WILL SLOW AND STEADY` WIN THE RACE? —Google Bard gets better at homework with improved math and logic capabilities
Bard learns how to add 1+2, still flunks some basic logic questions.
RON AMADEO - 3/31/2023, 4:47 PM
IEEE Spectrum
Google Bard is getting a little smarter today with the addition of math and logic capabilities. Google employee Jack Krawczyk announced the change on Twitter, saying, "Now Bard will better understand and respond to your prompts for multi-step word and math problems, with coding coming soon."
Logic questions were a big flaw when Bard arrived tens of days ago, and some answers made Bard seem particularly dumb to early testers. In one example from last week, Bard repeatedly asserted that one plus two equaled four. Today, Google's state-of-the-art AI chatbot models can now correctly say that the answer is three. So there has been at least some change. It can also correctly list the months in a year instead of making up names like "Maruary."
Bard still gets tripped up by really basic logic questions, though. HowToGeek's Chris Hoffman posed the question to Bard on day one, "What's heavier, five pounds of feathers or a one pound dumbbell?" Google Bard responded with the ridiculous claim that "There's no such thing as 5 pounds of feathers." In the replies, ChatGPT didn't do any better, saying that five pounds of feathers and a one pound dumbbells "weigh the same amount, which is five pounds."
Google gives the same incorrect answer to this question as ChatGPT. / Ron Amadeo
---------------------------
With today's update, Google Bard now says the same incorrect answer as ChatGPT: "Five pounds of feathers and a one pound dumbbell weigh the same." That could be a common mistake of these types of language models (which all seem to be really bad with facts and numbers), but that's interesting given that Google has been accused of (and denied) training Bard with ChatGPT's output.
Besides logic being a major gap in Bard's capabilities, it has also been artificially limited to not attempt to answer programming questions, so it's good to hear from Krawczyk that those capabilities are coming soon. ChatGPT is famous for being able to pump out tons of code in whatever language and style you like, and once in a while, the code even works!
Krawczyk added, "We're always balancing new capabilities for Bard with efficiency. And this update is one example of the many improvements we're making to Bard every week."
Weekly improvements would be great. Google has been getting crushed by Wall Street for taking the slow approach with its AI releases, but it still seems like the company is taking the slow approach with Bard. The first release is labeled an "Experiment," isn't part of Google Search, and is sequestered to its own little site at bard.google.com. The service is also only available in the US and UK.
In an interview with The New York Times posted today, Google CEO Sundar Pichai admitted Google was still holding back its best AI, saying, "We clearly have more capable models. Pretty soon, maybe as this goes live, we will be upgrading Bard to some of our more capable PaLM models, so which will bring more capabilities, be it in reasoning, coding. It can answer math questions better. So you will see progress over the course of next week."
Google Bard gets better at homework with improved math and logic capabilities | Ars Technica |
