mk_chan

This is just plain worse than giving the agent an instruction to just always read the code.

Documentation was always a proxy for the code so the people who didn’t have context could get started. Obviously if reading the code and reading the doc have the same cost, the doc is completely worthless. This was always the case. You could never depend only on the docs. It’s always the code that actually mattered.

Just shows the offset openai feels like it has to add to ‘equalize’ the average discourse of its training material

In my experience an LLM does 2 things: 1. Bring you up to some average-LLM level when you don’t have the skills/knowledge to actually do what you want. 2. Work at 80-90% of your capacity but WAY faster than you physically could depending on how much context you provide it. If you don’t provide it sufficient context to do what/how you want it to do, of course it might default to something you don’t want.

Microsoft ban any % speedrun soon.

Is this what the employees do nowadays while their AI is generating code?

This is very important to write on. A lot of people believe news is worth consuming for the truth and often cite it as a primary source of information. News producers may not necessarily lie but they cherry pick to maximize reach and that content plays on peoples belief that what they see on the news is all the information you need.

The news in a vacuum can actually be quite misleading and I too believe people should realize that it is not the ‘whole’ truth.

That should happen if the training dataset (which is presumably based on the real world) reflects that happening.

Following the paper, if you end up with a gender balanced workforce, it implies there is surely a bias in one of the variables - the candidate pool (like you say) or the evaluation of a candidate or other related things. However the bias must also reverse to equalize once the balance tips the other way or actually disappear once the desired ratio is achieved.

Edit: it should go without saying that once you hire enough people to dwarf the starting population of the startup + consider employee churn, the bias should disappear within the error margin in the real world. This just follows the original posted results and the paper.

Not sure why OP is getting downvoted. Going the premise of the cited doctrine, it follows precisely that even doing illegal things are acceptable as long as there is net gain until the expected future.

OP doesn’t even claim this is the one true doctrine or anything of the sort.

Diagnosis: Misaligned incentives correlate with acute miscomprehension

I don't actually use the program in my stead, I just put it up against other programs in computer chess tournaments.

It’s the same reason I stopped playing chess, but it only hit me later that I lost interest because of the program I wrote.

Now I just play casually if a friend wants to.

The way I put it was meant to be slightly amusing, but let me put it a different way for the sake of discussion.

For this purpose, I like to mentally model a country (or territory which decided to hold an election of some kind) as a collection of points in a multidimensional space of values or policies which are up for election. Each point represents where a voting person would stand according to their personal principles.

Now, I’d like to postulate that the act of voting corresponds to finding the center of the cluster of points. In this model, it’s easy to imagine the scenario where a very large portion of people (and possibly even the majority) are deeply dissatisfied with the result because they are too far on many dimensions from the center that was elected. This is actually further exacerbated by many factors in real life like the asymmetry of information between people.

The obvious solution would be to create more cluster centers instead of one fat cluster that leaves everyone not very happy. This corresponds to states or provinces within a country in the real world.

So coming from that line of thinking, it leads me to believe that sometimes to maintain a democracy, you need to cut the outliers away to move/make the center such that people are happy with the result. Essentially you enforce the will of the majority by cutting away the minority until it is no longer a tyranny to do so.

Of course this is gross simplification of real life. It is ignoring factors such as external threats, instability, unpredictability, information availability and so on. However, I think it’s useful to think this way for many purposes.

Addressing your extreme example of a criminal, that’s a point so far off in the space that you definitely want to cut it out of the system.

Ironically, tyranny of the majority is a very very good definition of democracy if you drill the majority down into subcategorical majorities recursively until you reach the individual who might be severely disconnected from the averages.