Not everything needs an LLM: A framework for evaluating when AI makes sense

Ask: Which product should the machine study (ml) use?
Project Manager Answer: Yes.

Jokes have increased our understanding that the arrival of generative AI, the best of use of the cases of use of the best. Historically, we have always accepted ML for ML repetitive, predictive examples In customer practices, but now, it is possible to use the ML form without all training information.

Nevertheless, “Which customer needs requires a Solution AI?” The answer to the question is still not always “yes”. Great language models (LLMS) can still be prohibited and as in all ML models, LLS is not always accurate. Always be used when a ML app does not move the right path. How do we assess the needs of our customers to implement the AI ​​as AI project managers?

The main considerations to help this decision are as follows:

1. Inputs and performances required to meet your customer’s needs: An entry is given to your product by the client and is provided by your product. Thus, the entrances for entries for the created playlist (output) of the Spotify ML can include client options and ‘liked’ songs, artists and music genre.
2. Input and combination of speeches: Customer needs may vary because they do not want the same or different access to the same or different access. We must contact ML against more permutations and combinations, scale, rule-based systems for entrance and results.
3. Patterns in entrances and exits: Patterns in the required combinations of entries or exits help you to decide which type of ML model you use for use for application. If there are patterns for entry and exit combinations (to obtain a partial account of customer anecdotes), consider control or semi-controlled ML models for more efficient.
4. Price and accuracy: LLM calls are not always cheap on the scale and the results are not always Accurate / accurateDespite fine regulation and fast engineering. Sometimes, instead of using an LLM, you are better with control models for neural networks that can make a fixed label set or even rule-based systems.

Summarizing the need for the customer needs of project managers and seeing a quick description, but I summarize the views, but determine that the customer needs to evaluate and implement the ML of ML.

Type of customer needs	Example	ML application (yes / no / dependent)	ML application type
Recurring tasks that a client needs the same exit for the same entry	Add my email online in various forms	No	Creating a system based on the rules is enough to help you in your speeches
Recurring tasks that a client needs different performances for the same entry	The client is in the “discovery mode” and is in the same action (such as the account of the account) waiting for a new experience: – Create a new piece of art per click –Dumbleupon (Do you remember?) Find a new corner of the Internet through casual search	Yes	– Generation LLMS -Morib algorithms (cooperation filter)
Repetitive tasks that the customer needs the same / similar exit for different entries	– Favorites Topics from customer feedback	Dependent	If the number of entries and output combinations is simple enough, the system based on the rules can still work for you. However, if you have started a number of entries and results, the rules-based system cannot be effectively scale, consider leaning: -Courseers -Popic modeling But if these entries have examples. If there is no example, consider using LLMS, but for one-time scenarios (LLS is not as accurate as controlled models).
Repeated tasks that the customer needs different performances for different entries	-Avçelerik Customer Support Questions -Ant to look for	Yes	It is rare to be found in the samples you can provide different performances for different entrances on a scale without ML. There are simply a lot of change for an application that is effectively applied in a scale order. Think: -Lms with a return generation (dwarf) Trees for products such as search
Repetitive tasks with different performances	Review of a hotel / restaurant	Yes	Pre-LLMS, this type of scenario was difficult to carry out without models taught for special tasks: -Hastropolitan Nerve Networks (RNNS) Short-term memory networks (LSTMS) to predict the next word LLS is an excellent suit for this type of scenario.

Bottom line: Do not use a lamp where a simple pair of scissors can do. Using the above matrix of your client, taking into account the cost and sensitivity of the above, rate the accuracy of the scale, taking into account the accuracy of the height and accuracy.

Shararya Rao Fintech Group is a product manager. The views shown in this article are the absence of the author and absolute companies or organizations.