Our Insights
Insight is the difference between what the data's saying and what it actually means.
A single test proves nothing about an AI system
Traditional tests are deterministic, so one assertion settles it. An AI system, on the other hand, changes each time you test it. This means modern systems need to test outcomes statistically.
More training data is not always a good thing
A model is only valid for the conditions its training data came from. When the underlying system keeps changing behaviour, older data can mislead the model. The trick is to identify the point of change and train only on data that represents the system's current behaviour.
Technique is no longer the constraint
The cost of running an experiment has fallen to almost nothing. What has not fallen is the cost of knowing which experiment is worth running. Knowing what is possible matters more than knowing how to do it.
Any software older than two years should be assumed exploitable
The absence of a known exploit is not evidence that a system is safe. Software left untouched for years is full of dependencies, assumptions, and code nobody remembers writing, and the tools for finding flaws in it keep improving. Age is evidence of exposure, not of stability.
Product delivery should now include the means to maintain
Handing over software and documentation leaves a client with a description of a system and no way to change it. What belongs in the delivery is the means to maintain what they have been given. That now means the agent that did the build, not just the code and the manual.
A good patent needs careful layering
A good patent is built in layers. If it comes under attack and the outer claims fall, the inner ones still hold. Most patents fail because they were built badly, not because what they were protecting was at fault.
The age of tailored software is here
Subscription software is priced on everything it includes, and most of it goes unused. Now that software production is 'cheap', we can afford to have our software fully customised to our way of working, building only what we need. Our software should adapt to our needs, not the other way around.