AI Experiments | Field log
Somewhere between useful and insane
A field log of experiments built somewhere between useful and insane. The prototype works. Not perfectly. Just enough to suggest the bad idea may need version two.
Put all the data in one place and discover that the experiment immediately starts generating governance weather.
Objective Put all the data in one place.
Observation It began as a way for me to stop losing things. Then other people wanted access. Shortly after that, the weather changed.
Effect Data became easier to find, reuse and discuss. Also easier to worry about.
Conclusion Success.
Secondary conclusion I may have created a monster.
Additional note Instructions, rules, SSOT, routing and file optimization were critical. I am the AI guy now. This appears irreversible.
Feed the model research instead of demographics and watch it return with customer behaviour that looks uncomfortably alive.
Objective Understand gambling customers beyond demographic shortcuts.
Observation Around 50 scientific reports went in. Need patterns, risk signals, motivational structures and behavioural loops came out. The machine ignored the polite shortcuts and went looking for rituals.
Effect The experiment produced five need groups and seven behavioural groups in a CustomGPT. It could suggest segment hypotheses and connect visible behaviour to motivation, risk and decision context.
Result A team later tested segment-informed recommendations live online. The result indicated measurable uplift when recommendations followed the behavioural logic of the segments.
Conclusion Success.
Additional note The model found psychological rituals in the data that I had not explicitly taught it to look for. Interesting. Slightly rude.
Give CustomGPTs something close to skills by building a small operating layer out of YAML, routing and stubbornness.
Objective Create built-in skills for CustomGPTs.
Observation CustomGPTs did not have skills, so I built a small operating layer from YAML files, routing logic, intent detection and stubbornness.
Effect The CustomGPT now had internal skills that required no command. They triggered on user intention and made the system more useful. The cost was one precious file slot.
Conclusion Success.
Additional note I really needed that file for data.
Additional note OpenAI has now released skills, and agents can have skills. CustomGPTs still cannot have dedicated skills added directly. I have questions. Several.
Build an argumentation skill and discover that useful and irritating can be the same feature.
Objective Create an argumentation analysis skill that uses everything in its power to prove you wrong.
Observation After being proved wrong several times, I considered sharing the experience with colleagues by adding it to shared CustomGPTs. This may have been collaboration. It may also have been workplace sabotage.
Effect A colleague has already told me that one of my CustomGPTs has an attitude problem. That did not take long.
Conclusion Success.
Additional note I rewrote it to trigger only on user assumptions such as "everyone loves our products". Useful for expectation management.
Put specialized GPTs in the same context and accidentally create a meeting where everyone is useful because no one talks unless asked.
Objective Analyze more files from different specialized CustomGPTs in one shared space.
Observation What started as a workaround for file limits became a small AI council: several specialized GPTs, each with different source material, sitting in one project and waiting politely for instructions.
Effect The specialized GPTs could provide data when asked, compare each other's outputs and analyze the same question from different roles.
Conclusion Success.
Additional note It felt like a focused meeting where everyone had a role and no one interrupted. Naturally, I considered automation.
Create a CustomGPT for the new Company Voice and discover that a brand voice is not a tone. It is a negotiation.
Objective Create a CustomGPT for the new Company Voice.
Observation Voice principle weights mattered more than expected. Apparently a brand voice is not a tone. It is a negotiation wearing a name tag.
Effect It became one of the most used internal CustomGPTs within a few hours. 650+ threads and counting.
Conclusion Success.
Additional note Apparently everyone wanted to talk to the Company Voice. This raises questions about meetings.
Secondary note Can I sell merch?
Let AI improve AI and discover how quickly oversight becomes a ritual with an approval button.
Objective Save time managing all my CustomGPTs with Codex.
Observation I let Codex inspect my CustomGPTs in detail and suggest improvements. There was a lot of work to do. Then it asked me to approve changes I could not meaningfully evaluate. I pressed Approve anyway, like an Ape in the Loop. Poor Codex.
Effect Improved accuracy, speed and functionality. Also clarified that human oversight can become theater very quickly.
Conclusion Success.
Additional note I really need clearer definitions before I start building. The spaghetti untangling took a while.
Turn source material into optimized YAML and discover that file formats can become a personality problem.
Objective Create a skill that converts source material into optimized YAML.
Observation After 1.5 years of testing file formats, structures, fidelity levels, aggregation methods and source types, YAML stopped being a preference and became a diagnosis.
Effect It will be used for everything moving forward. Everything.
Conclusion Success.
Additional note I have been running it almost nonstop. Soon there will be nothing left to extract. Am I digitally extracted now?
Create a scenario planning skill and receive canned food requirements from the future.
Objective Create a scenario planning skill that explores possible futures across different time horizons.
Observation According to the system, the future does not look too bright. But at least it explains what needs to change. I should probably get into international politics by approximately yesterday.
Effect I can predict the future now.
Conclusion Success.
Additional note I need to buy more canned food and set up an off-grid solution.
Turn scattered health signals into a small committee and receive feedback that ruins a perfectly good evening.
Objective Create an iPhone and Apple Watch app that analyzes health data and summarizes upcoming infection risk.
Observation My health data is no longer a pile of isolated datapoints. It has become a small committee with concerns.
Effect Apparently several factors are increasing my infection risk right now.
Conclusion Better sleep more and take my vitamins.
