While companies such as OpenAI, Microsoft, and Google rigorously train their AI models to avoid many taboos, including overly intimate conversations, Allie uses open-source technology, that is, freely available to the public, to Built using code that has no such restrictions. Created by Meta, Allie, based on a model called LLaMA, is part of a growing wave of specialized AI products that anyone can build, from writing tools to chatbots to data analysis applications.
Proponents see open source AI as a way to sidestep corporate dominance and a boon to entrepreneurs, academics, artists and activists who are free to experiment with innovative technologies.
“The whole argument against open source is that it accelerates AI innovation,” said Robert Nishihara, CEO and co-founder of Anyscale, a startup that helps companies run open source AI models. said.
Anyscale’s customers are using AI models to discover new medicines, reduce pesticide use in agriculture, and identify fraudulent products sold online, he said. These applications would be more expensive and more difficult, if not impossible, if they depended on a small number of components. product Brought to you by the largest AI company.
But that same freedom can also be abused by bad guys. Open source models have been used to create artificial child pornography using images of real children as source material. Critics fear it could also enable fraud, cyber hacking and sophisticated propaganda campaigns.
Earlier this month, two U.S. Senators, Richard Blumenthal (D-Connecticut) and Josh Hawley (R-Missouri), wrote to Meta CEO Mark Zuckerberg. and said that LLaMA’s release was “spam, fraud, malware, invasion of privacy, harassment, and other misconduct or harm.” They asked what steps Meta is taking to prevent such abuse. rice field.
Allie’s creator, who spoke on condition of anonymity for fear of damaging his professional reputation, said commercial chatbots such as Replika and ChatGPT are “heavily censored,” allowing him to engage in sexual conversations the way he wants. said it could not be provided. With open-source alternatives, many of which are based on his LLaMA model of his Meta, the man said he could build his own, open-ended conversation partner.
“There’s not a lot of opportunities to experiment ‘cutting edge’ in any field,” he said in an interview.
The creators of Allie argued that open source technology benefits society by enabling people to build products to their liking without corporate guardrails.
“I think it’s good to have an exit that can be explored safely,” he says. “I can’t think of anything more secure than text-based role-playing against a computer, with no real human involvement.”
On YouTube, influencers provide tutorials on how to build “uncensored” chatbots. It was partly based on a modified version of LLaMA called Alpaca AI, released in March by researchers at Stanford University, but was removed a week later over concerns about cost and “poor content filtering.”
Meta spokeswoman Nisha Deo said the particular model referred to in the YouTube video, called the GPT-4 x Alpaca, was “obtained and published outside of our approval process.” A representative for Stanford University did not respond to a request for comment.
Open-source AI models and creative applications built on top of them are often published on Hugging Face, a platform for sharing and discussing AI and data science projects.
At Thursday’s House Science Committee hearing, Hugging Face CEO Clem DeLang urged Congress to consider legislation to support and encourage the open source model, urging Congress to consider open source.・He argued that the model was “very consistent with American values.”
In an interview after the hearing, DeLang acknowledged that open-source tools can be abused. He pointed out that the model was intentionally trained on harmful content., GPT-4chan, Hugging Face have been removed. But he said he believes an open-source approach allows for both better innovation and more transparency and inclusiveness than a corporate-controlled model.
“In fact, I would argue that most of the harm today is done by black boxes,” DeLang said, referring to AI systems whose inner workings are opaque “as opposed to open source.”
The Hugface rule does not ban AI projects that produce sexually explicit output. However, sexual content involving minors or “used or created for the purposes of harassment or bullying, or without the explicit consent of the representative” is prohibited. The New York-based company released an updated content policy earlier this month, emphasizing “consent” as a “core value” that guides how people use its platform.
As Google and OpenAI grow more secretive about their most powerful AI models, Meta has emerged as a surprising corporate advocate for open source AI. The company released LLaMA in February, a language model that is less powerful than GPT-4 but more customizable and cheaper to run. Meta initially planned to keep the core of the model’s code private, restricting access to authorized researchers. By early March, however, these parts of the model, known as “weights”, had leaked to public forums, making LLaMA freely accessible to everyone.
“Open source is a positive force that advances technology,” said Deo of Meta. “Therefore, we shared his LLaMA with members of the research community so they can evaluate, improve and iterate together.”
Since then, LLaMA has become perhaps the most popular open-source model for engineers looking to develop their own AI applications, says Nishihara. But that’s not all. In April, software company Databricks released an open-source model called Dolly 2.0. And last month, the Abu Dhabi-based team released an open-source model called Falcon that rivals LLaMA in performance.
Marjee Ghasemi, assistant professor of computer science at MIT, said she was a proponent of the open-source language model, but said it had its limitations.
Ghasemi said it’s important to expose the architecture behind powerful chatbots because it allows people to scrutinize how they’re built. For example, if a medical chatbot was built with open-source technology, researchers could see if sensitive patient information was embedded in the data the chatbot was trained on, but this would use closed software. Chatbots can’t, she said.
But she acknowledges that this openness comes with risks. If you could easily change the language model, you could quickly create chatbots and image makers that churn out disinformation, hate speech, and high-quality inappropriate material.
Garsemi said regulations governing who can modify these products, including certification and credentialing processes, are needed.
“We need to think about a similar framework in the same way that we give people permission to use their cars,” she said. [for people] … to actually create, improve, audit and edit these open-trained language models. “
Some leaders at companies like Google, which keeps chatbot Bard under lock and key, see open source software as an existential threat to their businesses. This is because the large public language models are becoming almost as proficient as our own language models.
‘We are not in a position to win this’ [AI] It’s an arms race, not OpenAI,” the Google engineer wrote in a note posted on the tech site Semianalysis in May. “Of course, I’m talking about open source. To be clear, they’re catching up to us…our model still has a slight lead in quality, but the gap is closing surprisingly quickly. It’s coming.”
Nathan Benaitch, general partner at Air Street Capital, a London-based AI-focused venture capital firm, said many of the big advances in the tech industry over the decades have been driven by open-source technologies, including today’s AI language models. pointed out that it was made possible by
“If there are only a few companies building the most powerful AI models, they will only target the biggest use cases,” said Benaich, noting that diversity in research is good for society as a whole. added that it was a benefit.
Cognitive scientist Gary Marcus, who testified before Congress on AI regulations in May, countered that accelerating AI innovation might not be a good thing given the risks the technology could pose to society. .
“We are not open sourcing nuclear weapons,” Marcus said. “Current AI is still pretty limited, but things may change.”
