2/ Agents will understand context
“Claude needs to learn enough about your particular situation and the constraints that you operate under to be useful. Things like what particular role you’re in, what styles of writing or what needs you and your organization have.
ANTHROPIC
“I think that we’ll see improvements there where Claude will be able to search through things like your documents, your Slack, etc., and really learn what’s useful for you. That’s underemphasized a bit with agents. It’s necessary for systems to be not only useful but also safe, doing what you expected.
“Another thing is that a lot of tasks won’t require Claude to do much reasoning. You don’t need to sit and think for hours before opening Google Docs or something. And so I think that a lot of what we’ll see is not just more reasoning but the application of reasoning when it’s really useful and important, but also not wasting time when it’s not necessary.”
3/ Agents will make coding assistants better
“We wanted to get a very initial beta of computer use out to developers to get feedback while the system was relatively primitive. But as these systems get better, they might be more widely used and really collaborate with you on different activities.
“I think DoorDash, the Browser Company, and Canva are all experimenting with, like, different kinds of browser interactions and designing them with the help of AI.
“My expectation is that we’ll also see further improvements to coding assistants. That’s something that’s been very exciting for developers. There’s just a ton of interest in using Claude 3.5 for coding, where it’s not just autocomplete like it was a couple of years ago. It’s really understanding what’s wrong with code, debugging it—running the code, seeing what happens, and fixing it.”
4/ Agents will need to be made safe
“We founded Anthropic because we expected AI to progress very quickly and [thought] that, inevitably, safety concerns were going to be relevant. And I think that’s just going to become more and more visceral this year, because I think these agents are going to become more and more integrated into the work we do. We need to be ready for the challenges, like prompt injection.
[Prompt injection is an attack in which a malicious prompt is passed to a large language model in ways that its developers did not foresee or intend. One way to do this is to add the prompt to websites that models might visit.]