Meta Llama 2 Is Here! Everything You Need to Know About OpenAI's Biggest Open Source Competitor

0:00Today on the AI Breakdown, we're looking at everything you need to know about Meta's new Llama 2. Before that on the brief, new AI regulations on the docket and a tool that can generate an entire TV show from a single prompt. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, newsletter and Discord. Welcome back to the AI Breakdown Brief, all the AI headline news you need in five-ish minutes or less. Obviously, one of the big themes for the last few weeks has been these Hollywood strikes.

0:32first the Writers Guild strike and now the Screen Actors Guild strike as well, that while about a lot of different issues, have artificial intelligence and the future of their professions right at the heart of them. Now, as if designed to put a firm point on just how scary this is for some of the talent involved in these strikes, yesterday Simulation Inc. dropped an example of their new tool, which is a generative TV and showrunner agent. The promise that they offer is creating episodes of TV shows with a single prompt. From there, their tool, ShowOne, will write, animate, direct, voice, and edit the show for them.

1:07Now the example they gave, which really hit the whole issue right on the nose, was an AI-generated episode of South Park that was all about the SAG strike. Alongside it, they released a paper called To Infinity and Beyond, ShowOne and Showrunner Agents in Multi-Agent Simulations. The abstract reads, In this work, we present our approach to generating high-quality episodic content for intellectual property using large language models, custom state-of-the-art diffusion models, and our multi-agent simulation for contextualization, story progression, and behavioral control. Now, the big new dimension that they are adding to this is basically the role of showrunner.

1:41They point out that while current generative AI systems are great at short-term or specific tasks through prompt engineering, they don't have, quote, contextual guidance or intentionality to either a user or an automated generative story system as part of a long-term creative process. They point out that this is essential to producing, quote, high-quality creative works, especially in the context of existing IPs. Now, for those of you who are interested in story and the long-term capacity of AI to write and create stories, the paper is really interesting. They discuss, for example, the slot machine effect, which they define as a scenario where the generation of AI-produced content feels more like a random game of chance rather than a deliverative creative process, and they discuss how they try to address that with this new ShowOne model.

2:23In their announcement tweet, they write, Our goal at the simulation is AGI, AIs that are truly alive, not chatbots that pop into existence when we speak, but AI people living real daily lives in simulations growing over time. We built showrunner agents and are building show one model to give our AIs infinite stories. After sharing a set of sample South Park episodes, they write, We are working with creators and will be announcing several original IP simulations with attached AI TV shows later this year. A space exploration simulation, The Prize. A satire of Silicon Valley simulation, Exit Valley.

2:55a playful detective simulation about Charlie Jupiter. They conclude, Ultimately, we think single-agent chatbots will fail because they have no lives and can't empathize. Does anyone really want endless small talk with a brain in a jar? The AI should have their own lives, and for that, we need societies of AIs. Less her, more free guy. So this popped off on Twitter. Thousands and thousands of people have shared it. 700 ,000 people have viewed the original video. Now, while some people pointed out that this was a little inopportune at the moment, given the strike happening right now, Others were just focused on the creative possibilities that are coming down the pipeline.

3:28Bilawal Sidhu writes, We're going very quickly from doing the low-level stuff to orchestrating this all at a higher level of abstraction. It will be mind-blowing. Next up on The Brief, we officially have dueling open letters. This time, a new letter signed by more than 1 ,300 experts argues that AI is a force for good and that fears around its long-term existential risks have been overblown. The letter was organized by BCS, the Chartered Institute for IT in the UK, And as the BBC describes it, signatories to the BCS letter come from a range of backgrounds, business, academia, public bodies, and think tanks, though none are as well known as Elon Musk or run major AI companies like OpenAI.

4:05Speaking of AI for good, OpenAI and the American Journalism Project have announced a partnership through which OpenAI will give$5 million in cash, along with$5 million in OpenAI API credits, to local news publishers in order to help them both shape as well as use new generative AI tools in supporting local news efforts. OpenAI CEO Sam Altman says, We proudly support the American Journalism Project's mission to strengthen our democracy by rebuilding the country's local news sector. This collaboration underscores our mission and belief that AI should benefit everyone and be used as a tool to enhance work.

4:37Now, this comes a week after OpenAI announced a two-year deal with the Associated Press to use AP content to help train OpenAI's model. Meanwhile, other early attempts to use AI-generated content in publishers haven't gone so well. Geo Media, that owns companies like Gizmodo, has been roundly ridiculed over the last few weeks for error-ridden articles that they published that were written by AI. However, that score, along with antipathy from Geo staff, is not enough to change course. Meryl Brown, Geo's editorial director, said, It is absolutely a thing we want to do more of. And CEO Jim Spanfeller says, Over in the U.S., the regulatory march around AI continues.

