I would classify Atama’s outlook on AI as “reserved but aware.” It’s impossible to look at all of the exciting demos and influx of companies racing to add AI to their product offerings without wondering “how could Atama benefit from the use of LLMs or some form of generative AI?” At the same time our experiences have shown that there seems to be a wall somewhere between “awe at being able to ask ChatGPT to write you a catchy song about a camel living in Mariana Trench” and being able to count on business critical tasks being completed in a reliable and consistent manner.
This wall may very well disappear but it seems like today (May 2024) this wall is what many early consumer products (e.g. the Rabbit R1, the Humane Ai pin, things like Tesla FSD or Comma.ai’s openpilot, etc) are struggling with. Hallucinations manifest as bad or inconsistent data that makes you lose trust to the point of being unwilling to actually do anything important with a particular product. Or speed issues result in a significant degradation of user experience and the feeling of being part of some sort of experiment. Some really smart teams whose entire focus is on this type of problems are still struggling and so we’re hesitant to put too much emphasis on what AI could do for users of Atama.
At the same time, our reservation and skepticism around some of the big promises of AI doesn’t keep us from trying to think of practical uses for the technology. We’re trying to find small slices of Atama functionality that could be improved or made more enjoyable through the use of AI and where the inevitable and currently inherent inconsistencies introduced by current models can be controlled to avoid negatively impacting user experience.
One example of this is tooling discovery of libraries and content for a particular website. We have created a simple browser extension PoC that records information from a browsing session on a website and then passes various data points to Meta’s Llama 3 to provide a human readable breakdown of the tools and libraries most likely used on a given website.
“Guards” (from guardrailsai.com specifically) has a lot of potential and helps to ensure outputs from a LLM (again from Llama 3 in this example) conform to a particular machine readable specification. There are a number of tasks a business analyst type role needs to complete within Atama to configure source systems or business capabilities and the thought here is that some of this configuration could be potentially seeded with reasonable “best guesses” from an AI that would cut down on the time a BA would need to spend writing configuration or naming things (which can be hard!). Guardrails allows you to define a particular data structure (as in this below example) and it will help to prompt the system to generate compatible output (JSON in this case) and confirm the output conforms or take appropriate behavior (retry the model, fail, etc). So in this simple example you can define a format for a SourceSystem as below
And then write a prompt such as
> A SourceSystem that tracks customer sales. The source system uses graphql for all queries.
And the guardrail system can wrap the communication with the model and extract / confirm the structure conforms to the given data structure requirements:
transformers.js
We’re also looking at transformers.js, a way to run simple / small models completely inside of the browser without any network requests. This is most useful for giving an idea of sentiment or higher level tasks, but given the low level of complexity and relative ease in implementing some existing models even little quality of life improvements to Atama could be worth considering.
Take this simple example below of a classification problem. I typed up a description of what a particular business capability does in casual terms and the model running nearly instantly and in 100% in my browser was able to make a general guess as to what sort of category this business capability might belong to which could in term be used to create more useful auto-complete fields, automatically sort or rank a business capability inside of Atama Studio given a particular context on the page, etc.
Are the above examples life changing? We don’t think so. But we want to stay engaged with the technology and figure out if there could be “easy wins” where quality of life or quality of our customers' experiences could be improved through the low hanging fruit of AI.