" ... I can add it but expected the generated code to be correct in the first place..
AI forces me to clean up after it ... again.
#TheRealityOfAI
AI forces me to clean up after it ... again.
#TheRealityOfAI
April 19, 2025 at 9:56 PM
Only one way to make a worse suggestion: A function that generates purely random tokens. Here it suggests a non-existing identifier named "mockAlpaca" ... I can add it but expected the generated code to be correct in the first place..
AI forces me to clean up after it ... again.
#TheRealityOfAI
AI forces me to clean up after it ... again.
#TheRealityOfAI
Again, why would the AI suggest to call a non-existent method on a type? It has all the information it needs. A classic auto-suggest, auto-complete implementation easily does better.
I'd love for this work better as I want to make the tool work. It's just not that simple, though.
#TheRealityOfAI
I'd love for this work better as I want to make the tool work. It's just not that simple, though.
#TheRealityOfAI
April 19, 2025 at 9:27 PM
Again, why would the AI suggest to call a non-existent method on a type? It has all the information it needs. A classic auto-suggest, auto-complete implementation easily does better.
I'd love for this work better as I want to make the tool work. It's just not that simple, though.
#TheRealityOfAI
I'd love for this work better as I want to make the tool work. It's just not that simple, though.
#TheRealityOfAI
Why would it suggest a return type in a typed language if all other functions in the code base have a return type?
Why doesn't it suggest to return a value if all other functions do that?
AI still has a long way ahead of it. #TheRealityOfAI
Why doesn't it suggest to return a value if all other functions do that?
AI still has a long way ahead of it. #TheRealityOfAI
April 19, 2025 at 9:19 PM
Why would it suggest a return type in a typed language if all other functions in the code base have a return type?
Why doesn't it suggest to return a value if all other functions do that?
AI still has a long way ahead of it. #TheRealityOfAI
Why doesn't it suggest to return a value if all other functions do that?
AI still has a long way ahead of it. #TheRealityOfAI
Mind you: The AI should "know" that I'm trying to update the account so the command should be "new AccountUpdateCommand()" which is already in the code base. The AI has no deeper understanding. Just probabilities of what could be the next token. #TheRealityOfAI
April 19, 2025 at 1:22 AM
Mind you: The AI should "know" that I'm trying to update the account so the command should be "new AccountUpdateCommand()" which is already in the code base. The AI has no deeper understanding. Just probabilities of what could be the next token. #TheRealityOfAI
Spot the error .... The type parameter for the generic function call should be ".get()" not ".get()". As long as it gets simple things wrong like this, I am not concerned. Progress on the quality side has been flat in the last 6 months. #TheRealityOfAI
April 18, 2025 at 11:55 PM
Spot the error .... The type parameter for the generic function call should be ".get()" not ".get()". As long as it gets simple things wrong like this, I am not concerned. Progress on the quality side has been flat in the last 6 months. #TheRealityOfAI
Why would it suggest an assertion for it being "defined" if the return type of create() rules out "undefined"? Again, the AI does not "understand" the concept of a type. As a result it makes a suggestion that makes no sense. #TheRealityOfAI
April 18, 2025 at 11:33 PM
Why would it suggest an assertion for it being "defined" if the return type of create() rules out "undefined"? Again, the AI does not "understand" the concept of a type. As a result it makes a suggestion that makes no sense. #TheRealityOfAI
It definitely needs supervision. Here is another example of an endless suggestion loop produced by Copilot / GPT-4o / Agent mode. Attention span of a gold fish. Forgets what it wrote 6 lines earlier. I'm not convinced this technology will replace software engineers in 6 months time. #FunWithAI
April 18, 2025 at 8:41 PM
It definitely needs supervision. Here is another example of an endless suggestion loop produced by Copilot / GPT-4o / Agent mode. Attention span of a gold fish. Forgets what it wrote 6 lines earlier. I'm not convinced this technology will replace software engineers in 6 months time. #FunWithAI
And here is the AI being stupid again. The method clearly states "createAccount", HTTP method is POST, yet it generated code including "ChartCreateCommand()". This is hindering, not helping. #FunWithAI
April 13, 2025 at 5:04 AM
And here is the AI being stupid again. The method clearly states "createAccount", HTTP method is POST, yet it generated code including "ChartCreateCommand()". This is hindering, not helping. #FunWithAI
An example of Copilot giving a good answer. I spotted the problem, then highlighted the code, asked the question, including a request for an explanation. Copilot with o3-mini provided the correct answer. #HelpfulAIResponse
April 12, 2025 at 8:39 AM
An example of Copilot giving a good answer. I spotted the problem, then highlighted the code, asked the question, including a request for an explanation. Copilot with o3-mini provided the correct answer. #HelpfulAIResponse
More fun with AI: Here a snippet from a test generated by AI. As far as I am concerned lines 28 and 29 are not required. If either 28 or 29 fail, then 30 fill fail as well. Therefore lines 28 and 29 are superfluous. I'm wondering what the code looks like that was used for the training ... #FunWithAI
April 11, 2025 at 3:26 AM
More fun with AI: Here a snippet from a test generated by AI. As far as I am concerned lines 28 and 29 are not required. If either 28 or 29 fail, then 30 fill fail as well. Therefore lines 28 and 29 are superfluous. I'm wondering what the code looks like that was used for the training ... #FunWithAI
An example where a classic tool for auto-suggest and auto-complete at least "understands" that the constructor expects 4 arguments. Meanwhile the AI suggests random 6 arguments, none of which match the ones that are expected.
The out-dated tool beats the AI! 😂
#FunWithAI
The out-dated tool beats the AI! 😂
#FunWithAI
April 11, 2025 at 3:23 AM
An example where a classic tool for auto-suggest and auto-complete at least "understands" that the constructor expects 4 arguments. Meanwhile the AI suggests random 6 arguments, none of which match the ones that are expected.
The out-dated tool beats the AI! 😂
#FunWithAI
The out-dated tool beats the AI! 😂
#FunWithAI
... and it just keeps generating silly code ... 😂 With that kind of code it is easy to become a 10x developer! Without AI it would take me 10x as long to create technical debt. #FunWithAI
April 9, 2025 at 12:03 AM
... and it just keeps generating silly code ... 😂 With that kind of code it is easy to become a 10x developer! Without AI it would take me 10x as long to create technical debt. #FunWithAI
AI suggests code that gets a Types.ICommandService and assumes that it implements the IChartQueryService interface, which is not the case.
The AI has no deeper understanding of how a DI container works. AI - as least for the time being - requires constant supervision.
#FunWithAI
The AI has no deeper understanding of how a DI container works. AI - as least for the time being - requires constant supervision.
#FunWithAI
April 8, 2025 at 11:28 PM
AI suggests code that gets a Types.ICommandService and assumes that it implements the IChartQueryService interface, which is not the case.
The AI has no deeper understanding of how a DI container works. AI - as least for the time being - requires constant supervision.
#FunWithAI
The AI has no deeper understanding of how a DI container works. AI - as least for the time being - requires constant supervision.
#FunWithAI