Question: What product should machine learning (ML) use?
Project manager answer: Yes.
Apart from the jokes, the arrival of generative artificial intelligence raised our understanding, what cases of use ML are best suited. Historically, we have all the time used ML for repetitive, predictive patterns of customer experience, but now it is possible to make use of the ML form, even without a whole set of coaching data.
Nevertheless, the answer to the query “What need for customers need AI solutions?” It is still not all the time “yes”. Large language models (LLM) might be excessively expensive for some, and as with all ML models, LLM is not all the time accurate. There will all the time be cases of use in which the use of ML implementation is not the right path forward. How do we as AI managers assess the needs of our clients in the implementation of artificial intelligence?
Key considerations that may assist you make this decision include:
- Entrances and outputs required to fulfill the customer’s needs: The customer is delivered by the customer to your product and the output is delivered by your product. Thus, for the list of playback generated by Spotify ML (output), the input data may include customer preferences and “you like” songs, artists and music species.
- Combinations of entries and outputs: Customers’ needs may vary depending on whether or not they want the same or other way for the same or other entrance. The more permutation and mixtures we have to copy to enter and exit, the more we have to show to ML -based systems in comparison with the rules.
- Designs in entrances and outputs: Designs in the required mixtures of entrances or outputs help to come to a decision what kind of ML model you might want to use to implement. If there are patterns of input data and outputs (equivalent to review of customer anecdotes in order to acquire a results of sentiment), consider supervised or partially supervised ML models by LLM, because they might be more profitable.
- Cost and precision: LLM connections are not all the time low-cost on a scale, and outputs are not all the time precise/accurate, despite tuning and fast engineering. Sometimes it is higher with supervised neural networks models that may classify the input data using a everlasting set of labels and even systems based on rules, as an alternative of using LLM.
I submitted a short table below, summarizing the above considerations to assist project managers in assessing their clients’ needs and determining whether the implementation of ML appears to be the right path forward.
| Type of customer needs | Example | ML IMPLEMENTATION (yes/no/depends) | ML IMPLEMENTATION TYPE |
|---|---|---|---|
| Repetitive tasks in which the customer needs the same output for the same entrance | Add my e -mail in various online forms | NO | Creating a system based on rules is good enough to assist you with the results |
| Repetitive tasks in which the customer needs different exits for the same entrance | The client is in the “discovery mode” and expects a recent experience when he takes the same motion (for example, signing an account): – generate recent artistic endeavors to click – –Stumbleupon (Do you remember it?) Discovering a recent corner of the Internet by randomly searching | Yes | -Mimage generation llms -HomMendation algorithms (Filtering cooperation) |
| Repetitive tasks in which the customer needs the same/similar output for different input data | – Essemus -Degenering topics from customer opinions | Depends | If the variety of input and output mixtures is easy enough, deterministic, the system based on the rules can still work for you. However, if you begin to have a lot of mixtures of input data and outputs, because the system based on the rules cannot scale effectively, consider the tilt: – Classifiers But only when these entrances are patterns. If there are no patterns at all, consider the use of LLM, but only for one -off scenarios (because LLM is not as precise as supervised models). |
| Repetitive tasks in which the customer needs different exits for various input data | – Dringing questions about customer support -In | Yes | It rarely occurs on examples where you possibly can provide different options for different input data on a large scale without ML. There is just too much permutation for the implementation based on the rules to be effectively scaled. To consider: -LLMS with a recovered generation (RAG) |
| Non -relieving tasks with various outputs | Hotel/restaurant review | Yes | Initially, any such scenario was difficult to attain without models that were trained to specific tasks, equivalent to: -ReCrandrent Neural Networks (RNNS) LLMS matches perfectly with any such scenario. |
To sum up: do not use a light sword when a easy steam of scissors can do the trick. Evaluate the client’s need with the above matrix, taking into account the costs of implementation and precision of results to build accurate, profitable products on a large scale.
.