Raw LLM Responses

LLMs contributing to productivity is still up on the air, IMO. I still think it's just an easy way to sway VCs to spend money. I don't think the security issues will be solved so long as the customer facing agents are able to access sensitive data, while taking that away is going to severely hamstring their usefulness. You can parameterize responses by making the LLM call an API that will only give it data that's related to the current user, but again, you're hamstringing usefulness there and hiring humans will likely get better results. I've also been seeing white papers released about once a month using variations on similar techniques from previous methods to completely break through whatever safeguards there are in place, but my knowledge in this area is limited. Gpt5 is polished and optimized in your mind, but a lot of paying customers are expressing frustration at the apparent downgrade. The router choosing the "optimal route" is the nail in the coffin for a lot of them. To me, that's a cost saving move in preparation for a more corporation targeted business model. It seems like they're getting diminishing returns from pumping more into the models, so they're going to start charging business tier prices for what the regular consumer used to get in order to show they can make a profit and keep getting investors to throw money at them.