🤖 AI Brief: Meta’s Llama 2, Apple GPT, an AI South Park episode, and more.

Today is July 24, 2023.

This week’s issue is a tad longer than previous issues: in the past week, we’ve had a number of important AI news worthy of a deep dive, including the launch of Meta’s Llama 2 language model.

The top questions that I’m thinking about right now are:

• Will open-source beat closed-source AI models? Llama 2’s launch could turn up the heat here as companies adopt its open model.

• Can Google find its groove in AI? It still feels like they’re in the early innings of playing catchup.

• How painful will job disruption from AI be? AI seems like it’s going after creatives and journalists at a rapid pace. I cover several developments here.

• How will Apple’s entrance into generative AI go? The company likes to wait to get things perfect.

In this issue:

• 🧠 Meta releases Llama 2, its first commercially available model

• 🏃‍♂️ Google cofounder Sergey Brin is leading the creation of a GPT-4 competitor

• 📺️ Fable's AI tech generates an entire AI-made South Park episode

• 🍎 Apple has developed "Apple GPT" as it prepares for a major AI push in 2024

• 💾 LLMs are a "threat" to human data creation, researchers warn

• 🔎 OpenAI, Google, and other tech firms agree to voluntary safeguards, and other quick scoops

• 🧪 The latest science experiments, including the latest in face manipulation tech

Did someone forward this to you? Subscribe here!

🧠 Meta releases Llama 2, offering an open and commercially accessible LLM for the first time

Rumored for weeks, Meta finally released their latest LLM, LLaMA 2. It’s available for download and you can play around with it here on Hugging Face as well.

Here's what's important to know:

• For the first time, Llama 2 can be used for commercial purposes: apps with more than 700 million monthly users will have to apply for a license, but otherwise there are no other restrictions.

• It’s not truly “open-source”, but it’s close: while Llama 2’s license is relatively open, it falls short of standard open source licenses and contains certain limitations, some critics argue.

• The model was trained on 40% more data than Llama 1, with double the context length: this should offer a much stronger starting foundation for people looking to fine-tune it.

• Early results show Llama 2 outperforms other open-source models across a variety of benchmarks: MMLU, TriviaQA, HumanEval and more were some of the popular benchmarks where Llama 2 beat Llama 1, Falcon and MosaicML's MPT model.

Interestingly, Meta also announced a cozy partnership with Microsoft:

• Microsoft is our preferred partner for Llama 2, Meta announces in their press release, and "starting today, Llama 2 will be available in the Azure AI model catalog, enabling developers using Microsoft Azure."

• MSFT knows the open-source AI approach is going to be big, and they’re not putting all their eggs in one basket despite a massive $10B investment in OpenAI.

The takeaway: the open-source vs. closed-source wars just got really interesting. Google and OpenAI, both of which have closed-source models, are going to face increased pressure from an open-source alternative now.

🏃‍♂️ Google cofounder Sergey Brin leading creation of a GPT-4 competitor

Google's cofounder Sergey Brink, who notably stepped back from day-to-day work in 2019, is actually back in the office again, the Wall Street Journal revealed (note: paywalled article).

The reason? He's leading a push to develop "Gemini," Google's answer to OpenAI's GPT-4 large language model.

Why this matters:

• Concern about falling behind is clearly top of mind: Google was considered a tech and AI pioneer for much of their history, and sources speculate Brin is worried their recent missteps could leave them vulnerable.

• Brin views Generative AI as a pivotal moment of transformation in tech: it's enough to pull him away from other interests and back into day-to-day work.

• Speed is key as Google plays from behind: internal strategy at Google in recent months has focused on moving quickly (perhaps too quickly, some critics argue) and adding AI features to a broad range of products. Brin's involvement is helping catalyze this velocity, sources explained.