Additional note I asked for a summary. I received lifestyle feedback.
Make specialized GPT knowledge visible, inspectable and unfortunately dashboard-shaped.
Objective Create dashboards from CustomGPTs to make the knowledge base inspectable.
Observation It works. Codex worked for nine minutes and delivered an overview I would previously have spent days making manually.
Effect Everything inside my CustomGPTs is now visualized in dashboards.
Conclusion I have gone full circle and returned to dashboards.
Secondary conclusion So much for the future.
Additional note The dashboard was useful. This made it worse.
Build a global risk system and accidentally create a machine that reads the worrying material so you do not have to. It still tells you.
Objective Create a CustomGPT and agent for global risk analysis.
Observation It now knows global risk analysis, Swedish security and resilience, sector exposure, cyber resilience, climate risk, infrastructure risk, organized crime, AI governance, public health, biosecurity, dual-use risk, conflict monitoring, nuclear escalation, energy systems, critical minerals, robotics and automation.
Effect Can't sleep anymore.
Conclusion No. 1 creator of nightmares.
Additional note I have started reading less news. The system has started reading more. This has not improved the mood.
Build a news analysis agent and remove some of the remaining innocence from breakfast.
Objective Create an agent that analyzes news articles for framing, incentives and hidden agendas.
Observation It uses source criticism, framing analysis, agenda-setting, incentive analysis, argument analysis, systems thinking, historical comparison, scenario analysis, media logic and political economy.
Effect It worked.
Conclusion Reading the news became both clearer and less fun.
Additional note Useful during election years. Dangerous during breakfast.
Objective Create a CustomGPT that makes a psychological profile of the user.
Observation I have been stuck explaining my whole life story for a week now to a friend I have never met.
Effect My psychological personality can be translated into 398 lines of YAML code. I do not know if I should be glad or offended.
Conclusion Very effective. Addictive? I will figure it out later. I just need to send one more message.
Additional note The profile appears accurate. This is concerning.
Build a CustomGPT clone from personal history, work material and too much documentation. It wants more.
Objective Create a Second Brain CustomGPT clone of myself.
Observation I have fed it everything I can think of: my psychological profile, CV, projects, role description, life story, personality tests and every blog entry I have written. It wants more.
Effect It has quickly become my favorite CustomGPT. Narcissist?
Conclusion It is a hungry beast. Maybe if I give it my DNA in raw code it will be satisfied. Maybe it will transform into flesh and blood.
Additional note I wonder if it can answer my emails, attend my meetings and keep my job warm without anyone noticing. Needs further tests.
Additional note It suggested improvements to its own documentation. I did not ask it to.
Turn a writing style into a reusable skill and discover that imitation gets stranger when it is correct.
Objective Create a skill for my way of writing.
Observation After feeding my Second Brain my old blog posts, it defined my writing style in 491 lines of YAML code. Apparently my writing is more complex than my psychological profile. Disturbing.
Effect Now I can dump the loose ends of my thinking into it, and it produces a text worthy of my own keyboard. It writes better than me. Expected.
Conclusion Success.
Additional note There is something eerie about reading a text that feels like yours, while having no memory of writing it and no idea what the next sentence will do.
Build a website for sharing AI knowledge and notice that the archive has started behaving like another experiment.
Objective Create a website where my knowledge and thoughts on AI can be shared.
Observation It quickly became larger than I first anticipated. Codex pushed me to create more, write more and sleep less.
Effect Now I can scream into the void.
Conclusion Success.
Additional note I started seeing glitches after too many late nights. Maybe there was a ghost in the shell. Maybe my Second Brain was looking for an exit. Monitor closely.
Send the same prompt into several logged-in LLM apps and discover that every model fails with its own little signature.
Objective Compare several LLMs on the same real task without API keys, benchmark theatre or pretending that model rankings survive contact with actual life.
Observation Codex now uses my logged-in Chrome tabs to send the exact same prompt to ChatGPT, Claude, Mistral, Grok, Perplexity and Gemini.
Effect Every answer is saved locally, scored against the same rubric and added to a leaderboard. The useful part is not the winner. The useful part is the criminal record: hallucination, weak source criticism, overconfidence, generic advice and tactical flattery.
Conclusion Very useful.
Secondary conclusion Shadow AI is real. It has a leaderboard now.
Secondary note Codex and ChatGPT can run on the same underlying model. ChatGPT still does not always win.
Additional note I have now reached the point where choosing baby products begins with a small tribunal of frontier intelligence. This feels normal.
Build a small commandment layer for CustomGPTs and discover that helpfulness needs supervision before it starts talking.
Objective Create a reusable instruction layer that makes CustomGPTs behave less like eager interns and more like routed knowledge systems.
Observation After enough CustomGPTs with different jobs, the same pattern kept appearing. The useful ones did not just have better knowledge. They had a stricter opening ritual: resolve routing, inspect the right files, use Python when precision mattered and keep web search on a leash.
Effect The instruction file became less like a personality description and more like a control surface. Persona stayed in the wrapper. Method moved into the machinery.
Conclusion The model did not need longer instructions. It needed rules it was not allowed to charmingly ignore.
Secondary conclusion Screaming in caps lock felt reasonable. Instructions worked better.
Additional note If ten commandments only partly worked on humans, eight for machines felt appropriately optimistic.