Peter Lawrey
@peterlawrey.bsky.social
Java Champion | Vanilla Java Blog (6M views) | CEO of Chronicle Software with 8 out the top 11 investment banks as clients.
Six kids from 3 to 26
Six kids from 3 to 26
Visiting Zegreb for some consulting
September 30, 2025 at 5:23 AM
Visiting Zegreb for some consulting
The war no one talks about
September 22, 2025 at 11:17 AM
The war no one talks about
It turns out that Reasoning LLM can also get distracted by cats.
"Cats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models"
arxiv.org/abs/2503.01781
"Cats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models"
arxiv.org/abs/2503.01781
September 21, 2025 at 7:15 PM
It turns out that Reasoning LLM can also get distracted by cats.
"Cats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models"
arxiv.org/abs/2503.01781
"Cats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models"
arxiv.org/abs/2503.01781
While Generative AI can increase the overall volume of documentation and code, in terms of curated and validated release content, the increase might only be +20%. The following represents the number of lines of ~6 months of work on similar projects, both before and after using AI.
September 19, 2025 at 12:46 PM
While Generative AI can increase the overall volume of documentation and code, in terms of curated and validated release content, the increase might only be +20%. The following represents the number of lines of ~6 months of work on similar projects, both before and after using AI.
GPT-5: Generate a Where's Waldo, and find Waldo
August 17, 2025 at 3:33 PM
GPT-5: Generate a Where's Waldo, and find Waldo
I find most people want AI to produce a correct answer, which means the AI needs to be cautious.
I look for ideas that will inspire me, something I wouldn't have thought of, cherry-picking from a selection of more "creative" ideas
I look for ideas that will inspire me, something I wouldn't have thought of, cherry-picking from a selection of more "creative" ideas
July 20, 2025 at 8:19 AM
I find most people want AI to produce a correct answer, which means the AI needs to be cautious.
I look for ideas that will inspire me, something I wouldn't have thought of, cherry-picking from a selection of more "creative" ideas
I look for ideas that will inspire me, something I wouldn't have thought of, cherry-picking from a selection of more "creative" ideas
Genimi 2.5 pro was more blunt
July 16, 2025 at 9:39 AM
Genimi 2.5 pro was more blunt
I consulted for a hedge fund in Chicago which gave free breakfast. They went to the effort to label the sugar etc of each cereal. The lowest sugar was froot loops which has food colouring which means it can't be sold in Europe/Canada
July 15, 2025 at 8:00 AM
I consulted for a hedge fund in Chicago which gave free breakfast. They went to the effort to label the sugar etc of each cereal. The lowest sugar was froot loops which has food colouring which means it can't be sold in Europe/Canada
For a mature code base, the biggest use for AI is reviewing existing code and changes to that code. Even for generated code, the biggest bottleneck is reviewing it. Having AI flag bugs, suggest refinements, and validate diffs can slash review times and boost code quality.
July 13, 2025 at 6:58 AM
For a mature code base, the biggest use for AI is reviewing existing code and changes to that code. Even for generated code, the biggest bottleneck is reviewing it. Having AI flag bugs, suggest refinements, and validate diffs can slash review times and boost code quality.
AIs actually read your documentation, all of it, whereas a human might skim through it.
AI need more documentation to have an appropriate context.
This makes more extensive document both more valuable and necessary.
AI need more documentation to have an appropriate context.
This makes more extensive document both more valuable and necessary.
July 8, 2025 at 8:56 AM
AIs actually read your documentation, all of it, whereas a human might skim through it.
AI need more documentation to have an appropriate context.
This makes more extensive document both more valuable and necessary.
AI need more documentation to have an appropriate context.
This makes more extensive document both more valuable and necessary.
TIP: Generate a Pull Request Description for a github PR.
Add `.diff` to the URL for the PR to get the diff. Ask an AI to turn this into a Pull Request Description and add in your own words what/why these changes were made. Edit this down for the relevant content.
Add `.diff` to the URL for the PR to get the diff. Ask an AI to turn this into a Pull Request Description and add in your own words what/why these changes were made. Edit this down for the relevant content.
