1/6 To assess performance at the individual level, at the end of every project you will ask managers NOT about the skills of each team member but about their own future actions with respect to that person! Here is a quick clickable example: bit.ly/3NGXvB1

Full article: bit.ly/3UMv0Ws

1/5 One approach to overcome this problem is to use - standardized, systematic project-based assessment...

1/4 Scullen et al. (2000) found that 62% of rating differences came from the evaluator’s own quirks and preferences, and only 21% was based on real performance! (grinning troll face)

1/3 These factors tend to colour their (our) judgment more than the employee's actual work!

1/2 When managers rate employees, their assessments often reflect (1) their own experience, (2) personal values, and (3) whatever data they can recall - usually just recent events (hello, recency bias) instead of systematically collected historical data points.

It’s ridiculous that a convicted felon could even have a shot at the presidency. Imo. If these criminal cases get swept under the rug, it could mean the end of democracy in the U.S.

Honestly, a social network is nothing without a critical mass of people, and Bluesky just doesn’t have it. It’s like throwing a party where no one shows up. Sure, the sky’s blue and all, but without enough people, there’s no point sticking around.

I bet it’s #1, i had the same exp.

To conclude, LLM-based simulations of experiments could offer significant value in areas such as (1) intervention design, (2) minimizing harm to human participants, (3) pilot testing study materials, and (4) predicting subgroup effects, (5) pre-testing product hypothesis, etc.

A recent study titled "Predicting Results of Social Science Experiments Using Large Language Models" by Ashokkumar et al. (2024) found a strong alignment (r = .85) between simulated and observed effects across 70 pre-registered studies, 476 treatment effects, and over 100K participants.

…the results show that the LLM crowd outperformed a simple no-information benchmark and is not statistically different from the human crowd.

Another study by Schoenegger et al. (2024) on the "wisdom of the silicon crowd" used an LLM ensemble approach consisting of a crowd of 12 LLMs. They compared the aggregated LLM predictions on 31 binary questions to the predictions of 925 human forecasters from a three-month forecasting tournament...

For example, last year’s study, “Can AI language models replace human participants” by Dillion et al. (2023), focuses on moral psychology and suggests that GPT-3.5 (text-davinci-003) generates judgments about a variety of moral scenarios that strongly correlate with average human judgements.

Similarly, other studies suggest that #SyntheticUsers mean values tend to be highly similar to those of their human counterparts.

Multiple shreds of evidence suggest that #LLMs are pretty effective at providing answers to questions that closely reflect those collected from real humans, hence effectively simulating human answers, behaviours, and psychological traits.

With #o1 out, we’ve got System 2 (slow thinking) alongside System 1 (fast thinking). #SyntheticUsers will now better mimic human behavior, showing more human-like cognitive patterns and moving beyond simple reactions to more thoughtful, context-aware actions.

Indeed!

Why coffee in the US sucks so badly? :) or I’m wrong? :)