TechScape: Will Metas’ Open Source LLM Make AI Safer or Put It in the Wrong Hands? | Apple

[ad_1]

The summer of AI is truly upon us. (This joke may not play quite as well for Southern Hemisphere readers.) Whether we call this period the peak of the hype cycle or simply the moment when the curve goes vertical will only be obvious in hindsight, but the cadence of big news in the field has gone from weekly to almost daily. Let’s see what the biggest AI players Meta, Microsoft, Apple and OpenAI are doing.

Apple

Always one to keep his cards close to his chest, don’t expect to hear about many R&D breakthroughs from Cupertino. The AI ​​work that has led to product shipments is also being hidden rather than shouted from the rooftops, with the company speaking about machine learning and transformers at its annual worldwide developer conference (WWDC) last month, but clearly avoiding AI.

But that doesn’t mean he isn’t playing the same game as everyone else. For Bloomberg():

The iPhone maker has created its own framework for building large language models, the AI-powered systems at the heart of new offerings like ChatGPT and Google’s Bard according to people familiar with the efforts. With that foundation, known as Ajax, Apple also built a chatbot service that some engineers call Apple GPT.

In recent months, the AI ​​push has become a major effort for Apple, with several teams collaborating on the project, people said, who asked not to be identified because the matter is private. The job includes trying to address potential technology-related privacy issues.

On the one hand: Of course they are. It’s hard to remember, because it’s been on the sidelines for so long, but Apple led the industry with voice assistants when it launched Siri in 2011. But surely within a few years of the Echo smart speaker’s launch in 2014 it had fallen behind, and now it’s been relegated to near-joke status. Fixing Siri is hard work, but that’s what the Vanguard of Work LLM is perfect for. So it’s no surprise that the company is working on it.

On the other hand: Building a model foundation is difficult, expensive and perhaps unnecessary. Apple has already built on open source roots (each of its operating systems, for example, ultimately sits on top of the open source Darwin kernel) and has licensed technology from third parties (notably, these days, Arm, which still provides the underlying designs for its chips). And there are plenty of opportunities for both of these approaches

Meta and Microsoft

The Metas Llama foundation model has become the accidental foundation, ahem, of an entire research community. Competitor GPT was released for download to a select group of researchers, who had signed NDAs and promised not to share it more widely when it was promptly leaked. Copies of Samizdat were shared across the network, as was a whole system for collaborating without ever publishing the stolen LLM openly. The whole thing was against Meta’s terms, but the company didn’t seem too unhappy with being at the center of a computer revolution.

And now, it’s official. Metas released Llama 2 with terms of service that legitimize that ecosystem. From his announcement:

We are now ready to open source the next version of Llama 2 and are making it freely available for research and commercial use. Model weights and initial code have been included for the pre-trained model and also for the refined conversational versions.

And the company has partnered with Microsoft to expand access:

Starting today, Llama 2 is available in the Azure AI template catalog, enabling developers using Microsoft Azure to build with it and leverage their own cloud-native tools for content filtering and security features. It’s also optimized to run locally on Windows, giving developers a seamless workflow while delivering generative AI experiences to customers across platforms

However, the model is free as in beer, rather than free as in speech. Metas’ commercial terms essentially require a license from any company with more than 700 million monthly active users, every other company discussed in today’s newsletter, and very few others. It also prevents anyone from using Llama 2 to enhance other LLMs. It may be free, in other words, but it’s not open source.

Open AI

Do measures to make LLMs safer also make them dumber? OpenAI Photography: Olivier Morin/AFP/Getty Images

But it is still more open than the competition. Allowing users, researchers, and (smaller) competitors to download the full model and poke around to see how it works obviously helps anyone who wants to build on what you’ve done, but it also helps build trust with potential partners. For a sign of the pitfalls that come with the opposite approach, check out OpenAI. From Ars Technica:

In a study titled How Does ChatGPT Behavior Change Over Time? posted on arXiv, Lingjiao Chen, Matei Zaharia, and James Zou, questioned the consistent performance of OpenAI’s large language models (LLMs), especially GPT-3.5 and GPT-4. Using API access, they tested the March and June 2023 versions of these models on tasks like solving math problems, answering sensitive questions, generating code, and visual reasoning. Notably, GPT-4’s ability to identify prime numbers plummeted dramatically from 97.6% accuracy in March to just 2.4% in June. Oddly enough, GPT-3.5 has shown improved performance over the same period.

The findings contribute to the widely held fear that efforts to improve the security of the GPT are making it dumber. OpenAI regularly releases changes to GPT, and given how regularly CEO Sam Altman talks about the security of AI, it’s perfectly plausible that those changes are largely security-focused. And so if the system is getting worse, not better, maybe it’s because of that compromise.

skip past newsletter promotion

Alex Hern’s weekly dive into how technology is shaping our lives

“,”newsletterId”:”tech-scape”,”successDescription”:”We’ll send you TechScape every week”}” clientOnly>Privacy Policy: Newsletters may contain information about charities, online ads, and externally funded content. For more information, see our Privacy Policy. We use Google reCaptcha to protect our website and the Google Privacy Policy and Terms of Service apply.

after the promotion of the newsletter

But the paper itself does not hold. Ars Technica, again:

Artificial intelligence researcher Simon Willison also disputes the papers’ conclusions. I don’t find it very convincing, he told Ars. A decent part of their criticisms are whether or not the code output is wrapped in Markdown backticks or not. So far, Willison thinks any perceived changes in GPT-4’s capabilities stem from the novelty of the LLMs fading away. After all, GPT-4 unleashed an AGI panic wave shortly after launch and was once tested to see if it could take over the world. Now that the technology has become more mundane, its flaws seem apparent.

But the allegations strike at the heart of OpenAI’s (ironically) closed model. The company rolls out changes to GPT on a regular basis, with little explanation and no way for users to understand why or how each new model differs. Inspecting any LLM is a black box problem, with little ability to peek inside and see how it thinks, but those problems are much worse when your only way of interacting is through an API to a version hosted by a third party.

Ars Tehcnica, one last time:

Willison agrees. Honestly, the lack of release notes and transparency might be the bigger story here, he told Ars. How can we build reliable software on a platform that changes in completely undocumented and mysterious ways every few months?

X marks tweet spotsTwitter blue bird no more. Photograph: Mateusz Sodkowski/ZUMA Press Wire/Shutterstock

So Twitter has a new name — here’s everything we know so far.

Each TikTok is worth 1,000 words: In response to Twitter and Threads, the video-sharing platform is now offering the ability to create long, text-only posts.

If you want to read the full version of the newsletter, sign up to receive TechScape in your inbox every Tuesday.

[ad_2]