5:16And yet, as Senate Majority Leader Chuck Schumer focuses on comprehensive legislation, other senators are focused on smaller, more defined measures. Michigan Senator Gary Peters has introduced legislation called the AI LEAD Act, which is scheduled for a markup on Wednesday of this week, and is focused exclusively on the federal government itself, in terms of how it builds, buys, and deploys AI-driven systems. Daniel Ho, a member of the White House's National AI Advisory Committee, said, The government is going to be one of the largest purchasers of AI systems, So the standard that it sets will have a pronounced impact on responsible AI innovation.

5:48Meanwhile, just like we covered antipathy from Gary Gensler in the SEC towards AI, on yesterday's show, a different financial regulator, this time the Fed's banking regulator, Michael S. Barr, the Fed's vice chair for supervision, has made another warning about AI, saying that it could lead to illegal lending practices such as excluding minorities. Barr said, while these technologies have enormous potential, they also carry risks of violating fair lending laws and perpetuating the very disparities that they have the potential to address. The example that he gave was digital redlining, where minority communities are denied access to credit or housing opportunities.

6:19The fear is, of course, that AI trained on prejudiced or biased data could end up reinforcing and extending that prejudice or bias. So just another example of how basically every department in the government is trying to figure out how AI is going to impact what they have particular oversight into. That is going to do it for today's AI Breakdown Brief. If you're enjoying it, you should go subscribe to the AI Breakdown newsletter. It comes out every morning and features the five most interesting or important stories in AI. You can find a link down below in the show notes. Thanks again for listening or watching, and I'll be back soon with the main AI breakdown.

6:55Hey guys, before we dive into the main part of the episode, I want to share a little bit about today's sponsor, Supermanage. A truly great one-on-one should be about celebrating wins, solving problems, and deepening the connection between two human beings. But what if you miss those wins, never heard about those problems, and spent your whole meeting avoiding the hard stuff? That's where Supermanage comes in. Supermanage AI distills your public Slack channels into a one-on-one brief that highlights everything you need to know to jump right in. Because let's face it, you want your team to do the best work of their lives.

7:26And that starts with world-class conversations. Visit supermanage.ai slash breakdown today to start making the most of your one-on-ones. Thanks again to Supermanage for sponsoring the AI Breakdown. hmm meta has officially announced the launch of llama 2 it's an updated more powerful still open and now commercially available version of their large language model and represents not only a significant competitor to gpt and bard but is also flaring up significant conversations about the risks and opportunities of open source ai welcome back to the ai breakdown yesterday meta crushed the rest of the news of the week when they announced their much anticipated llama 2 there were a number of big parts of this announcement.

8:09The first is that Llama 2 remains an open source approach to LLMs. The second is that it's free not only for research but for commercial use. A third is that Meta is deepening a partnership with Microsoft through which developers using the Azure cloud will be able to natively access Llama. And finally, there is a huge emphasis on safety, which makes sense given the controversy around whether AI should be open sourced at all. Before we get into Llama 2, let's go back and actually look at Llama 1 because it's had a pretty important role in the development of this space over the last six months. Now, going back to Llama 1, in March it was released as an open-source package, although it wasn't complete.

8:44Basically, the weights in the model weren't included. However, within about a week of announcing it, the full model was leaked online, and almost immediately people were concerned about the implications. Jeffrey Ledish, who we'll hear from again later in the show, said, Get ready for loads of personalized spam and phishing attempts. Open-sourcing these models was a terrible idea. Now, while some of the more dire warnings about scams and attacks might not have been borne out quite yet when it comes to that leak, that's not to say that there weren't serious implications for how Lama's open model being available would impact the development of the AI space.

9:16In May, another leak, this time from someone inside Google, argued that the real competitor for Google and OpenAI was not another big company developing a model based on huge amounts of training data, but instead was the insurgency coming from the open source ranks. The piece starts, we've done a lot of looking over our shoulders at OpenAI. Who will cross the next milestone? What will the next move be? But the uncomfortable truth is we aren't positioned to win this arms race, and neither is OpenAI. While we've been squabbling, a third faction has been quietly eating our lunch. I'm talking, of course, about open source.

9:46Plainly put, they are lapping us. Things we consider major open problems are solved and in people's hands today. Now, the important part of this analysis for our story today comes in the What Happened section. The anonymous author writes, At the beginning of March, the open source community got their hands on the first really capable foundation model as Meta's Llama was leaked to the public. It had no instruction or conversation tuning and no RLHF. Nonetheless, the community immediately understood the significance of what they had been given. A tremendous outpouring of innovation followed with just days between major developments.

