Is AI Unreliable Compared to the Competition?

In a world of everyone accusing AI of being unreliable, I wanted to take the time to talk about the competition.

Below is an article written by a Stanford grad who even attended a class on the subject matter. Yet their analysis misses a fundamental idea of the topic itself: randomness does not exist in computers that you and I will use, it can't.

Any discussions about algorithms being used to generate random numbers is moot as those are only pseudo-random number generators. If you know the initial value, or the seed, you can calculate the output. The only known truly random processes humans have observed are natural phenomena such as radioactive decay.

Similarly, ai is deterministic as most are calculating the gradient descent to determine the most likely word that should appear next. While the output might change, that is not what random means.

Yes, ai is still not 100% reliable, but frankly neither are people. Even a source from an ivy league graduate has noticable errors that impact the results he presents. Don't just check the work of ai, fact check all of the information you receive.

Modern AI requires that you understand the field you are using it for, you should look at ai like an overly enthusiastic intern who is extraordinarily fast. Aai can speed up your workflow; engineers, lawyers, and even customer service are leveraging ai to do better work faster. Do you plan on being left behind, or will you ride the rising tide to new heights?

Source: https://towardsdatascience.com/llms-are-randomized-algorithms/

Why Models Hallucinate

This might also have implications outside of new models. Including training data multiple times for an output to increase the likelihood it catches the correct answer might be a decent strategy. Hopefully the files won't get too big and cause other issues, but for generating a timeline for instance, having multiple copies of medical charts might increase accuracy without additional time or resources.

Assuming that the increased size fits within the token limit so that you do not run into other issues... (here's hoping token limits will be a think of the past soon! Maybe the trillions in infrastructure will help:/)

Source: https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4aaa5/why-language-models-hallucinate.pdf

Agents vs Models

I have found that using advanced agentic like prompts directly into GPT5 over 4o agent mode in Copilot shows a dramatic difference in the quality of the output. In my workflow, it was able to generate numerous litigation discovery documents with extremely few errors that were mostly stylistic or formatting. GPT5 can reference hundreds of pages of documents across multiple providers, and formats in order to complete these tasks.

The experience of using it has almost been a bit spooky. The depth of knowledge that it has on tap is extroadinary, and now the hallucinations are down to a useable level in certain usecases with proper review techniques. I've been developing the new company website and wanted to have a knowledge base for people to go to thats focused on business owners with an action link at the bottom if a reader has that sort of case. But, it was spooky because it must have lifted some of its tips from m&a textbooks and biographies. GPT5 was using ebitda optimazation strategies that I learned working at Peregrine... The research and strategy implications over prior models can not be understated.

It can also read medical charts with extreme accuracy, with some results as low as a 1.6% error rate... that is almost expert level review in seconds which can be scheduled the moment you get access to the records.

Source: https://mpgone.com/is-gpt-5-accurate/