Both major systems, Azure and Google, are not yet stabilized, especially their customer support. It is essential to have a robust customer support system to get quick resolutions as they're frequently making changes. Working on multiple indexes is also a challenge with both systems.
It should improve the model efficiency, which is just training the model over and over again. Building an application on Agent Builder and grounding it with our internal documents and internal data does improve its accuracy when used in generative AI. Some of the tools should have more advanced settings available. Some are very locked into certain features and settings, and there is no customization. A more advanced button to fine-tune some of those settings would be good. Having it more available throughout all the other applications on GCP would be good.
I believe that Vertex AI is a robust platform, but its effectiveness depends significantly on the domain knowledge of the developer using it. While Vertex AI does offer support through the console UI in the Google Cloud environment, it is better suited for technical members who have a deeper understanding of machine learning concepts. The platform may be challenging for business process developers (BPDUs) who lack extensive technical knowledge, as it involves intricate customization and handling of numerous parameters. Effectively utilizing Vertex AI requires not only familiarity with machine learning frameworks like TensorFlow or PyTorch but also proficiency in Python programming. The complexity of these requirements might pose challenges for less technically oriented users, making it crucial to have a solid foundation in both machine learning principles and Python coding to extract the full value from Vertex AI. It would be beneficial to have a streamlined process where we can leverage the capabilities of Vertex AI directly through the BigQuery UI. This could involve functionalities such as creating machine learning models within the BigQuery UI, providing a more user-friendly and integrated experience. This would allow users to access and analyze data from BigQuery while simultaneously utilizing Vertex AI to build machine learning models, fostering a more cohesive and efficient workflow.
While it employs chat activity for answering queries, the available options might be somewhat limited. I've noticed that using chat activity often presents a broader range of options and insights for a well-constructed question. Improving the knowledge base could be a key aspect for enhancement—expanding the information sources to enhance the generation process.
It would be beneficial to have certain features included in the future, such as image generators and text-to-speech solutions. They could offer some simplified system to train the samples other than LLM models.
Both major systems, Azure and Google, are not yet stabilized, especially their customer support. It is essential to have a robust customer support system to get quick resolutions as they're frequently making changes. Working on multiple indexes is also a challenge with both systems.
It should improve the model efficiency, which is just training the model over and over again. Building an application on Agent Builder and grounding it with our internal documents and internal data does improve its accuracy when used in generative AI. Some of the tools should have more advanced settings available. Some are very locked into certain features and settings, and there is no customization. A more advanced button to fine-tune some of those settings would be good. Having it more available throughout all the other applications on GCP would be good.
The tool's documentation is not good. It is hard.
I believe that Vertex AI is a robust platform, but its effectiveness depends significantly on the domain knowledge of the developer using it. While Vertex AI does offer support through the console UI in the Google Cloud environment, it is better suited for technical members who have a deeper understanding of machine learning concepts. The platform may be challenging for business process developers (BPDUs) who lack extensive technical knowledge, as it involves intricate customization and handling of numerous parameters. Effectively utilizing Vertex AI requires not only familiarity with machine learning frameworks like TensorFlow or PyTorch but also proficiency in Python programming. The complexity of these requirements might pose challenges for less technically oriented users, making it crucial to have a solid foundation in both machine learning principles and Python coding to extract the full value from Vertex AI. It would be beneficial to have a streamlined process where we can leverage the capabilities of Vertex AI directly through the BigQuery UI. This could involve functionalities such as creating machine learning models within the BigQuery UI, providing a more user-friendly and integrated experience. This would allow users to access and analyze data from BigQuery while simultaneously utilizing Vertex AI to build machine learning models, fostering a more cohesive and efficient workflow.
Google Vertex AI is good in machine learning and AI, but it lacks optimization.
While it employs chat activity for answering queries, the available options might be somewhat limited. I've noticed that using chat activity often presents a broader range of options and insights for a well-constructed question. Improving the knowledge base could be a key aspect for enhancement—expanding the information sources to enhance the generation process.
It would be beneficial to have certain features included in the future, such as image generators and text-to-speech solutions. They could offer some simplified system to train the samples other than LLM models.