The product must provide better documentation. I don't see a lot of documentation, particularly on the Studio feature. In general, there is not a lot of information about how to use Feature Store. I can see it there, but the documents are not very explanatory.

The solution requires a lot of data to train the model.

The training modules could be enhanced. We had to take in-person training to fully understand SageMaker, and while the trainers were great, I think more comprehensive online modules would be helpful. 

Additionally, the user manuals can be difficult to navigate without prior knowledge. We often test new features for clients in small groups, and I've heard feedback that the documentation could be more user-friendly.

Amazon SageMaker could improve in the area of hyperparameter tuning by offering more automated suggestions and tips during the tuning process. Having integrated intelligence to suggest hyperparameters would be beneficial for optimization.

Amazon SageMaker is expensive. It is more expensive than Databricks because it has a pay-as-you-go model with SageMaker.

The pricing of the solution is an issue.

In SageMaker, the monitoring does not support data captured in a protocol called protobuf since it only accepts JSON and CSV. We use protobuf to exchange messages. We have to write our own logic to serialize protobuf, which is more work, and it'd be better if SageMaker provided out-of-the-box support for that.

In SageMaker, monitoring could be improved by supporting more data types other than JSON and CSV. Amazon SageMaker can start with Protobuf since it is a popular messaging protocol.

The solution is complex to use.

Some additional functionality would be for them to provide sample end-to-end card formation templates and try to unify the setup. At the moment, as you move from one set of documentation to the next, some of the documentation is for bringing your model, some of the documentation is for SDK, some of it's for API, some of it's for command lines, and some of it's for step functions. None of the documentation seems to be end-to-end. There are gaps in the documentation. It proves to require a lot of digging from the user to figure it out. I did get through the AWS machine learning specialty certification, but that proved to be a bit superficial. Though it covered a lot of ground, it didn't have the detail one would need for Sagemaker. I am self-taught in a lot of the stuff. I could dive deeper into some code and take time to get examples running. But I was consulting a startup, and they needed to move quickly.

I was hoping SageMaker would be easier to work with because I was expecting there would be examples we could repurpose that were more complete.

The new functionality I'd like to see is Amazon tuning attention to the documentation sets and the templates.

I feel that the area around the interface in AWS is overall confusing. The payment and monitoring metrics are a bit confusing not only for Amazon SageMaker but also for the range of other products that fall under AWS, especially for a new user of the product. The tool is not simplified enough for beginners to use. From an improvement perspective, the tool needs to be simplified enough for beginners to use.

Creating notebook instances for multiple users is pretty expensive in Amazon SageMaker.

Amazon SageMaker should concentrate and get the performance of the ensemble model to be good enough for its users.

Improvements are needed in terms of performance for not all but some of the models, especially whenever we use the product for image classification or something. In general, improvements are needed on the performance side of the product's graphical user interface-related area since it consumes a lot of time for a user.

The product must improve its documentation. The documentation must be made clearer and more user-friendly. Sometimes, we run into issues with setup. However, it's not that often.

The pricing is complicated and should be simplified.

I would suggest that Amazon SageMaker provide free slots to allow customers to practice, such as a free slot to try out working with a Sandbox. This would be beneficial for newcomers, especially those who are getting into the cloud space. They could explore this area and get all of the aspects including data engineering, data recognition, and data transformation.

SageMaker would be improved with the addition of reporting services. In addition, the models available in SageMaker are not enough for most of our use cases and require customization to be useful.

AI is a new area and AWS needs to have an internship training program available. This is one place where I see this solution lagging. There is high-level training available, but when you consider that people have been working with Windows, Linux, and various applications for the past 20 years, they know those products inside and out. SageMaker, on the other hand, is a completely new tool. It can be very hard to digest.

AWS needs to provide more use cases for SageMaker. There are some, but not enough. They should collect or create more use cases and then distribute them free of charge to the customers.

I would like to see a more graphical, low-code interface that can be used to customize SageMaker.

The solution needs to be cheaper since it now charges per document for extraction. 

The product has come a long way and they've added a lot of things, but in terms of improvement I would like to probably have features such as MLflow embedded into it.

Additional features I would like to see would include, as mentioned, MLflow and ML Pipelines which are more of a feature rich support of machine learning pipelines as well as scheduling machine learning pipelines, and visualization of machine learning pipelines.  

There are other better solutions for large data, such as Databricks.

I would say the IDE is quite immature, but it is still in its infancy, so I expect it to get better over time.

The interface and the IDE could have some improvement. UX isn't bad but could be better.

Orchestration of the ML flow can be made easier (like ETL etc.)

Scalability to handle big data can be improved by making integration with networks such as Hadoop and Apache Spark easier.

Adding certain AI functionalities similar to what DataRobot or Azure AI has would be really great.