DeepMind's Sparrow vs OpenAI's ChatGPT
Once again, both companies are set to fight for the lead on AI
OpenAI and DeepMind are so alike from afar yet so different when you zoom in. And the same can be said about their CEOs, Sam Altman and Demis Hassabis.
This article has two intentions: First, I want to review the recent news about Sparrow, a ChatGPT-like model that DeepMind has had up its sleeve before the actual ChatGPT was even the seed of an idea (well, maybe that's a slight exaggeration, but three months is a lifetime in AI time).
Hassabis has revealed in an interview with TIME that they’re “considering releasing … Sparrow, for a ‘private beta' some time in 2023.” Can it outperform ChatGPT? Is this how Alphabet (Google’s and DeepMind’s parent company) will compete against Microsoft?
Second, I want to explore the commonalities and differences between DeepMind and OpenAI. Sparrow was announced half a year ago—why didn't DeepMind release it then? Why is the Londoner company, despite being older (founded in 2010 vs 2015) and with a larger resource pool at its disposal, arguably less popular than OpenAI? How do Altman and Hassabis approach AI?
Let's untangle the mysteries of the two global leading AI labs.
The Algorithmic Bridge is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
How Sparrow could challenge ChatGPT
Sparrow is DeepMind's version of ChatGPT. A language model (LM) trained on internet text data and optimized for dialogue (also similar to google’s LaMDA). It leverages reinforcement learning with human feedback to improve its behavior (less bias and discrimination and safer responses than its non-reinforced counterparts).
It was first introduced as a research model in September 2022 in a paper entitled “Improving alignment of dialogue agents via targeted human judgments.” As the researchers explain in the associated blog post, Sparrow is designed to mitigate some of the typical problems of LMs, it's “useful and reduces the risk of unsafe and inappropriate answers.”
You could say the same thing about ChatGPT—OpenAI did a good job with RLHF—but it's pretty clear it was insufficient. Sparrow is no different. As DeepMind researchers acknowledge, despite the measures, "Sparrow isn't immune to making mistakes, like hallucinating facts and giving answers that are off-topic sometimes."
But one of Sparrow's key features stands out. DeepMind explicitly trained it to be able to retrieve information from the Internet to improve factuality and reliability. "Because we show answers with and without evidence retrieved from the internet, this model can also determine when an answer should be supported with evidence," they explain.
All else being equal, Sparrow would theoretically top ChatGPT as a source of information. A retrieval mechanism enhanced by the ability to select relevant sources takes an LM's natural language mastery to the next level—a must-have characteristic to revolutionize search engines.
That's how Sparrow could, if turned into a product, challenge ChatGPT's current dominance.
But it hasn't happened yet. Although ChatGPT doesn't have this ability, it's the chatbot we're all talking about. It's the one that went viral and gave OpenAI even more popularity. It's the one that Microsoft is planning to integrate with its existing offering of software products, including Office and Bing—and the reason behind the $10-billion deal.
Let me ask you this, had you heard about Sparrow before Hassabis mentioned it during his interview with TIME?
Sparrow went under everyone's radar and the reason is straightforward: Most people don't care about AI research. What they care about is whether they can use AI. OpenAI makes good use of this fact. DeepMind knows it too but then, why did they hold back? Why didn't the company leverage this edge over its main competitor?
Sparrow was conceived earlier than ChatGPT. The resulting popularity—and subsequent investor interest—that OpenAI amassed could belong to DeepMind.
Keep reading with a 7-day free trial
Subscribe to The Algorithmic Bridge to keep reading this post and get 7 days of free access to the full post archives.