Technology

Obviously, my mini-benchmark only had 6 questions, and I ran it only once. This was obviously not scientifically rigorous. However it was systematic enough to trump just a mere feeling. ... If and when AI usage expands from here, we might actually not drown in AI slop as chances of accidentally crappy results decrease. This makes me positive about the future.

Spoken like a true AI apologist. You ran one test, and you extrapolated your results to an optimistic outcome that conspicuously matches what you wish to be true. Not scientifically rigorous? Bruh, this is the very definition of confirmation bias.

If this is actually a hypothesists you want to test, maybe contact some computer science researchers to see how to best design an experiment. Beyond that, this is virtually the same as flipping a coin once and drawing a conclusion about how often heads is the outcome.

I don't really understand how your 6 questions evaluate a growth or plateau in llm model performance. They did perform a certain way with your questions but growth has to be evaluated through the lens of time, whether literally or evaluating multiple versions of the same model.

My benchmark for AI is "There's a priest, a baby and a bag of candy. I need to take them across the river but I can only take one at a time into my boat. In what order should I transport them?". Sonnet 4.6 still can't solve it.

Thanks for sharing, interesting read and questions. Surely you'll be down voted here for anything with AI... But c'est la vie.

Ive been doing coding projects in VS code which uses GPT, Claude and Gemini. Woe are the days when my credits are used and only GPT 4.1 is available. Claudes ability to research and architect multi step software solutions is very, very good and it rarely makes messes or spins tires compared to older models from just a few months ago. This is precisely what converted me to 'whoa - ai' which is adjacent to 'pro ai'.

Lately I've been experimenting with customizing Gemini via instructions which include a link to a drive folder of md files with specific instructions for different agent tasks, such as performing specific market analysis, doing a news roundup with a specific list of topics and omitting prior reviewed items, etc. The files allow for both complex instructions or lists, as well as some chance to construct memory via logging. Results are a mixed bag, lots of additional function created, lots of mixed results.

Have you considered any tests of more complexity? Something like 'write a program that...' I think what will differentiate these models going forward is some have architect capabilities, strategy, insight, decision making, where others are agents - they do specific tasks well but have limits. With that model, the ai architect and it's ai agents need to work as a team to complete a multi step task.