Reliability and predictability
The best way we work together with computer systems right this moment is predictable. For example, once we construct software program programs, an engineer sits and writes code, telling the pc precisely what to do, step-by-step. With an agentic AI course of, we don’t present step-by-step directions. Quite, we lead with the result we wish to obtain, and the agent determines the way to attain this objective. The software program agent has a level of autonomy, which suggests there may be some randomness within the outputs.
We noticed an analogous difficulty with ChatGPT and different LLM-based generative AI programs once they first debuted. However within the final two years, we’ve seen appreciable enhancements within the consistency of generative AI outputs, because of fine-tuning, human suggestions loops, and constant efforts to coach and refine these fashions. We’ll must put an analogous stage of effort into minimizing the randomness of agentic AI programs to make them extra predictable and dependable.
Knowledge privateness and safety
Some firms are hesitant to make use of agentic AI attributable to privateness and safety considerations, that are much like these with generative AI however may be much more regarding. For instance, when a person engages with a giant language mannequin, each bit of data given to the mannequin turns into embedded in that mannequin. There’s no approach to return and ask it to “overlook” that data. Some forms of safety assault, equivalent to immediate injection, make the most of this by making an attempt to get the mannequin to leak proprietary data. As a result of software program brokers have entry to many various programs with a excessive stage of autonomy, there’s an elevated threat that it may expose non-public information from extra sources.