Tuesday, July 1, 2025

Not all the things wants an LLM: A framework for evaluating when AI is smart


Be part of our each day and weekly newsletters for the newest updates and unique content material on industry-leading AI protection. Be taught Extra


Query: What product ought to use machine studying (ML)?
Undertaking supervisor reply: Sure.

Jokes apart, the appearance of generative AI has upended our understanding of what use circumstances lend themselves greatest to ML. Traditionally, we’ve all the time leveraged ML for repeatable, predictive patterns in buyer experiences, however now, it’s doable to leverage a type of ML even with out a complete coaching dataset.

Nonetheless, the reply to the query “What buyer wants requires an AI answer?” nonetheless isn’t all the time “sure.” Giant language fashions (LLMs) can nonetheless be prohibitively costly for some, and as with all ML fashions, LLMs should not all the time correct. There’ll all the time be use circumstances the place leveraging an ML implementation will not be the fitting path ahead. How will we as AI mission managers consider our clients’ wants for AI implementation?

The important thing concerns to assist make this resolution embrace:

  1. The inputs and outputs required to satisfy your buyer’s wants: An enter is supplied by the shopper to your product and the output is supplied by your product. So, for a Spotify ML-generated playlist (an output), inputs might embrace buyer preferences, and ‘preferred’ songs, artists and music style.
  2. Mixtures of inputs and outputs: Buyer wants can differ based mostly on whether or not they need the identical or totally different output for a similar or totally different enter. The extra permutations and mixtures we have to replicate for inputs and outputs, at scale, the extra we have to flip to ML versus rule-based methods.
  3. Patterns in inputs and outputs: Patterns within the required mixtures of inputs or outputs assist you to determine what sort of ML mannequin it’s essential to use for implementation. If there are patterns to the mixtures of inputs and outputs (like reviewing buyer anecdotes to derive a sentiment rating), think about supervised or semi-supervised ML fashions over LLMs as a result of they is perhaps cheaper.
  4. Value and Precision: LLM calls should not all the time low-cost at scale and the outputs should not all the time exact/actual, regardless of fine-tuning and immediate engineering. Typically, you might be higher off with supervised fashions for neural networks that may classify an enter utilizing a hard and fast set of labels, and even rules-based methods, as a substitute of utilizing an LLM.

I put collectively a fast desk beneath, summarizing the concerns above, to assist mission managers consider their buyer wants and decide whether or not an ML implementation looks as if the fitting path ahead.

Kind of buyer wantInstanceML Implementation (Sure/No/Relies upon)Kind of ML Implementation
Repetitive duties the place a buyer wants the identical output for a similar enterAdd my e-mail throughout numerous varieties on-lineNoMaking a rules-based system is greater than enough that can assist you along with your outputs
Repetitive duties the place a buyer wants totally different outputs for a similar enterThe shopper is in “discovery mode” and expects a brand new expertise once they take the identical motion (resembling signing into an account):

— Generate a brand new art work per click on

StumbleUpon (keep in mind that?) discovering a brand new nook of the web by means of random search

Sure–Picture technology LLMs

–Advice algorithms (collaborative filtering)

Repetitive duties the place a buyer wants the identical/comparable output for various inputs–Grading essays
–Producing themes from buyer suggestions
Relies uponIf the variety of enter and output mixtures are easy sufficient, a deterministic, rules-based system can nonetheless be just right for you. 

Nevertheless, in case you start having a number of mixtures of inputs and outputs as a result of a rules-based system can not scale successfully, think about leaning on:

–Classifiers
–Matter modelling

However provided that there are patterns to those inputs. 

If there aren’t any patterns in any respect, think about leveraging LLMs, however just for one-off eventualities (as LLMs should not as exact as supervised fashions).

Repetitive duties the place a buyer wants totally different outputs for various inputs –Answering buyer assist questions
–Search
SureIt’s uncommon to return throughout examples the place you’ll be able to present totally different outputs for various inputs at scale with out ML.

There are simply too many permutations for a rules-based implementation to scale successfully. Contemplate:

–LLMs with retrieval-augmented technology (RAG)
–Resolution bushes for merchandise resembling search

Non-repetitive duties with totally different outputsEvaluate of a resort/restaurantSurePre-LLMs, one of these state of affairs was tough to perform with out fashions that had been skilled for particular duties, resembling:

–Recurrent neural networks (RNNs)
–Lengthy short-term reminiscence networks (LSTMs) for predicting the following phrase

LLMs are a fantastic match for one of these state of affairs. 

The underside line: Don’t use a lightsaber when a easy pair of scissors might do the trick. Consider your buyer’s want utilizing the matrix above, taking into consideration the prices of implementation and the precision of the output, to construct correct, cost-effective merchandise at scale.

Sharanya Rao is a fintech group product supervisor. The views expressed on this article are these of the creator and never essentially these of their firm or group.


Related Articles

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Latest Articles