Perplexity breaks out, becomes a unicorn

Perplexity, the most successful of the AI-native search startups, is having a breakout moment. As Google stumbles badly at putting AI into its search results (see story below), Perplexity is dazzling users with its just-announced ability to answer questions with custom-made shareable websites, while at the same time it is raising a rumored $250+ million at a $2.5 - $3 billion valuation. Doubters (including moi) have pointed out that Perplexity is primarily a thin wrapper of OpenAI’s AI joined to a thin wrapper of Google’s search. What, we asked, was Perplexity’s secret sauce? Now we know: Perplexity can walk and chew gum at the same time, able to put AI and search together in compelling ways, something that neither OpenAI or Google have yet achieved.

Perplexity answers your questions with shareable web pages

Perplexity raising $250 million at a valuation of $2+ billion

Google’s AI search stumbles, is put on the back burner

Frantically chasing OpenAI’s lead in AI while trying to preserve its own lucrative hegemony in search, Google keeps stepping on rakes and smacking itself in the face. The latest episode in “What did Google do wrong today?” revolves around the company’s attempt to put its newly released Gemini AI into Google searches. In addition to the usual list of ad-heavy links, a Google search can return an AI-generated summary of the results, a feature called AI Overview. Google began rolling out AI Overviews a week or so ago, and immediately there was trouble. AI Overview began returning some bizarre, even dangerous results, such as urging the user to eat rocks, or to use Elmer’s Glue to keep cheese on a pizza. In the firestorm that followed, Google at first staunchly defended the quality of the “vast majority” of its AI Overview results, but has now quietly reduced the feature’s use to a small minority of searches while they try to iron the bugs out.

Google’s AI search stumbles, is put on the back burner

Amazon uses AI to prevent shipping defective merch

Amazon ships a whole lotta stuff - estimated at 3.5 million packages per day in the US. Inevitably, some of that stuff is damaged or defective. To prevent damaged merchandise from being shipped to the customer, Amazon has rolled out an AI system it calls Project P.I. (for “private investigator”, get it?). Project P.I. works by having each item to be shipped run through an imaging tunnel, where optical character recognition scanners check the expiration date, while AI-powered computer vision models check the color to make sure it is what the customer ordered, and compare the item to reference images in the product database to look for defects and damage. Any item that fails inspection by the AI is blocked from shipping. Project P.I.’s AI model is also frequently updated with feedback from customers on returned items, in an attempt to get the defect rate as close to zero as possible.

Amazon uses AI to prevent shipping defective merch

Ex-OpenAI staff call for AI safety and transparency

Another day, another drama from OpenAI. This time, some ex-employees of OpenAI spilled the tea to the New York Times about how (in their opinion) the company has run roughshod over safety concerns in the rush to monetize the technology. Then yesterday, they released an open letter ringing the alarm about AI safety more widely, and calling on all the major AI companies to establish greater protections for whistleblowers. No doubt this Strongly Worded Letter will show the AI behemoths the error of their ways, and peace and harmony will prevail - not.

Ex-OpenAI staff call for AI safety and transparency

Showrunner aims to be the ‘Netflix of AI’, allow users to create

An AI text-to-video startup calling itself The Simulation has announced the rollout of its text-to-TV-episode platform known as Showrunner. The company released a proof of concept South Park episode generator last year which went viral. Now they say they want to be the “Netflix of AI”, promising to release a bevy of AI-generated animated shows, which the user can generate additional episodes for on their platform. Whether this particular company becomes successful or not, someone is bound to find the winning formula for this idea.

Showrunner aims to be the Netflix of AI

FAA approves Amazon drone delivery expansion

Amazon announced last Thursday that they have received approval from the FAA to expand their drone delivery service in the US. The FAA has approved Amazon’s technology for piloting drones beyond visual line of sight, a key component of an expanded drone delivery service.

FAA approves Amazon drone delivery expansion

GPT 4 outperforms humans at financial statement analysis

Researchers at University of Chicago’s Booth School of Business have published a paper comparing the performance of GPT-4 at analyzing the financial statements of public companies, with the performance of human professional financial analysts. GPT-4 was given anonymized historical financial statements of actual public companies, and instructed to forecast future earnings. These earnings forecasts were compared to the actual future earnings of the company, and to the consensus forecasts of professional financial analysts at the time of the financial statements. The result was that GPT-4 made more accurate predictions of each company’s future earnings than did the contemporary human professional analysts, despite not knowing anything about the company or the industry other than what was included in the anonymized financial statements.

GPT 4 outperforms humans at financial statement analysis

Suno 3.5 now lets you craft 4 minute songs with AI

Suno, often called the “ChatGPT of music”, is a text-to-music AI model that is currently freely available on the web. In the past, users could only generate short snippets of music, but the latest upgrade, version 3.5, allows users to craft songs as long as 4 minutes. (Try it, it’s great fun.)

Suno 3.5 now lets you craft 4 minute songs

GPT-4 assesses acuity in ER patients as well as humans

Researchers at UC San Francisco tested GPT-4’s ability to triage ER patients into acuity categories, from most urgent to least urgent. The AI was presented with pairs of actual records of former ER patients, and was tasked with choosing the patient of higher severity. The AI’s performance was found to be comparable to that of human raters, as seen in the figure below.

GP-4 assesses acuity in ER patients as well as humans

AI brain maps helps decode Alzheimer’s and autism

Neuroscientists at the University of South Florida are developing an AI brain map, which they hope will help unravel the mechanisms of alternate brain function in conditions such as Alzheimer’s disease and autism.

AI brain map helps decode Alzheimer’s and autism

MGB develops AI that predicts cancer outcomes

Mass General Brigham researchers have developed an AI model that identifies patients with non-small cell lung cancer who are at risk for poor outcomes from immune checkpoint inhibitor chemotherapy, based on changes in body composition (such as loss of muscle mass) detected with routine scans used in monitoring treatment.

MGB develops AI that predicts cancer outcomes

AI uses brain implant to translate brain waves into 2 languages

An article in Nature Biomedical Engineering from May 20 details the success of researchers in translating signals from a brain implant in a longtime paralyzed patient who is unable to vocalize, into both the patient’s native Spanish, as well as the English he learned after becoming paralyzed.

AI uses brain implant to translate brain activation patterns into 2 languages