Perhaps the biggest news of the last week is the launch of Amazon Q, its competitor to ChatGPT, at AWS Re:Invent in Las Vegas. Amazon has been lagging in the AI race and OpenAI was able to establish a significant lead while Amazon had no competitive product for over a year. This is not a situation AWS is used to, and is unusual given its undeniable technical prowess and resources. However, it is now clear that Amazon has been thinking in a much deeper and more sophisticated fashion about what a more capable and more general AI system looks like and how it operates. The narrative that emerged around ChatGPT was that Large Language Models (LLMs) like GPT-4 were “super algorithms” capable of anything provided they were big enough (1 trillion parameters!) and trained on enough data. AI practitioners could point out that GPT-4, like all GPTs before it, is unable to reliably multiply integers and the response would be something like “well, just wait until the next release with a bigger model and more training” rather than just using a standard calculator for that kind of thing. Amazon’s approach with Q is to blend together an ensemble of AI systems that are specialized for different tasks, using different models that match the problem space. Part of this is because AWS wants to be a model broker for user traffic, but it also enables Q to do more tasks at higher accuracy while retaining the simple chat-based interface. Think Alexa vs. Alexa Skills. The result is a better general intelligence than a single model because it can harness the power of specialization. At Diffblue we have deliberately specialized our machine learning approach to write unit test code, and as a result it’s so accurate that it can code completely autonomously. In contrast, general LLMs produce unit tests that don’t work >60% of the time while also doing a very good job of synthesizing a Shakespearean sonnet. In the natural world, specialization has massive competitive benefits and Amazon Q just tilted the playing field for AI systems by using that approach. When it comes to AI, bigger is not always better; sometimes the power of specialization prevails. [Googley-eyed Shakespeare image below via Jasper.AI]
Gustaf Cavanaugh
8mo
"In contrast, general LLMs produce unit tests that don’t work >60% of the time while also doing a very good job of synthesizing a Shakespearean sonnet." 😂 Thought provoking and funny -- you haven't lost your tech journalist touch
Anne Johnson
8mo
Specialization is going to be the key to useful applications; defining the proper scope for the relevant vertical and hence the scope of the specialization is going to be critical, and require some experimentation. Very interested to see how Diffblue approaches this.
Raha Rotabi
8mo
I agree and think the same is true even within AI models, i.e. DallE will create gibberish text in images rather than use something like greedy search.
Kevin Bailey
8mo
Nice overview qf quality vs quantity
20+ yrs in Sales and Marketing Automation | Microsoft MVP | Improving Executive Visibility into Pipeline | Co-author of Implementation as a Subscription Service Model | COO | Entrepreneur
8moThe question is when would Alexa get it