While Brin didn't comment on the article, the WSJ revealed he's been quite active in the AI community, and has been attending various AI events. This marks a shift from Brin's earlier beliefs:

• Early on, Brin "expressed skepticism that they could crack artificial intelligence," the WSJ reports, noting that he "ignored the work of the Brain Team" that he originally helped start.But recently his mindset has shifted as AI progress has picked up.

The main takeaway:

• Founders coming back to their companies can often inject a new sense of urgency and mission. The most famous example is probably how Steve Jobs reinvigorated Apple.

• While Google won't say it publicly, it's likely they're treating this moment as an existential crisis. All the internal signals (the "code red" memos, Brin's involvement, the dropping of AI safeguards) are signs that they are reorienting to move quickly here.

How this will play out, though, is ultimately unknown. Google also has the threat of open-source AI to contend with (Google's PaLM 2 remains closed-source).

📺️ Fable's AI tech generates an entire AI-made South Park episode

Fable, a San Francisco startup, just released its SHOW-1 AI tech that is able to write, produce, direct, animate, and even voice entirely new episodes of TV shows.

Fable's first proof of concept? A 20-minute episode of South Park entirely written, produced, and voiced by AI. Watch the episode and see their Github project page here for a tech deep dive.

Their tech critically combines several AI models: LLMs for writing, custom diffusion models for image creation, and multi-agent simulation for story progression and characterization.

Why this matters:

• Current generative AI systems like Stable Diffusion and ChatGPT can do short-term tasks, but they fall short of long-form creation and producing high-quality content, especially within an existing IP.

• Hollywood is currently undergoing a writers and actors strike at the same time; part of the fear is that AI will rapidly replace jobs across the TV and movie spectrum.

• The holy grail for studios is to produce AI works that rise up the quality level of existing IP; SHOW-1's tech is a proof of concept that represents an important milestone in getting there.

• Custom content where the viewer gets to determine the parameters represents a potential next-level evolution in entertainment.

In a nutshell: SHOW-1's tech is actually an achievement of combining multiple off-the-shelf frameworks into a single, unified system. This is what's exciting and dangerous about AI right now – the right tools, with just enough tweaking and tuning, can start to produce some very fascinating results.

The main takeaway:

• Actors and writers are right to be worried that AI will be a massively disruptive force in the entertainment industry. We're still in the "science projects" phase of AI in entertainment -- but also remember we're less than one year into the release of ChatGPT and Stable Diffusion.

• A future where entertainment is customized, personalized, and near limitless thanks to generative AI could arrive in the next decade. But as exciting as that sounds, ask yourself: is that a good thing?

🍎 Apple has developed "Apple GPT" as it prepares for a major AI push in 2024

Apple has been relatively quiet on the generative AI front in recent months, which makes them a relative anomaly as Meta, Microsoft, and more all duke it out for the future of AI.

The relative silence doesn't mean Apple hasn't been doing anything, and a Bloomberg report (note: paywalled article) sheds light on their master plan: they're quietly but ambitiously laying the groundwork for some major moves in AI in 2024.

Driving the news:

• Apple is internally testing a chatbot dubbed "Apple GPT" right now. After being caught "flat-footed" by ChatGPT, they're playing catch up.

• The company has also built a framework for creating LLMs, dubbed "Ajax". Ajax is designed to accelerate Apple's ability to move quickly on the generative AI front heading into next year. Their overall plans are not public, but the leak about Ajax is a confirmation their ambition is wide in scope.

Why this matters: trillions of dollars in market cap are at stake.

• While Apple has moved ahead with imbuing their products with AI (maps, search, photos etc.), they're worried about losing the race in generative AI.

• Their cautious approach towards AI and privacy means products like Siri have stagnated, giving up their early mover advantage in the assistant space.

• Apple regards generative AI as a "paramount shift in how devices operate," and see this as an existential threat to the company's ability to sell devices.

• Tim Cook acknowledged he's using ChatGPT in a recently interview. It's something Apple is "looking at closely," he confirmed. But generative AI has a "number of issues that need to be sorted," he noted.