10:17Here we are barely a month later and there are variants with instruction tuning, quantization, quality improvements, human evals, multimodality, RLHF, etc, etc, many of which build on each other. Now the author also does talk about what feels to them like the irony of Facebook being the leader in this new environment. They write, paradoxically, the one clear winner in all of this is Meta. Because the leaked model was theirs, they have effectively garnered an entire planet's worth of free labor. Since most open source innovation is happening on top of their architecture, there is nothing stopping them from directly incorporating it into their products.

10:48The value of owning the ecosystem cannot be overstated. Now that letter came out in May and importantly, we're now a couple months later seeing the impacts in a big way. If you're a regular listener, you will have heard me read Sam Hogan's big essay tweet the other day, where he argued basically that AI was not the savior to the venture startup ecosystem that people had thought, because the two big winners were on the one end of the spectrum, open source indie developers, and on the other end of the spectrum, big enterprise companies. Now his argument for why the enterprises were doing better than anyone thought, had a lot to do with these open source models.

11:21As a reminder, he wrote, executives at enterprise companies are excited about AI, and they have been vocal about this from the beginning. This led a lot of founders and VCs to believe these companies would make good first customers. What the startups building for these companies failed to realize is just how aligned and savvy executives and the engineers they manage would be at quickly getting AI into production using open source tools. An engineering leader would rather spin up their own Langchain and Chroma infrastructure for free and build tech themselves than buy something from a new, unproven startup.

11:47So this was the situation heading into the last week, and lots and lots of rumors had been swirling that Llama 2 was on the way, and that this time it would come with a license ready for commercial use. Well, as of yesterday, Llama 2 is here, it is indeed ready for commercial use, it picked up an interesting partner in Microsoft, and it's generating some serious discussion around issues of open source AI. Let's talk first about how Llama compares to other open source models. TLDR is its way out ahead. For those of you who are listening, on the screen I'm showing a chart that shows benchmark comparisons of LAMA 2, both its 7 billion parameter version and its 13 billion parameter version, outcompeting many other open source LLMs.

12:27Now there's also a chart that they shared in the white paper that shows how LAMA 2 compares in various benchmarks to other commercially available models like GPT-3.5, GPT-4, POM, and POM-2L. While LAMA remains pretty meaningfully behind GPT-4 as a 4 example, it's coming up pretty close to the levels of GPT-3.5. What's more, as NVIDIA's Dr. Jim Phan points out, model tests that involved humans suggested that LLAMA performed even better. Jim writes, Meta's team did a human study on 4K prompts to evaluate LLAMA-2's helpfulness. They use win rate as a metric to compare models in similar spirit as the Vakuna benchmark.

13:01The 70 billion parameter model roughly ties with GPT-3.5 and performs noticeably stronger

13:13Now in that same tweet, Jim also points out that Lama 2 is not yet at the GPT 3.5 level, and that the big thing holding its back is its coding abilities. Speaking of quirky human tests, Professor Ethan Mollick from Wharton writes, Out of the box, Lama 2 beats Bard at the insane memo test. Write a corporate memo in a serious style explaining and justifying the following points. One, the floor is now lava. Two, promotion will be by staring contests. Three, we have merged with a hive of bees. The queen is your new CTO. Now, as we mentioned, in terms of upgrades from Llama 1, the biggest one is the commercial availability.

13:46If you go back and look at how the developer community was discussing and talking about the first iteration of Llama, a lot of it was about trying to assess whether Meta would actually sue if people used it for commercial products. For example, this Hacker News post says, Can Llama weights be used for commercial products? And the top-rated comments is all about the difference between what the terms literally say, which did exclude commercial use, versus what they would actually do because the optics of suing might be terrible. Well, that has now been resolved as this model, again, is available for commercial use.

14:15And importantly, again, from a commercial standpoint, Meta isn't charging directly for its usage. They'll make money by selling the program as a paid hosted service through various cloud computing partners. That's, for example, where Microsoft comes in. Now, there are a couple commercial limitations to note. The terms prevent Llama 2's data or output from being used to train other LLMs. And second, if the monthly active users of the product that is using Llama 2 exceed 700 million users, Llama is requiring a special commercial license. Obviously, there's a very small handful of companies for whom that would apply.

14:46Now, going back to Microsoft for a moment, people were fairly surprised by this announcement featuring Microsoft so prominently. Matt Wolf writes, So Microsoft has partnered with OpenAI on their closed-source LLM, and now they're partnering with Meta to release an open-source LLM with Llama 2. I love that things are moving towards more open source. I'm just really confused by where Microsoft is going with all this. For market observers, though, the answer is pretty clear. Barron's writes yesterday, Microsoft shows investors the money from AI. Why its meta deal threatens Google. The piece starts, Microsoft has just closed the gap between the hype and the reality when it comes to AI.

