That was fast. In lower than per week since Meta launched its AI model, LLaMA 2, startups and researchers have already used it to develop a chatbot and an AI assistant. It is going to be only a matter of time until corporations start launching products built with the model.
In my story, I have a look at the threat LLaMA 2 could pose to OpenAI, Google, and others. Having a nimble, transparent, and customizable model that’s free to make use of could help corporations create AI services and products faster than they may with a giant, sophisticated proprietary model like OpenAI’s GPT-4. Read it here.
But what really stands out to me is the extent to which Meta is throwing its doors open. It is going to allow the broader AI community to download the model and tweak it. This might help make it safer and more efficient. And crucially, it could exhibit the advantages of transparency over secrecy relating to the inner workings of AI models. This might not be more timely, or more essential.
Tech corporations are rushing to release their AI models into the wild, and we’re seeing generative AI embedded in an increasing number of products. But essentially the most powerful models on the market, corresponding to OpenAI’s GPT-4, are tightly guarded by their creators. Developers and researchers pay to get limited access to such models through a web site and don’t know the main points of their inner workings.
This opacity may lead to problems down the road, as is highlighted in a brand new, non-peer-reviewed paper that caused some buzz last week. Researchers at Stanford University and UC Berkeley found that GPT-3.5 and GPT-4 performed worse at solving math problems, answering sensitive questions, generating code, and doing visual reasoning than that they had a few months earlier.
These models’ lack of transparency makes it hard to say exactly why that is perhaps, but regardless, the outcomes must be taken with a pinch of salt, Princeton computer science professor Arvind Narayanan writes in his assessment. They’re more likely attributable to “quirks of the authors’ evaluation” than evidence that OpenAI made the models worse. He thinks the researchers didn’t have in mind that OpenAI has fine-tuned the models to perform higher, and that has unintentionally caused some prompting techniques to stop working as they did previously.
This has some serious implications. Corporations which have built and optimized their products to work with a certain iteration of OpenAI’s models could “100%” see them suddenly glitch and break, says Sasha Luccioni, an AI researcher at startup Hugging Face. When OpenAI fine-tunes its models this manner, products which were built using very specific prompts, for instance, might stop working in the best way they did before. Closed models lack accountability, she adds. “If you’ve a product and you alter something within the product, you’re speculated to tell your customers.”
An open model like LLaMA 2 will not less than make it clear how the corporate has designed the model and what training techniques it has used. Unlike OpenAI, Meta has shared the complete recipe for LLaMA 2, including details on the way it was trained, which hardware was used, how the information was annotated, and which techniques were used to mitigate harm. People doing research and constructing products on top of the model know exactly what they’re working on, says Luccioni.
“Once you’ve access to the model, you’ll be able to do all styles of experiments to be sure that that you just recuperate performance otherwise you get less bias, or whatever it’s you’re searching for,” she says.
Ultimately, the open vs. closed debate around AI boils all the way down to who calls the shots. With open models, users have more power and control. With closed models, you’re on the mercy of their creator.
Having a giant company like Meta release such an open, transparent AI model looks like a possible turning point within the generative AI gold rush.
If products built on much-hyped proprietary models suddenly break in embarrassing ways, and developers are kept at nighttime as to why this is perhaps, an open and transparent AI model with similar performance will suddenly seem to be a way more appealing—and reliable—alternative.
Meta isn’t doing this for charity. It has so much to realize from letting others probe its models for flaws. Ahmad Al-Dahle, a vp at Meta who’s leading its generative AI work, told me the corporate will take what it learns from the broader external community and use it to maintain making its models higher.
Still, it’s a step in the best direction, says Luccioni. She hopes Meta’s move puts pressure on other tech corporations with AI models to think about a more open path.
“I’m very impressed with Meta for staying so open,” she says.
Deeper Learning
Face recognition within the US is about to satisfy considered one of its biggest tests
By the top of 2020, the movement to limit police use of face recognition within the US was riding high. Around 18 cities had enacted laws forbidding the police from adopting it, and US lawmakers proposed a pause on the federal government’s use of the tech. Within the years since, that effort has slowed to a halt. Five municipal bans on police and government use passed in 2021, but none in 2022 or in 2023 thus far. Some local bans have even been partially repealed.
All eyes on Massachusetts: The state’s lawmakers are currently thrashing out a bipartisan bill that may allow only state police to access a really limited face recognition database, and require them to have a warrant. The bill represents a significant test of the prevailing mood around police use of those controversial tools. Read more from Tate Ryan-Mosley here.
Meanwhile, in Europe: Police use of facial recognition technology can be a significant sticking point for European lawmakers negotiating the AI Act. EU countries want their police forces to make use of the technology more. Nonetheless, members of the EU Parliament desire a more sweeping ban on the tech. The fight will likely be an extended, drawn-out one, and it has change into existential to the AI Act.
Bits and Bytes
The White House 🤝AI corporations
The Biden administration announced last week it had made a pact with Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI that they’d develop recent technologies in a protected, secure, and transparent way. Tech corporations pledged to watermark AI-generated content, spend money on cybersecurity, and test products before releasing them to the market, amongst other things. But that is all completely voluntary, so the businesses will face no repercussions in the event that they don’t do it. The voluntary nature of this announcement shows just how limited Biden’s powers are relating to AI.
ChatGPT’s surprising skill: Facial recognition
OpenAI is testing a version of ChatGPT that may recognize and describe people’s faces from pictures. The tool could aid visually impaired people, but might be a privacy nightmare. (The Latest York Times)
Apple has built its own generative AI model and chatbot
Higher late than never, I assume. Apple executives have still not decided how they’ll release their model, Ajax, and chatbot, Apple GPT, to consumers. (Bloomberg)
Meet the Google engineers who pioneered an AI revolution
A pleasant have a look at the origin story of the transformer, the AI technology powering today’s generative AI boom, and the team of engineers who built it. Notably, none of them work at Google anymore. (Financial Times)