The end-of-year AI rush

Google and the rest of the tech industry are racing to release what they can, even if the announcements are half-baked.

by Alex Heath

Dec 8, 2023, 11:23 PM UTC

Alex Heath is a contributing writer and author of the Sources newsletter.

It’s safe to say that Google wanted a splashier moment for the unveiling of Gemini, its new suite of AI models that are intended to address the rise of OpenAI.

I’ve heard that multiple in-person launch events for Gemini were recently shelved due to concerns about the model’s performance. The real Gemini — the “Ultra,” multimodal model that can interact across text, visuals, and audio — still doesn’t have a public release date. Instead, Google has shipped a text-only, mid-tier “Pro” version to its Bard chatbot, with Google Cloud customers getting access to it on December 13th.

These large language models still hallucinate a lot, as early reactions to the new Bard illustrate. Google doesn’t want people to think of Bard as a search replacement, and with Gemini Pro powering it, the chatbot does appear to be better at certain tasks. But the problem of accuracy remains, leaving Bard in a particularly awkward spot thanks to its close association with Google.

The full, multimodal Gemini we saw this week turned out to essentially be a magic trick: a highly edited video designed to make the model look more impressive than it is. If Gemini really is, as CEO Sundar Pichai put it, “one of the biggest science and engineering efforts we’ve undertaken as a company,” why not wait to properly show it working in the real world?

If I had to guess, Google’s leadership is desperate to no longer appear behind in the generative AI race, and the company wanted to have something to show this year after teasing Gemini at its developer conference in the spring. But the way it demoed the multimodality by embellishing its capabilities has already backfired, with the media catching on and employees calling out the needless misdirection internally.

The stakes for Google to get Gemini right are obviously pretty high. Beyond the competitive dynamics at play, the more than 700-person team behind Gemini includes Google co-founder Sergey Brin, who came out of retirement this year to work on the project.

“My first real/long interaction with him when he started coming in to help on Gemini was staying at the office until 1am staring at a Colab, fixing a gnarly numerics bug,” Enrique Piqueras, one of the team members, posted on X (formerly Twitter). “He grasped my code in minutes. Top tier engineer, scientist, and a pleasure to work with.” Another posted: “Sergey was in with us basically every day, often pairing!”

Does Brin still have it? We’re going to find out eventually.

While we wait for the full release of Gemini, here are my notes on some of the other AI news this week:

While it didn’t get anywhere near as much attention as Gemini, Apple’s quiet release of a machine learning framework for its own silicon shows that it’s at least thinking about what more it can do and that on-device AI models will be its focus. I’m expecting a lot more from Apple on the large language model front next year. There’s a rumor floating around that teams working on generative AI have been given a deadline of next WWDC to ship something. Stay tuned…
As I predicted in a previous issue, Elon Musk confirmed that he’s raising a big round of funding for xAI, his OpenAI rival that increasingly seems designed to bail out the investors who lit money on fire for his takeover of Twitter. X Corp. investors will get 25 percent of the equity in xAI, according to Musk. Unfortunately, money can’t buy a sense of humor.
AMD is finally coming for Nvidia’s perch atop the generative AI craze with the release of its MI300 chip. The specs seem impressive, but anyone who works with this stuff will tell you that Nvidia’s real lock-in is its CUDA software platform, which AMD still doesn’t have a viable answer to.
It baffles me that those directly involved in the OpenAI board drama keep giving interviews without saying why Sam Altman was fired! I tried my best to get him to give a reason during our brief interview last week. Now, former board member Helen Toner is speaking on the record without saying much at all. There are a bunch of really important, unanswered questions here: Why did Ilya Sutskever initially vote to fire Altman? What’s his role at the company going forward? Will the board investigation, which is being led by the law firm WilmerHale, be made public? The sheepishness from OpenAI on all of this makes me think there is another big shoe to drop…

Meta’s encryption push

Another big story this week is Meta finally rolling out default encryption in Messenger, which will shield billions of messages from the eyes of law enforcement, hackers, and the company itself.

What Meta has done by fundamentally rearchitecting an app used by billions of people is no small feat, especially when you consider that it’s rolling out encryption in a way that won’t cause people to lose their chat history. Shortly after the news broke, I spoke with Ime Archibong, Meta’s VP of product for Messenger.

“I had a conversation with somebody who had worked at another one of our industry peers that has built sizable messaging experiences over the course of the last decade or two, and they were blown away that we had taken on this challenge,” he told me. Meta is initially turning on default encryption for one-on-one chats, which he said are the most popular in Messenger, before flipping it on for groups in the near future.

Now that Messenger has been encrypted, the next step for Meta will be to encrypt Instagram direct messages, Archibong confirmed. As CEO Mark Zuckerberg told me a few months ago, the goal is still to make it so that people can communicate across all of Meta’s messaging apps once they’re encrypted.

A Cybertruck owner’s reaction

Reddit co-founder Alexis Ohanian was at Tesla’s Cybertruck unveiling in Texas to receive the third customer delivery.

Having now owned it for a week, I wanted to know what he thought, so I texted him to find out:

People moves

Some interesting job moves I noticed this week:

First here: Meta’s AR division, which is responsible for its Ray-Ban smart glasses, a forthcoming neural smartwatch, and full-fledged AR glasses, has undergone a significant reorg. The team, led by Andrew Bosworth deputy Alex Himel, has shifted from a vertical, product-focused org structure to a horizontal, discipline-focused one. The new org goes as follows: experiences (aka apps) lead Ming Hua, product management lead Sue Young, Li-Chen Miller and Heidi Young leading a new platform team, and Rafa Camargo leading hardware.
Another previously unreported Meta move: Bill Jia, VP of engineering for AI infrastructure and several other key areas, recently informed colleagues that he’s leaving.
It has been a really bad week for Spotify: a few days after laying off 17 percent of the company, CEO Daniel Ek axed CFO Paul Vogel. The PR explanation: Spotify “is entering a new phase and needs a CFO with a different mix of experiences.” Translation: this is all your fault, Paul. I’m sure his poorly timed stock sales didn’t help.
Is more M&A on the horizon for Airbnb? Dave Stephenson is moving to a newly created chief business officer role, which CEO Brian Chesky is positioning as a way to help the company “expand beyond our core.” One of Stephenson’s direct reports, Ellie Mertz, will replace him as CFO.
More chaos in Muskworld: Ganesh Venkataramanan, who led Tesla’s Dojo supercomputer project, quietly left last month. The role has since been filled by Peter Bannon. Bill Chang, a principal engineer who appeared onstage at Tesla’s AI day last year, has left as well.
Jerry Dischler, Google’s VP of ads product, is leaving and being replaced by Vidhya Srinivasan.
Anima Anandkumar, Nvidia’s senior director of AI research, is leaving to “start something new.”
The VP of product design for the iPhone and Apple Watch, Tang Tan, is leaving.
David Friedberg of All-In podcast fame has become CEO of Ohalo, a gene editing startup incubated by his Larry Page-backed investing firm, The Production Board.
After a decade at Google Ventures, M.G. Siegler is moving from general to venture partner and not making new investments.