Anthropic published a profile on what we're building at Kepler. Sharing because the architectural argument (LLM for intent, deterministic code for retrieval and computation, every number traceable to source) is the part I'd actually want HN to push on. Happy to answer questions in the thread.
I'm on a very similar train. You cannot dump all the data into an LLM (for many reasons) and we also already have clearly defined rules that an LLM doesn't have to figure out.
So keep organizing data (LLM powered, of course), so that you can query data as usual (multi modal, so not just graphs, but also time series, relational, etc). Feed that to deterministic computations. Let an LLM reason about the outcomes.
Give the LLM the freedom to orchestrate the retrieval and computations. Make sure the way it orchestrates it is auditable.
The key thing I want to achieve is beyond this system: I want to uncover hidden things in the system (missing in the ontology, computations, etc) and propose to add these. This will effectively give you a generic approach to create ever evolving systems aliging with reality while being fully auditable.
From a systems engineering standpoint, the purpose of LLMs is to construct, verify, and "push down" abstractions and deterministic layers. Deterministic layers are able to cope reliably with the law of medium numbers.
The title is misleading. They achieved a 94% accuracy rate which in financial services is a far cry from acceptable without a human-in-the-loop verifier.
Reassuring to see this approach coming out consistently. I’ve been doing the same for high volume data pipelines, extracting the deterministic actions from markdown instructions and leaving the LLM to do the analysis/act as the fluid coupling between deterministic parts.
Over time you can refine this to be more and more codified, handle edge cases with agents/LLMs then turn them into first class deterministic branches too.
This pattern seems to be emerging everywhere, the chain of thought and intent capture to improve it seems to be the next big thing
The idea is to take markdown instructions and "compile" them into a Prolog-based DSL that orchestrates both deterministic and LLM-based components. The (meta-)interpreter of the DSL automatically tracks the entire execution process, so that the final ouput becomes observable and more explainable. Still at an early stage, but I am having lots of fun with it and would love to explore possible use cases.
13 comments
[ 4.7 ms ] story [ 41.4 ms ] threadSo keep organizing data (LLM powered, of course), so that you can query data as usual (multi modal, so not just graphs, but also time series, relational, etc). Feed that to deterministic computations. Let an LLM reason about the outcomes.
Give the LLM the freedom to orchestrate the retrieval and computations. Make sure the way it orchestrates it is auditable.
The key thing I want to achieve is beyond this system: I want to uncover hidden things in the system (missing in the ontology, computations, etc) and propose to add these. This will effectively give you a generic approach to create ever evolving systems aliging with reality while being fully auditable.
On the one hand, very encouraging to see plain old deterministic infra w/o using slop machines.
On the other hand, this is a recognition that LLMs are just additional friction in the system that we would better off without in the first place!
We are living in an age of hot air.
But the https://kepler.ai website says 10M+
Over time you can refine this to be more and more codified, handle edge cases with agents/LLMs then turn them into first class deterministic branches too.
This pattern seems to be emerging everywhere, the chain of thought and intent capture to improve it seems to be the next big thing
The idea is to take markdown instructions and "compile" them into a Prolog-based DSL that orchestrates both deterministic and LLM-based components. The (meta-)interpreter of the DSL automatically tracks the entire execution process, so that the final ouput becomes observable and more explainable. Still at an early stage, but I am having lots of fun with it and would love to explore possible use cases.