Workflow Planner
Looking into the tech that powers Alchemysts to beyond Turing Test levels
Motivation
Pushing Alchemysts to pass the Turing test and mimic human-level behaviour, especially in a context-intensive way, is what we always do. But as human-like as AI responses might seem, getting tasks done fully through generative AI at one go is still a far cry. This is exactly where our Workflow Planner comes in.
Background
We take a page out of the organizational behaviour to get work done: to collaborate, reflect, iterate and improve. For example, in most teams, work usually gets done through multiple iterations, as shown in the diagram below:
In an iteration in a team, we have the following steps:
The Junior (henceforth referred as J) receives the work.
J does an iteration of the work
J sends the work to a senior (henceforth referred as S) for review.
S reviews the work and sends it back to J. This also includes any feedback
J receives the review done by S.
If the review approves for J's work to be sent, J stops iterating further, otherwise he goes back to step 2 again.
J sends the work.
A pseudocode for a review iterations can be essentially thought of as:
def doWork(goal, previous_state, approved) -> output:
if approved:
return previous_state
else:
# Logic to reiterate
def senior(goal, output) -> approved_or_not: # The review process done by the senior
if isOptimal(goal, output) return None # Optimality check
else:
return feedback(goal, output)
def workflow(goal, MAX_RETRIES) -> work:
initial_work = doWork(goal, previous_state=None, approved=False)
work = initial_work
while not senior(goal, work) and MAX_RETRIES > 0: # Do it until the senior approves or MAX_RETRIES exhaust, whichever comes first
work = doWork(goal, work, False)
MAX_RETRIES -= 1 # Reduce the number of MAX_RETRIES
rememberWork(goal, work) # Remembers the output
return work
Well of course it's not as easy as it looks - these functionalities span across multiple segments throughout our tech stack, as shown as a high-level overview in the below figure:
The Alchemyst C1 and X1 models are the first iterations of our Alchemyst C and Alchemyst X model families respectively. You can read more about them here.
Results
An example of how different an Alchemyst's output is from a simple model query to ChatGPT is shown below. Here, we are using Maya The SDR, our first Alchemyst. Click on the picture to open it in fullscreen.
Conclusion
What this means for us
This means that we don't suffer from third-party vendor lock-ins. This implies that we can serve you results from the best models and context processing techniques as soon as they come out with better performance - you can rest assured that you will be having the latest innovations on AI directly served to you as an AI employee.
What this means for you
The adaptive and automated nature of this incremental self-learning has quite a few ramifications:
Bolster your understaffed team: You can think of Alchemysts as a junior in your team - but those that iterate and adaptively learn fast, incrementally over time.
Scale productivity at the speed of thought: This also essentially implies that the more data you have to provide to Alchemyst, the more easily it will adapt to your tasks. Our usage based pricing model allows you to scale your usage trivially.
Incremental RoI over time: The more you scale and use Alchemysts, the better the Return of Investment (RoI) gets on every task.
Data Privacy: All data you send to an Alchemyst stays between us, since we have our own in-house models.
Comparison
The prices for the GPT models has been taken directly from OpenAI Pricing Page as of 20th June, 2024.
Follows Instructions
⚠️
✅
✅
✅
Can reason through complex tasks
⚠️
⚠️
✅
✅
Context Window Size
16K
128K
128K
128K**
Gets context
❌
⚠️
⚠️
✅
Multilingual
⚠️Shaky at best
✅
✅
✅Natively multilingual, further bolstered using Indic models by AI4Bharat
Multimodal
❌
✅
⚠️Using extensions
✅
Adaptive Learning
❌
⚠️Partly through the memory feature
⚠️Partly through the memory feature
✅
Data Privacy
⚠️Through opaque data retention agreements
⚠️Through opaque data retention agreements
⚠️Through opaque data retention agreements
✅ You have full control of the data through configurable data policies
Cost (Input + Output)
$4/1M tokens
$20/1M tokens
$40/1M tokens
$3/1M tokens**
TL;DR
Alchemyst models are cheaper on a dollar-to-dollar basis than the GPT models
They can also take in much more information in their context than similarly priced models.
They are multilingual and multimodal.
They can learn adaptively.
Last updated