partypete

The idea that a home loan is risk-free ignores decades of recent history. A quick google search for "2008 GFC" will bring you up to speed.

Given that home loans are not, in fact, risk-free, how do you expect anyone to risk the large up front cost of financing a home if they do not have any expectation of profit? The small profit that landlords make off of these investments is mostly that risk premium. As a tenant, I am not liable to the bank for hundreds of thousand of dollars if any one of hundreds of calamities outside of my control were to occur rendering my home worthless. In exchange for this relative lack of risk, I gladly pay the 2% premium to my landlord. This is not to mention the benefit of mobility, among others.

Home loans are one of the most accessible loans in modern western civilization, so no, banks don't "always choose people who are already wealthy."

I assure you that if we were to remove the profit motive from financing the development of housing stock, there would be far less of it and home prices would certainly increase.

I think this is an excellent way to think about LLM's and any other software-augmented task. Appreciate you putting the time into an article. I do think your points supported by the graph of training steps vs. response length could be improved by including a graph of (response length vs. loss) or (response length vs. task performance), etc. Though # of steps correlates with model performance, this relationship weakens as # steps goes to infinity.

There was a paper not too long ago which illuminated that reasoning models will increase their response length more or less indefinitely toward solving a problem, but the return from doing so asymptotes toward zero. My apologies for missing a link.

Author here. I just used a free pdf version of the book and the legacy software package which they link to on the website. Great experience. The authors have since created a self-contained web IDE version of the book/course, but if you know the basics of working with a terminal, you'd probably prefer the software package.

Keep in mind that the actual profit if the bot were not randomly choosing between each model version would be about 600k/hour. You're conflating the experimental profit, wherein each model only has access to 1/3 of the potential trades, to what the profit would be if only the best model were making trades.

Also these bots have been running for over a year almost every day without being banned.

That’s right - I retrained the model every 7 days. These experiment results are from the most recent one-week period.

The training data is composed of the bot’s previously executed trades from the last 63 days joined on the price/volume data. So the model discovers patterns in the price/volume data that have historically lead to well-performing trades. The dependent variable is gold/second generated from a historical trade.

For what it’s worth, I am a relatively young engineer who has recently been working very closely with other engineers over the age of 55 on a daily basis. It has been one of the more enriching periods in my career so far. I often wonder if the ageism that I see so often referred to on Hackernews is more projection than reality because I consistently find it pleasant to work with individuals who have a very large body of experience to draw from. Ageism in general baffles me as one could very well ask the question: “would you rather have a carpenter with 2 years of experience working on your house, or one with 30 years of experience?” The answer seems obvious to me, and the question could easily be applied to software engineering.

You may be interested in some of the foundational papers exploring game theory models similar to the Ultimatum game[1][2]. These are known as Iterated Prisoner's Dilemmas.

---

[1] The Evolution of Cooperation (https://ee.stanford.edu/~hellman/Breakthrough/book/pdfs/axel...)

[2] Evolutionary Dynamics of Spatial Games (https://www.sciencedirect.com/science/article/abs/pii/016727...)

Came here to say this. They only just recently got an AMD GPU on MLPerf thanks to a (different company), Tinycorp by George Hotz. I guess basic ML performance is too hard a problem.