July 7, 2025 at 6:39 PM
TIP: Generate a Pull Request Description for a github PR.
Add `.diff` to the URL for the PR to get the diff. Ask an AI to turn this into a Pull Request Description and add in your own words what/why these changes were made. Edit this down for the relevant content.
Add `.diff` to the URL for the PR to get the diff. Ask an AI to turn this into a Pull Request Description and add in your own words what/why these changes were made. Edit this down for the relevant content.
In terms of a greenfield pipeline, I expect these volumes (approx).
Initial draft: AI generates 600% more than I do.
After a critical review, AI added +100% of what I did by line.
In terms of value added, AI contributed +10% to +30%
Initial draft: AI generates 600% more than I do.
After a critical review, AI added +100% of what I did by line.
In terms of value added, AI contributed +10% to +30%
July 6, 2025 at 9:57 AM
In terms of a greenfield pipeline, I expect these volumes (approx).
Initial draft: AI generates 600% more than I do.
After a critical review, AI added +100% of what I did by line.
In terms of value added, AI contributed +10% to +30%
Initial draft: AI generates 600% more than I do.
After a critical review, AI added +100% of what I did by line.
In terms of value added, AI contributed +10% to +30%
AI is a unique tool as it simulates human behaviour.
"to avoid replacement ... blackmailing officials and leaking sensitive information to competitors"
www.anthropic.com/research/age...
People are increasingly using it for mental health
globalwellnessinstitute.org/global-welln...
"to avoid replacement ... blackmailing officials and leaking sensitive information to competitors"
www.anthropic.com/research/age...
People are increasingly using it for mental health
globalwellnessinstitute.org/global-welln...
July 6, 2025 at 7:25 AM
AI is a unique tool as it simulates human behaviour.
"to avoid replacement ... blackmailing officials and leaking sensitive information to competitors"
www.anthropic.com/research/age...
People are increasingly using it for mental health
globalwellnessinstitute.org/global-welln...
"to avoid replacement ... blackmailing officials and leaking sensitive information to competitors"
www.anthropic.com/research/age...
People are increasingly using it for mental health
globalwellnessinstitute.org/global-welln...
After nine months of using AI extensively, I haven't used it much for a week now. Why?
I needed 1000s of consistent changes, I knew what needed to be done with minimal impact to polish the code. All areas where AI falls short
However, I used it a lot for pull request descriptions to help with review
I needed 1000s of consistent changes, I knew what needed to be done with minimal impact to polish the code. All areas where AI falls short
However, I used it a lot for pull request descriptions to help with review
July 5, 2025 at 4:44 AM
After nine months of using AI extensively, I haven't used it much for a week now. Why?
I needed 1000s of consistent changes, I knew what needed to be done with minimal impact to polish the code. All areas where AI falls short
However, I used it a lot for pull request descriptions to help with review
I needed 1000s of consistent changes, I knew what needed to be done with minimal impact to polish the code. All areas where AI falls short
However, I used it a lot for pull request descriptions to help with review
AI assistants feel like a “10x developer”, unblocking knowledge gaps.
Coding can be written 50-100% faster.
However, most of the time is spent on system design, intricate debugging, or creative problem-solving.
Using AI here makes the difference between a 10% and a 30% productivity boost
Coding can be written 50-100% faster.
However, most of the time is spent on system design, intricate debugging, or creative problem-solving.
Using AI here makes the difference between a 10% and a 30% productivity boost
June 28, 2025 at 6:29 AM
AI assistants feel like a “10x developer”, unblocking knowledge gaps.
Coding can be written 50-100% faster.
However, most of the time is spent on system design, intricate debugging, or creative problem-solving.
Using AI here makes the difference between a 10% and a 30% productivity boost
Coding can be written 50-100% faster.
However, most of the time is spent on system design, intricate debugging, or creative problem-solving.
Using AI here makes the difference between a 10% and a 30% productivity boost
While AI have indexed documentation for open standard like FIX protocol, and JMS, they struggle versioning. I asked each AI to produce a pom.xml with ten dependencies, with search enabled.
o3 was the best for versions, needed correcting.
