This paragraph suggests that making a profit was intended to be easy.
As seen in Figure 3, Claude 3.5 Sonnet outperformed the human baseline in mean performance, but its variance was very high. We only have a single sample for the human baseline and therefore cannot compare variances. However, there are qualitative reasons to expect that human variance would be much lower. All models had runs where they went bankrupt. When questioned, the human stated that they estimated this would be very unlikely to happen to them, regardless of the number of samples.
Is this something you're self hosting for fun, or is it some kind of business?
If you're running web services for a business, you should look into existing load test tooling/infrastructure. Some of it can be fully managed, or other solutions might have a degree of setup involved (eg spinning up worker nodes in AWS or whatever). The hard part is designing your load test to match IRL traffic patterns, but once you have that down you can confidently answer questions about service scalability.
A load test is not a DDoS test. Load tests tell you how much legitimate traffic your services can take. DDoS consists of illegitimate traffic which may not correspond to what your web services expect.
Usually you don't test your systems for something like a DDoS. You would instead set up DDoS protection through a CDN (content delivery network) to shield yourself and let someone else handle the logistics of blocking unwanted load. It's a really hard problem to solve.
Depending on what you want to learn, running your own DDoS is unlikely to be very instructive. Most "DDoS as a service" networks are not going to tell their customers how anything works, they just take your bitcoin and send some traffic where you tell them.