The main takeaway:

• Apple's recent previews of their Vision Pro show that they really want to get something right, in a way that can exceed existing consumer expectations.

• If their release of generative AI tech to consumers doesn't turn out like Apple Maps did (a complete disaster of a launch), things could get very interesting in the LLM space.

• But Apple is under the gun here. The AI space is moving fast, and they don't have years of time to get things perfect.

💾 LLMs are a "threat" to human data creation, researchers warn, as traffic from sites like StackOverflow declines.

LLMs rely on a wide body of human knowledge as training data to produce their outputs. Reddit, StackOverflow, Twitter and more are all known sources widely used in training foundation models.

A team of researchers is documenting an interesting trend: as LLMs like ChatGPT gain in popularity, they are leading to a substantial decrease in content on sites like StackOverflow. Here's the paper on arXiv for those who are interested in reading it in-depth. 

Why this matters:

• High-quality content is suffering displacement, the researchers found. ChatGPT isn't just displaying low-quality answers on StackOverflow.

• The consequence is a world of limited "open data", which can impact how both AI models and people can learn.

• "Widespread adoption of ChatGPT may make it difficult" to train future iterations, especially since data generated by LLMs generally cannot train new LLMs effectively.

This is the "blurry JPEG" problem, the researchers note: as JPEGs are resaved and shared, their quality degrades over time. The same could happen to the internet’s content as AI content increasingly takes over and AI models use AI content for training.

The main takeaway:

• We're in the middle of a highly disruptive time for online content, as sites like Reddit, Twitter, and StackOverflow also realize how valuable their human-generated content is, and increasingly want to put it under lock and key.

• As content on the web increasingly becomes AI generated, the "blurry JPEG" problem will only become more pronounced, especially since AI models cannot reliably differentiate content created by humans from AI-generated works.

🔎 Quick Scoops

OpenAI, Google, and other tech firms agree to voluntary safeguards to mitigate risk of AI. Details remain murky, but the White House announced this set of voluntary commitments as an initial step in the AI regulatory journey. (Forbes)

Is ChatGPT’s performance degrading over time? This Stanford study documents some fairly significant performance shifts between different versions of GPT-3.5 and GPT-4. (arXiv)

G/O Media intends to flood the internet with more AI-generated content. The owner behind Jezebel, AV Club and more talks openly about using more AI tools. (Vox)

Google is pitching an AI for writing news articles. Media orgs who saw it found it "unsettling" as the journalism field faces a rapidly changing landscape. (New York Times - note: paywalled)

Microsoft and OpenAI test synthetic data to train LLMs. Web data is "no longer good enough" to produce more refined models, they’re saying. (Financial Times - note: paywalled)

World of Warcraft fans tricked an AI into writing about a non-existent feature. As AI tools mine Reddit for news content, lack of human editing leaves them vulnerable to misinformation. (BBC)

🧪 Science Experiments

FaceCLIPNeRF: Text-driven 3D Face Manipulation now possible

• Recent advances in Neural Radiance Fields (NeRF) have enabled high-fidelity 3d face reconstruction. These researchers layer in a single text prompt to manipulate a face reconstructed with NeRF, offering a new level of access and ease for face manipulation.

• arXiv paper here.

STEVE-1: A Generative Model for Text-to-Behavior in Minecraft

• Constructing AI models that respond to text instructions is challenging, especially for sequential decision-making tasks. This pretrained model for Minecraft offers a novel pathway to creating advanced behavior from a simple text prompt.

• arXiv paper here.

Stability AI introduces FreeWilly 1 and FreeWilling 2 LLMs

• These are finetuned on Llama 1 and Llama 2, respectively, against Stabilty AI’s Orca dataset

• Download and try it on Hugging Face.Chart showing how FrugalGPT scales performance against cost.

😀 A reader’s commentary

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at [email protected] 

As always — have a great week!