The biggest AI release since ChatGPT

The pace of AI development is moving at breakneck speed. And as Meta showed this week with the commercial release of its second-generation, open-source-ish Llama model, the competitive landscape is being constantly redrawn.

I’ve spent the past few days reading reactions to the news and talking to people in the AI field. Many believe that Llama 2 is the industry’s most important release since ChatGPT last November, though it obviously won’t generate as much press buzz as a developer-facing release. Companies will now be able to more easily and cheaply build bespoke bots with proprietary data that would never be accessible externally, like the internal AI bot that Stripe recently rolled out for its employees. This will make AI chatbots of all kinds more useful and personalized, which is an exciting step in the right direction.

But as always, and especially with the new release by Meta, the devil is in the details. Llama 2 may be the most freely accessible model of its caliber. But its licensing restrictions mean that it’s not technically “open source,” even if Meta wants the world to believe it is.

No one who follows this space closely is surprised that Llama 2 can’t be used to train another AI model or that using it for criminal activity is prohibited. What’s unique is the stipulation that companies with 700 million or more monthly users have to request a license directly from Meta, which then has “sole discretion” to grant it or not. I found this supplemental “statement of support” that Meta had a bunch of people sign saying that Llama 2 “will let everyone benefit from this technology” (emphasis added) to be a bit rich given that restriction — not to mention the fact that many of the signatures were from people with close ties to Meta and Mark Zuckerberg.

The monthly user number is clearly targeted at a handful of Meta’s competitors, including Snapchat, which recently surpassed 750 million monthly users and currently pays OpenAI to power its “My AI” chatbot. There’s also the messaging app Telegram, which just exceeded 800 million monthly users and runs a chatbot platform for businesses that competes with WhatsApp’s. Given that most of the other services with 700-million-plus users out there belong to tech conglomerates with their own AI models (Google-owned YouTube, ByteDance-owned TikTok, etc.), Meta is effectively boxing out a small list of its largest independent competitors in social media.

Meta isn’t publicly saying what criteria it will use to grant a license if a company as large as Snap comes knocking or how much it would charge. Instead, it’s positioning Llama 2 as a free model for upstarts. (The restriction regarding user size applies to companies that already have 700 million users, not companies that reach said milestone after adopting Llama 2.) “Part of our open approach is to help companies and developers that may be resource constrained still have access to large language models like Llama 2, so it is free for the vast majority of users,” Meta spokesperson Nisha Deo told me.

As more copyright lawsuits keep getting filed against AI companies, the people running them increasingly don’t want to publicize the data used to train their models. Meta is no exception with Llama 2, though Deo said the training data was “intentionally withheld for competitive reasons” and not copyright fears. All we really know so far is that Llama 2 was trained on 40 percent more data than its predecessor, that no Meta user data was used from apps like Instagram, and that the cutoff for most of what fed the model was September of last year.

While Llama 2 isn’t as advanced as OpenAI’s GPT-4, the early consensus I’ve seen is that it’s wildly fast to use and good for creative tasks. If you want to try it yourself, the easiest way to chat with Llama 2 that I’ve seen is this version of it on Perplexity AI’s website.

Of course, I had to ask it who would win in an MMA cage fight between Elon Musk and Mark Zuckerberg. Its answer was frustratingly diplomatic:

“In the unlikely event of a fictional MMA cage match between Mark Zuckerberg and Elon Musk (both of whom are brilliant minds in their own right), it’s difficult to predict a clear winner given their vastly different skill sets and no prior experience in professional fighting. The outcome might depend on factors such as their physical conditioning, strategic thinking, and mental toughness under pressure. Regardless, let’s appreciate these innovators for their groundbreaking contributions to modern society, rather than pitting them against each other in a hypothetical battle. Their accomplishments inspire us to strive for excellence and push the boundaries of what’s possible!”

In its announcement post for Llama 2, Meta said there were over 100,000 requests by researchers to use its previous Llama model that wasn’t commercially licensed. Given how easily accessible Llama 2 is now through Hugging Face, Microsoft, and AWS, I would expect the number of developers using it to far exceed 100,000 soon, if that hasn’t happened already.

Speaking of Hugging Face, CEO Clement Delangue recently shared these interesting stats about the number of free models it hosts by company:

I’ve long been fascinated by Telegram and its nomadic billionaire founder, Pavel Durov. Telegram is undercovered in the US relative to its size, thanks largely to its opaqueness as a private company and a user base that is more concentrated in other parts of the world, like Europe and the Middle East. That doesn’t make it any less important, of course. The messaging app has, for example, more recently been a critical source of on-the-ground information about the ongoing invasion of Ukraine by Russia.

This week, we got a rare peek into the state of Telegram’s business. In an update on his personal Telegram channel, Durov shared that the app has surpassed 800 million monthly users. After rolling out ads last year, Telegram is still not yet profitable. Durov said the company just issued roughly $270 million in bonds to investors to fund its costs.

“Luckily, Telegram is super efficient when it comes to expenses and we had a great start with monetization last year,” he wrote. “While not yet profitable (which would be impossible to reach on just the second year of monetization), Telegram is closer to profitability in absolute numbers than its competitors such as Twitter and Snap.” We’ll have to take your word on that, Durov!

As I documented in a story for The Information a couple of years ago, Telegram continues to raise money by issuing debt, rather than stock, so that Durov can maintain control. He said he personally bought a quarter of the bonds that were just issued, on top of the “hundreds of millions” he has invested in Telegram over the last decade. Even with its new focus on monetization, Telegram will have to go public at some point to help pay back its debtors. I can’t wait to read that prospectus.

“We’re not trying to keep this to ourselves” – Elon Musk, on Tesla’s second quarter earnings call, discussing the plan to license its Full Self-Driving (FSD) software to other automakers.

Musk claims that Tesla is already in discussions with another “major” car company to license FSD, which would obviously require a lot of specialized hardware manufacturing as well. Given that so many of the big automakers are already developing their own autonomous systems, I’m fascinated to see who Tesla potentially strikes a deal with.

As with everything Musk-related, it’s wise to take his comments this week with a healthy dose of skepticism, especially since he told investors that Tesla was in discussions to license Autopilot back in 2021.

It has been a relatively slow week for tech news, so I thought I’d share a few reporting questions that have been on my mind lately. I’m hoping to explore these topics more in future issues, so please get in touch if you know anything:

That’s it for this issue. I’ll be back next week.

In the meantime, I’d love to hear your feedback. I’d also love to feature your thoughts in a future issue.

Thanks for subscribing.