AI^2: writing scripts and plugins so that AI prompts are programmatically produced and monitored. Your CLI AI is running other AIs and monitoring those.

The widespread use of AI has led to expectations for best-of-breed solutions that far exceed productivity gains, meaning major tasks can take much longer. www.linkedin.com/feed/update/...

Sureal moment: I told Codex the way /prompt worked wasn't right, and now it's looking through the ~/.nvm Codex code to figure out why

Yesterday, I was reminded that AI creates new classes of bugs. Gemini 3.0 Pro (preview) CLI kept “hallucinating” my config as code edits. Too much AI gap-filler kills determinism – we need Codex/Claude-style smart glue, not opaque agents. www.linkedin.com/feed/update/...

Tip for AI CLI: ask it to suggest topics to research to fix a problem, get a chat AI to do the research with web search, and paste the results into the CLI, save it to an adoc|md for future reference

Oddly specific rety time for Gemini CLI

to avoid an empty block complaint. while (methodReader.readOne()) { continue; }

A test that never failed doesn't add much value. Ask an AI to write a failing test, and it will write a broken test. The solution I have found is: 1. Ask it to find/fix a bug and write a test to exercise it 2. Keeping only that test, check it fails without a fix 3. Write the fix

"Much of the most valuable Gen AI usage will grow out of existing business applications, rather than be wholly new. These applications will draw on existing domain models and infrastructure, and be more robust and useful as a result." medium.com/@springrod/d...

I find AI is more useful as a simulated developer user of the software. It doesn't find bugs as often as it stumbles on usability issues.

The AI boom isn’t just about algorithms — it’s about money, power, and a race to build infrastructure on a scale we’ve never seen before youtu.be/NbL7yZCF-6Q?...

Running codex cli via ssh from my phone, works surprisingly well while doing house work and other tasks

Visiting Zegreb for some consulting

On 20 Sep 2025, the UK had its third mass shooting for the year, the highest number in a year since 2006. Also on 20 Sep 2025, the US had 3 mass shootings that day, which has the highest in one day since 18 Sep 2025. The US has 5x the population of the UK, not 365x.

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

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.

Reporter: The president of Venezuela called the strike on the boat illegal Trump: What’s illegal is the fact that 300 million people died last year from drugs

Parwill was the name of an Australian food product from 1928 to 1935, originally branded as Vegemite, which was a failed attempt to compete with the popular British brand Marmite. The name was a pun on Marmite. uk.pcmag.com/ai/159686/el...

Coding is a mystery story where you are the detective, perpetrator, and the victim

The minimum tarrif for goods mailed into the US will be $80 to $200 depending on the country. A $5 cup direct from China will have a $200 tarrif.

#!/usr/bin/tail +2 Hello World

GPT-5: Generate a Where's Waldo, and find Waldo

AI coding tools are powerful for certain tasks with significant caveats. They can enhance developer learning and enjoyment The human developer’s role is still central. Understanding the problem, ensuring quality, and making architectural decisions remain human responsibilities that AI can’t shoulder

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 refined a prompt across multiple AI models to consistently generate low-latency Java code for formatting timezone offsets, moving beyond one-shot requests to detailed instructions that bridge performance gaps blog.vanillajava.blog/2025/07/impr...

I walk through a single “low-latency” Java in six different AI models (Gemini 2.5 Pro, OpenAI o3-Pro, o4-Mini-High, Claude 4, Grok 3 Think, Copilot) each rework it. I consider how models deliver different trade-offs blog.vanillajava.blog/2025/07/aski...

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

It's a mistake to assume AI saves time, especially for experienced developers. For senior developers, "analysis reveals that AI actually increased task completion time by 19%. ... www.techradar.com/pro/using-ai...

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.

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.

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%

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...

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

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

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

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

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.

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

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

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

Hot take: If AI can generate documentation, there is >80% chance it isn't worth writing. a) It's something an expert will already know this, or b) You can ask an AI to produce that content. Treat AI output as an indicator that it *doesn't* need to be explained

Developers consistently fail to realise that understanding the problem is 90% of the effort, and writing code is only 10%. They compound that mistake when using generative AI, making the writing code easier at the expense of understanding. Use multiple AIs to explore the problem

AI compresses experience curves. < 5‑yr devs will fill today's 10‑yr roles; 5‑10 yrs will step into 10‑20; 10‑20 yrs into 20‑40. That frees 20+ yr veterans to tackle the big, risky projects we keep postponing

o3-pro is low on details of it's thinking. My favourite comment is "Progress is steady", then nothing for 10 mins.

One of my favourite websites now with AI making up explains www.tylervigen.com/spurious-cor...