Olivier Corradi

Most feature lists are too long and rating grids rarely separate signal from noise. MaxDiff is a practical alternative: show small sets of items, ask for one most and one least, repeat across screens, then convert those observed choices into preference shares that support clear decisions. Design starts with the decision you need to make. If you want a strategy ranking, ask about importance. If you care about behavior, ask about frequency of use. Keep item wording short and concrete. Choose the design numbers up front. Total items, items per screen, how many times each item appears, total number of screens computed as items times appearances divided by items per screen. Aim for balanced exposure so each item appears equally often and with a fair mix of neighbors. Pretest for accessibility and place the most and least controls together. Export one row per respondent. Include the design matrix that lists which items appeared on each screen. Keep only complete respondents. Code most as 2 and least as 1. Preserve a respondent identifier for later joins. Begin with a quick diagnostic. For each item, compute Best minus Worst divided by Times shown. Use this as a sanity check. It often mirrors the model ranking at a high level but it is not sufficient for final decisions. Move to a choice model for decision ready outputs. The model estimates a utility for each item so that the observed most and least picks are likely under the model. Convert utilities to preference shares that sum to one. Two established paths are common. Multinomial Logit gives a clean sample level ranking and is fast to run. Hierarchical Bayes estimates a distribution of utilities across people and returns individual level scores along with a robust total, which is valuable when you expect heterogeneity or want segmentation. In R, mlogit is a practical entry point for Multinomial Logit. Apollo supports richer specifications. ChoiceModelR and bayesm are popular for Hierarchical Bayes. For an end to end workflow that reads Qualtrics style exports and returns decision friendly shares and segments, flipMaxDiff is convenient. For best worst data preparation, bwsTools is helpful. Report results as a ranked Total share chart with clear labels and short explanations. Add class or respondent views when the decision warrants it. Include a short methods note that states the design summary, the model you used, and at least one stability check such as a split sample comparison. Answer what if questions with a simple simulation. Remove the items under consideration, rebase shares within each class to sum to one, then recompute the Total as a class size weighted average. This turns potential scope changes into concrete impact estimates.

80

10 Kommentarer

"With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks using only 1000 training samples. The model operates without pre-training or CoT data, yet achieves nearly perfect performance on challenging tasks including complex Sudoku puzzles and optimal path finding in large mazes. Furthermore, HRM outperforms much larger models with significantly longer context windows on the Abstraction and Reasoning Corpus (ARC)" https://lnkd.in/gZU-AkKp https://lnkd.in/g9dqj_WA

173

19 Kommentarer

I love deep-diving on the technical stuff, but finally I wrote a more 'high-level' overview on how AI is impacting engineering industries: https://lnkd.in/e8pN7CtE The case is clear for LLMs (chatbots, agents, etc.), but here I focus on 'everything else'. Video: GE Aerospace engine testing for flying through ‘ice clouds’ in their specialized test facility. Credit to GE Aerospace LinkedIn page as the video source. #ai #chatgpt #engineering #automotive #aerospace

465

12 Kommentarer

Water boils at 100°C. Not at 98°C. That final push? It changes everything. Product development is no different. Idea validation — At 98°C, you assume users will need it. — At 100°C, you validate with real users and iterate from their feedback. (Think Airbnb listing their own apartment before raising funds.) MVP build — At 98°C, you rush to release something minimal. — At 100°C, you still move fast, but polish enough for someone to use it daily without friction. User experience — At 98°C, the app works, but feels clunky. — At 100°C, small design details turn first-time users into long-term users. (Think of Notion’s obsessive focus on UX before it took off.) Go-to-market — At 98°C, you launch and wait. — At 100°C, you show up, talk to users, fix what’s broken, repeat. Tech & team — At 98°C, you hire help. — At 100°C, you partner with people who think like co-builders and care about your vision. The difference between a good product and a great product often comes down to that extra 2 degrees of effort. Always go all in

42

16 Kommentarer