Prospective students seeking professional degrees presumably want to get hired and start working off those student loans.
Employers hiring for entry-level positions for those folks earning professional degrees are presumably more likely to hire candidates that have better/more relevant training.
So, if you’re a school looking to make your graduates more attractive (and thus be more attractive to prospective students)…
Thought this article was interesting…
With this quote:
Because of this, it’s now apparent that using AI is more expensive than hiring people, especially since it offers only limited productivity gains at the moment.
Isn’t that true of candidates who fail hiring tests even without AI?
Do you have a link for a study on “games based” hiring?
I found one from Stanford, but no mention of games.
A well know feature of highly accurate machine learning classifiers is that they can tend to overestimate the “spread” between classes. For example, if i’m trying to predict whether a person will default on a loan, and the actual probably of default is 80%, the model might internally predict 90%. This must then be “recalibrated” if the model is supposed to predict a probability rather than a class (say, default or not). Gradient boosted machines will do this.
Given that, it maybe should not be surprising, that LLMs tend to exacerbate “bias”, such as generating “him” for the pronoun for doctors.
Wouldn’t the LLM just maintain the bias? So if 60% of doctors were male, it would return “him” only 60% of the time?
I think it would depend on the type of algorithm you’re using.
I think it technically depends on the “temperature” setting. A lower temperature means tossing out unlikely words. A temperature of 0 means always choosing the most likely word (him), while a higher temperature means sometimes randomly including less likely words. Typically (always?) some words are tossed out.
But, also I think they do a lot to try to “unbias” LLMs. And also make them more ‘aware’ of their biases, and other errors.
The article has the linked paper (its the same one I think)
Link and paper below:
https://giftarticle.ft.com/giftarticle/actions/redeem/1bc0de47-ad6f-4a92-9fb5-a7915d654bea
paper_260526_200247.pdf (1.8 MB)
Thanks. It looks like the same paper. The summary I found didn’t talk about “games”.
I haven’t applied for a job in a very long time. I don’t have any idea what kind of computer games employers use to evaluate applicants.
A model maximizing a likelihood with a whole lot of data and no other problems would reproduce the 60%.
In practice, i think a lot of models will tend to overestimate the probability.
I have never done a deep dive on this, and am going off memory, but i think it’s probably related to this:
If you are trying to maximize the number of correct predictions then you want to always pick the more likely class. So you’d predict doctors to be male 100% of the time, not 60% of the time. This would exacerbate the existing “bias”.
The same one used to find Ender.
I’m waiting five minutes for a prompt. Ctrl-C and type ‘What’s taking so long?’.
" Nothing should be — I just hadn’t run anything yet. Let me implement now."
Yeah right, you were chewing through tokens like they’re free lol.
Remember, friends don’t let friends use Meta products.
claude code is doing some stuff reallyreally well. It’s really good for big overviews, or taking a lot of data and sorting through stuff quickly. It’s also letting me do some enterprise level stuff - stuff I’ve always wanted to do, couldn’t because manpower. Right now we’re rebuilding my entire website, mostly from a background technical viewpoint. Stuff I struggled with before. Every page technically correct for SEO purposes. Overlaying a secondary linking structure for seo purposes. tracking of all leads from ads through to the crm into a reporting utility so I can track sales - then push data back into the ads. There’s something called ‘EAT’ which is basically dispalying authority online. I have a LOT of authority, but it’s a bit of a technical mess of linking, schema’s and stuff, to make this obvious to google - I dind’t have that before but I’m going t shortly and that may make a noticeable difference in my rankings (i.e. I have 150 articles on my website that are written by me, now google’s going to know that I’m the author,And I’ve been mentioned in every daily national newspaper and on tv a bunc of times, now googles’ going to know that).
What it still does badly, to my benefit, is it still does a crappy job if you don’t know what you’re doing, and it still goes on tangents that have to be yanked back. That’s a benefit because it means anyone can create a website, but not everyone can craete a great website. LLM/s/claude code is allowing me to create a really greate website, something that was very difficulty for anyone before, and almost impossible for me.
