It seems like every day brings new announcements in generative AI. Two weeks ago, I was at AWS re:Invent, where Amazon had no shortage of news around artificial intelligence & machine learning (AI/ML). The star of the show was Amazon Q — a new B2B chatbot they called an application that “manifests” in different ways. However, I consider it to be much more of a platform rather than an application, which we’ll dig into later in this piece. Amazon also had a variety of other AI/ML announcements from its latest training chip to SageMaker (for training & deploying custom ML models) to Bedrock (its foundation-model service) and beyond.
Not even a week later, Google announced its new Gemini models. From the pre-announced Gemini Ultra, to the now-available Gemini Pro, to the phone-sized Gemini Nano, Google’s done a lot of work (and spent a lot of money!) to not only catch up but even pull slightly ahead.
Amazon Q: an application, or a platform?
Amazon Q is, in short, a chatbot app (in Amazon’s view) but more of a chatbot platform, in my opinion. It “manifests” in different ways, such as embedded within the AWS management console, within VS Code, within Amazon’s business-intelligence app QuickSight, and more. Although it’s shown mixed results to date, improving quality is only a matter of time. Those manifestations, to me, are the first clear indication that it’s a platform rather than an application — even if that platform might be internally facing.
If we look further, however, we see that Q has multiple “connectors” to interact with a variety of corporate data sources from Amazon’s own to databases to ServiceNow, Salesforce, Slack, and more. That also includes a custom connector, meaning you could index basically any data that you can write the code for.
In addition, it supports “plugins” to take action, such as within Jira or ServiceNow. There’s no custom plugins yet, but you could certainly imagine them.
Finally, there’s an API for Q. This opens up the potential for embedding its intelligence into other applications besides those Amazon explicitly decides to support, including your own in-house custom apps.
Coupled with Amazon’s clear emphasis on protecting its customers’ corporate data, this opportunity for flexible incorporation to solve customer problems more effectively is a very compelling look at the future, for when Q becomes GA.
Amazon’s approach to “best tool for the job” with a large set of models in Bedrock (and consumed by Q) contrasts with the dedicated providers such as OpenAI and Anthropic, as well as other cloud providers such as Microsoft and Google that are focused largely on a single model from themselves (Google) or a strategic partner (Microsoft with OpenAI).
Amazon’s philosophy holds promise not only in the context of doing better at a variety of use cases where some models are better suited than others, but also to improve efficiency, performance, and cost. Why use a massive, expensive model when a simpler one could do? Why use a model best-suited for prose when generating code, or vice versa?
Its biggest challenge in this front will be aiding its customers in the appropriate selection and combination of all these models. Amazon has historically not done a strong job of helping customers parse through the complexity of its offerings, so this will be a steep barrier to surmount.
A closer look at Gemini
Google’s Gemini Ultra incrementally advances the state of the art over OpenAI’s latest GPT-4 with Vision (GPT-4V). Across a variety of benchmarks, it shows improvements generally in the range of a couple of percentage points (e.g. from 52 to 54, or 90 to 92).
Gemini Pro, available last week in Bard and this week via API, is the best you can get at the moment. Its performance significantly surpasses GPT-3.5 and Meta’s Llama 2, while being competitive with Anthropic’s Claude 2. As for availability in Google’s other services such as Duet AI for working directly within your docs and emails, you’ll have to wait for “the coming months.”
I see the incorporation of Gemini into Google Workspace as a particularly large opportunity for Google. Considering the dominance of Gmail for personal use and in startups & education, this is a great opportunity for them to build increased stickiness with both users and companies early in their lifetimes or their company’s growth, and increase the retention and conversion rates of Google Workspace customers as a result.
Price pressure and moving up the stack
However Google’s still using the price point of $20/month for GPT-4 access as their anchor and raising it even higher ($30/user/month), just as Microsoft has done with Copilot for Office 365. This looks prohibitive at scale when compared to the base Google Workspace or Office 365 cost ($12–22/user/month for business plans). Expecting business customers to pay almost triple the price just to add a productivity plugin, when they’re looking at what they’ve paid for an entire SaaS-based office suite, seems very challenging to scale across their customer base.
Recent news from Mistral and Together.ai (offering Mistral’s own new model at a lower price) continue to create price pressure at the API layer. This is going to push differentiation and value creation further up the stack into business applications and use cases. That’s a place where IBM has historically been extremely strong, and their partnership with HuggingFace opened up great possibilities for a “best tool for the job” approach as well. I expect them to continue their strength in understanding the world’s largest enterprises as they apply generative AI.
Key takeaways
- Amazon Q has the potential to become a chatbot platform for custom, secure enterprise data, but it’s not there yet in performance or capabilities.
- Google Gemini has a Pro model now available that’s compelling in comparison to freely available GPT 3.5 and Llama 2, but Anthropic’s Claude 2+ outperform it. Google Workspace and GCP customers should consider it, if they’re focused on a strategic vendor strategy or have risk concerns about earlier-stage startups.
- Prices for AI in business applications remain very steep, so value will need to be proven at scale to drive enterprise-wide purchasing decisions.
- Innovation remains rampant in the foundation models, and price pressure from competition as well as open-source models will continue to create new price floors. This is good for consumers who are flexible on where they get their LLM chatbots and have low switching cost.
Disclosures: Amazon covered my travel costs & a portion of onsite expenses to attend AWS re:Invent.
Platify Insights 