Develop new experiences with the all-encompassing Vertex AI | by Ankit Sinha | Sep, 2023

Thomas Kurien, as CEO Google Cloud, is delivering on Sundar’s AI strategy with utmost intent. So, no surprises that Vertex AI overshadowed every other Google Cloud service throughout the conference.

For the uninitiated, Vertex AI is a Platform to build and deploy ML models, and now used frequently to customize LLMs that get AI apps ready in rapid time.

Integrated view to building GenAI apps

At Searce, we adopted this McKinsey framework to untangle GenAI. And Vertex AI is the focal point to bring it all together to adhere to this methodology. One does not need dozens of tools to approach AI systematically. Searce Architects and ML specialists can reduce the unknowns, draw from our battle scars and outline the path to production. Our deep expertise in each of the 6 layers is why customers love solving with us.

At the base, all serious model training and inference work need accelerators — TPUs & GPUs. The new TPU is 2X+ more price performant. Customers can now create Kubernetes node pools containing TPU Pods. All popular open source frameworks, such as Ray and Slurm, can leverage TPUs in their preferred way.

The new Cloud TPU v5e on display.

No one needs compute to become cost efficient as much as Google. Their internal demand is going to be a big driver for innovation in chip design. But then the question arises.

Does GCP have a significant AI-headstart over others? Is GenAI really going to sustain CEOs’ attention and deliver on the early hype? Will workloads of the future really be different than what gets built today? Will AI be central to all of Software?

Would an Azure, AWS or an On-Prem client switch to Google Cloud for the AI services alone?

Well, the Nvidia stock rally indicates the promise is real. Jensen Huang, made us aware that the Nvidia DGX Cloud will be available on GCP Vertex AI. Do note that this is the all powerful AI Supercomputer, and it is on Oracle and Microsoft Azure already. Also, the very popular NVIDIA H100 GPUs will be made available on A3 VMs in GCP starting soon. These chips shall continue to be elusive till 2024 end in markets outside.

Midjourney art from the Google Blog

I’m expecting the GPUs to become unavailable on GCP as soon as they get listed. Midjourney has to create art for the whole planet using GCP. And not just Midjourney, in excess of 70% of the GenAI startups are building on GCP. So the competition for such prized resources is going to be heavy.

But all is not lost! Amongst the 1000s of announcements, Compute FUTURE RESERVATIONS got hidden. This little feature is going to relieve so much anxiety. Companies can plan their Big Billion Day sales better, campaign spikes better. This Capacity Planner feature is in Private Preview currently.

Vertex AI has upgrades all the way up, not just the infrastructure layer. Let’s see what’s in store for practitioners.

Vertex AI Search and Chat features are allowing faster development cycles.

This attempt at simplification is important, as there are more software engineers than mathematicians in this world. Engineers are going to discover shorter ways to build AI apps. Mutli-Modal inputs allow for text and images as the source prompt. Multi-Turn flows keep a conversation running. It would appear like all the customer experience platforms will get awesome in no time.

Gen App Builder is now called Vertex AI Search and Conversation

Easy access to the Catalog of Large Language Models

The very handy Model Garden on Vertex AI allows power users to select from over a 100 models. This includes Google’s own 1st party of course, and also the popular 3rd party models. Meta’s Llama 2, Code Llama, TII’s Falcon and Anthopic’s Claude 2 made their debut at the Next to much cheer.

Meta’ Llama2 on Model Garden

Softbank CEO, Masayoshi Son predicts we shall Singularity by 2047. On the contrary, research by Ari Allyn-Feuer and Ted Sanders suggest the odds are below 0.4% that transformative AGI is achieved by 2043. You know which side I am rooting for! Just 72 hours at Google Cloud Next, and I believe Artificial General Intelligence (AGI) is not a hallucination.

Well processed LLMs should be able to encapsulate all of our human intelligence, and surpass it in dimensions beyond what we can process. Cloud Next did not delve into the ethics, morality or consciousness topics — so I shall also leave that out till I read more about it from AI Ethics observers.

All I can do is advocate for businesses to start arming themselves with the most potent piece of technology available today, and do it for good purposes.

The keynote underpinned the value from advancements in Google’s Med-PaLM.

“An AI Doctor for Everyone”

These custom-built models push such critical sectors forward. An AI Doctor for everyone, wouldn’t that be a gift to the world where we are short by over 7 Mn in Healthcare staff?

Google’s Codey and Imagen got some airtime to show how developers and designers respectively, will get wings.

Reduces hallucination problem plaguing LLMs by GROUNDING

Google emphasized on best practices, like grounding, that will facilitate real world acceptance. Enterprises expect their AI bots to be factual, not get too creative. Helps to have experience in the Embeddings API and Matching Engine (now called Vector Search in Vertex AI).

And these grounded LLMs that work on Enterprise Data, will serve such meaningful results on your shopping website, on your travel app, on CRM systems — everywhere.

It was initially feared that LLMs would struggle with Long Term Memory.

Vector DBs have been all the rage ever since. The Vector Search on Vertex AI is truly world class when it comes to similarity identification.

Now our favorite DBs & DWH on GCP — CloudSQL for PostGre, AlloyDB, Redis, Neo4j and even BigQuery — all support these vector embedding. Plus, they retain their original goodness for sure, don’t demand the need to re-architect your stack from scratch.

The demo in the DB Spotlight sessions was exhilarating. No Python, no SQL, no prepping the business layer, no moving data, no prerequisites of database modernisation, no knowledge of vector embeddings, heck, not even a full understanding of the underlying data-schema. Just plain English.

“Which of my users are likely to churn? Identify the most relevant Coupon and apply it on their accounts.”

It’s not going to be long before we can forget the math, forget the plumbing. Just instruct machines in plain English.

Read more here: Source link