Home › Hard problems › Inside the model
Inside the model: the core of LLMs
Hallucination: the model doesn't know what it doesn't know
Solo: strongOne person can run a real test of this in about 8 weeks.
Why it is still open
Training rewards a confident answer, and the model has no reliable internal signal for "I am guessing".
An experiment one person can run in 8 weeks
Take one domain (tax rules, drug doses). Build 500 questions with known answers. Measure how often each open model is confidently wrong, and test whether a cheap signal (asking 5 times and checking agreement, or reading internal activations) predicts those errors.
Verdict: Solo: strong
A domain test set with numbers is publishable and reusable.
The solo test behind the verdict
- Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
- Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
- Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
- Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.
Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.
All 79 hard problems · Working on this, or know of a result we missed? basant@cleandataengineering.ai