- Atomic AI, which uses machine learning to explore the potential of RNA in drug discovery, raised a $35 million series A round.
- They have developed a machine learning model that can accurately predict the structure of RNA molecules based on a limited set of data.
- They have since improved their model and are using it to pursue their own drug discovery program, which produces candidate molecules that might work to treat conditions that are drug resistant or notoriously difficult to treat.
Having written the articles, I think they’re pretty solid summaries, but brief enough that they (hopefully) pique your interest. Not everyone goes through and clicks links in an article, because who knows what’s even relevant? Are you going to open 10 tabs just to find out? Or will you try to guess from the URL? The popup seems pretty nice to me, and it only appears when you want it to:
Here, try it live on this post — you may or may not care about football salaries, but you can see it pop up and disappear. That appearance is custom, too; the length of the summary, bullets, whether and when it’s triggered, all that can be changed. So what happens when link previews get this rich AI summary treatment?
“We’re seeing like 50 percent more pageviews,” said Shrager. “I’m a pretty analytical person and I was nervous about what we were seeing, because it’s a little counter-intuitive. But Wikipedia did this, and it was very successful. It reduces the cost of exploration to the user.”
By getting a bigger hint as to what they’re hovering over is going to be, it reduces the hump of “should I click this or not?” World’s smallest hump, sure, but if you told a web publisher you could increase clicks by a single percent — let alone 50 percent — they’d jump on it. Time on site and engagement are valuable metrics, and finding ways to increment those upwards is a big part of any product manager’s job.
Different links, like affiliate links, can provide different previews, for instance pros and cons of a product summarized from the last hundred reviews. Or external links can be left naked — perhaps (just being honest) to prevent the same clicky effect from benefiting them.
Shrager noted that they’ve done a lot of work under the hood specific to making sure summaries don’t exhibit the kind of “creativity” language models are infamous for — names, dates, quotes and other things are always retained, for instance, and changing wording is limited to places where it won’t change meaning. “All of our valuable IP is all the know how and knowledge that are before and after the AI model,” he said.
Ultimately, although users benefit, Summari’s customers are now website operators. The company charges a flat fee for access to the tool and then a small usage fee.
“If your average article has 1,000 words, and you have five of them, we summarize 5,000 words and write 500, and we charge maybe 50 cents,” Shrager offered as a very loose example of the scale. “We try to make the overall price de minimis so it’s not a block to a sale — the only way we can scale fast is to figure out big distribution channels.”
After its initial partners, Summari is set to go live on a major academic publisher and large news site, both unnamed. “We’re definitely noticing FOMO across the industry,” Shrager said. “There are people sniffing around, starting to see the value. There’s a natural network effect because backlinks are being summarized. Once TechCrunch has summarized using us, you wouldn’t want to turn it off and do it all over again.”
I suggested that, with major tech players like Microsoft and Google making huge plays in AI, the company should not be surprised to get an offer or two slipped under the door. After all, summaries like these would be great in search engines or an algorithmically curated news site. But Shrager said they aren’t looking for a quick exit.
“My job is to maximize value for shareholders. If I got an enticing offer from Google or Yahoo, it might be of massive strategic benefit… I’m not foolish, I’d go for that. But everybody here is going for the big win.”