15:18The tech giant unveiled its plan to monetize the technology Tuesday, answering a key question surrounding the recent AI stock boom. The company plans to charge businesses$30 a month for its artificial intelligence-powered Microsoft Office apps. In response to these updates, yesterday Microsoft's stock hit an all-time high. Now another big emphasis of the announcement of Llama 2 was around its approach to safety. Lewis Martin tweets, I am proud to have led the safety effort behind Llama 2. Our fine-tuned models are deemed safer and more helpful compared to other open and closed-source models such as ChatGPT.

15:48Safety was evaluated by human annotators on a set of 2K adversarial prompts. We improved the safety of our models using supervised fine-tuning, RLHF, context distillation, and continuous red-teaming. In particular, we noticed that RLHF makes our model more robust on the long tail of adversarial prompts. Thanks to context distillation, we have improved our model's responses to adversarial prompts. We first generate answers by prefixing a prompt with safety guidelines, then fine-tune the model on these safe responses without these guidelines. We proactively test our model's weaknesses with continuous red-teaming.

16:19We conducted a series of red teaming events with various teams of over 350 people, including domain experts. They also included individuals representative of a variety of demographic groups. One thing that many have noticed is that Meta took a slightly different approach to dealing with these safety issues by actually training Llama with two separate reward models. One was based on its helpfulness and one was based on its safety. This allowed them to have more fine control over how the model should respond in different contexts and scenarios. Now that said, some saw the very release of information, particularly the weights of the model, as undermining all of this focused on safety.

16:51Stanford PhD student Chris Cundy writes, I appreciate all the emphasis on safety in the LLAMA 2 paper, but I'm not sure how that squares with releasing the weights. If I want CrimeLLAMA for effective phishing emails, can't I just fine tune to remove safety guardrails? Jeffrey Ledish said something similar. If you have access to the weights, you can fine tune away any safety controls. And this gets us to the discussion of open source more broadly. On the one hand, it's hard to deny how much Lama 2 advances the open source LLM ecosystem. Nathan Lambert wrote on his substack, quote, The base model seems very strong beyond GPT-3, and the fine-tuned chat model seem to be on the same level as chat GPT.

17:26It is a huge leap forward for open source and a huge blow to closed source providers as using this model will offer way more customizability and way lower costs for most companies. Remember what we had discussed before, how enterprises had changed the way that they engaged with AI because of the availability of this type of open source model. And indeed, a lot of the mainstream media coverage focused on the risks of open sourcing. The Washington Post writes, Facebook to make its AI free to use, expanding access to powerful tech. The social media giant is doubling down on its open source approach, potentially boosting competition, while also raising the risks of malicious actors using the tech.

18:01From the Post, quote, the decision will deepen the divide forming in the tech world over whether to make new AI tech open source or not. Google and OpenAI have rejected full transparency, citing the risks of bad actors using their tech or developing it in ways that increase risks to people. Facebook and a group of startups, including Hugging Face and Stability AI, have said open source is key to making sure the powerful new technology doesn't further entrench the tech giants and stifle competition. The Post also writes, earlier this year, Meta released Lama to a select group of researchers only for the model to be leaked and later used for applications ranging from drug discovery to sexually explicit chatbots.

18:33Last month, Senators Richard Blumenthal and Josh Hawley, wrote to Zuckerberg arguing that in the short time generative artificial intelligence applications have become more widely available, they have already been misused for problematic content, from pornographic deepfakes to malware and phishing campaigns. Now, Meta itself has pushed back on this idea. Nick Clegg, who is the president of global affairs at Meta and is a former UK deputy prime minister, said on BBC4 yesterday, my view is that the hype has somewhat run ahead of the technology. I think a lot of the existential warnings relate to models that don't currently exist.

19:03So-called super-intelligent, super-powerful AI models. The vision where AI develops an autonomy and agency on its own where it can think for itself and reproduce itself. The models that we're open sourcing are far, far, far short of that. In fact, in many ways, they're quite stupid. Now, it's pretty clear that the release of Llama 2 sets up this open source debate to move even more into the mainstream. A quick search on Twitter for Llama and Open will see just how much disagreement there is even within the often mono-thinking Silicon Valley tech culture. However, for now, for most developers, those debates can wait because they have an incredibly powerful new tool and they're outbuilding the next wave of AI innovation.

19:38That's going to do it for today's AI Breakdown. If you're enjoying this and you're watching, please go listen to the podcast. If you're listening to this, go check out the YouTube. You can get information about everywhere this content lives at breakdown.network. And until next time, peace.

19:58Thank you.