System One + Jev – TypeSafe AI

A Modern Solution for a classical problem

What is Jev

TypeSafe AI says this is a system 1 model quoting the thinking patterns from famous book of Daniel Kahneman

For this context it is good to understand ,

“System 1” is fast, instinctive and emotional; “System 2” is slower, more deliberative, and more logical

So it rules out systems own reasoning efforts with any custom chain-of-thought datasets during its training , More than that its,

• Terse – Output tokens are simply free!
• Non Hallucinating – speaks only from finite set of given options
• single parallel pass – Non recursive and faster

It is said to have trained in “Reinforcement Learning for Calibrated Decisions (RLCD)” But Actually there is no explanations and TBH, no one except for TypeSafe AI knows what it means. Most of the online explanations are interpretation of docs and trying to read between the lines of the word “Calibrated”. For implementation the following understanding pivotal i feel.

The Classical Problem

Machine Learning for classification and regression. Supervised learning where we take “Structured data” or use the Unstructured data and preprocess it and get it structured , share it among many models and teams as feature stores , version them with a clean DVC , and build model with model pipelines, run nightly for training , Track the drift of model and data and then monitor it for continual optimization ? .. Thats the classical problem this is solving but ,

• Without any Training
• Without any Explanation
• No control on Model updates , so no way to measure drifts over time
• With no way (untill now) to add seed to tame the stochasticity

Also , one can bolt on some MLops on top of JEV , but it will be more of O&M than MLops.

Inputs and Outputs

Inputs are all of your data , if you come from LLM its all of your context , Throw all your pdfs via Docling and bundle your texts to what TypeSafe AI calls input as “State” . Its also easy to see they are adding multimodal states to this model types in future.

Output Types says everything about this approach,

• Choice
  • A multi class classification equivalent along with prob scoring
• Score
  • A regression equivalent along with prob scoring
• Noul
  • Basically Bool output

Seeing this as a simple function call is easier to get it to implement , seeing this as a Remote Procedure Call is much better to design network loading patterns and handling exceptions.

Why this will be popular

• Cheaper
• Faster
• Brute force successes – Throw all unstructured data into it and it will do its job. This will get any app to prod faster.
• No need to train or finetune!

Best Practices if JEV is used in prod

• Use Probabilistic scores instead of end to end choices and cook your own thresholds and custom logics.
  • This helps your code to document and leave breadcrumbs in your Otel setup of what happened and why it happened.
• Avoid Nouls- this makes most of the logic and explainability opaque
• In Classfication tasks, Always use a “I dont know” as a choice to allow the model to breathe .
• After you collect enough data about your application pattern attempt to see if simple ML models can be used for the problem as now there will be labelled data from Jev.
• Agentic use cases is where this shines ! .. Quick check to find the right tool call , quick decisions to get into and out of a loop

Donts

• Dont hand over your business critical “states” to JEV to quickly get results. As of now there is no GDPR or HIPAA compliance mentioned out clearly. So dont do it , atleast Yet
• Dont Promote any QA implementation without smoke test with selected few dataset to prod as you do not have control over the model versions as of now.