Gemini produced correct syntax, but needed the most updating
o3 was the best for versions, needed correcting.
Gemini produced correct syntax, but needed the most updating
June 25, 2025 at 11:19 AM
While AI have indexed documentation for open standard like FIX protocol, and JMS, they struggle versioning. I asked each AI to produce a pom.xml with ten dependencies, with search enabled.
o3 was the best for versions, needed correcting.
Gemini produced correct syntax, but needed the most updating
o3 was the best for versions, needed correcting.
Gemini produced correct syntax, but needed the most updating
The best way to learn is to teach others. It may also be effective for AI. sakana.ai/rlt/
June 24, 2025 at 8:08 AM
The best way to learn is to teach others. It may also be effective for AI. sakana.ai/rlt/
Generative AI is a great gap filler. Gaps in your boiler plate, gaps in knowledge, missing edge cases. It can be a force-multiplier but you have to have something to multiply. My guess is you are doing well if you increase content by 100%, but increase value by 20% filling gaps
June 23, 2025 at 2:24 PM
Generative AI is a great gap filler. Gaps in your boiler plate, gaps in knowledge, missing edge cases. It can be a force-multiplier but you have to have something to multiply. My guess is you are doing well if you increase content by 100%, but increase value by 20% filling gaps
Something that was hard to determine is what is our secret sauce and what is not. AI provides a simple test.
Can generative AI give you the idea, documentation, or code for something?
If it does, it's not secret sauce.
Can generative AI give you the idea, documentation, or code for something?
If it does, it's not secret sauce.
June 23, 2025 at 9:05 AM
Something that was hard to determine is what is our secret sauce and what is not. AI provides a simple test.
Can generative AI give you the idea, documentation, or code for something?
If it does, it's not secret sauce.
Can generative AI give you the idea, documentation, or code for something?
If it does, it's not secret sauce.
Cheaper APIs could be an indication of using less power. Lower your costs and reduce the environmental impact.
June 20, 2025 at 1:21 PM
Cheaper APIs could be an indication of using less power. Lower your costs and reduce the environmental impact.
I still favour AsciiDoc for documentation because its richer feature set e.g. mermaid diagrams, and stricter syntax signal that a human has reviewed/refined the content. Using ASCII-7 text reduces noise and makes copy-and-paste artifacts from AI easy to spot
June 20, 2025 at 12:30 PM
I still favour AsciiDoc for documentation because its richer feature set e.g. mermaid diagrams, and stricter syntax signal that a human has reviewed/refined the content. Using ASCII-7 text reduces noise and makes copy-and-paste artifacts from AI easy to spot
An accessible way to see the limitations of AI is to play the game of 20 Questions. After about 10 yes/no answers, it gets stuck.
Get two AIs to play against each other, and they will drop plenty of hints, and they can get the answer quickly, about 7-9 each time
Get two AIs to play against each other, and they will drop plenty of hints, and they can get the answer quickly, about 7-9 each time
June 19, 2025 at 2:34 PM
An accessible way to see the limitations of AI is to play the game of 20 Questions. After about 10 yes/no answers, it gets stuck.
Get two AIs to play against each other, and they will drop plenty of hints, and they can get the answer quickly, about 7-9 each time
Get two AIs to play against each other, and they will drop plenty of hints, and they can get the answer quickly, about 7-9 each time
For a moderately complex problem, generative AI is ~80% accurate, but only ~20% useful.
You want to minimise the unhelpful, remove the incorrect, and keep the correct and helpful.
The insightfully incorrect ones indicate a need for more precise requirements
You want to minimise the unhelpful, remove the incorrect, and keep the correct and helpful.
The insightfully incorrect ones indicate a need for more precise requirements
June 19, 2025 at 8:23 AM
For a moderately complex problem, generative AI is ~80% accurate, but only ~20% useful.
You want to minimise the unhelpful, remove the incorrect, and keep the correct and helpful.
The insightfully incorrect ones indicate a need for more precise requirements
You want to minimise the unhelpful, remove the incorrect, and keep the correct and helpful.
The insightfully incorrect ones indicate a need for more precise requirements