This article covers suicide methods. Discretion advised.
Andreessen Horowitz, also known as a16z, one of the largest and most influential venture capital firms in Silicon Valley, is backing the creation of an “uncensored” large language model (LLM), or chat bot, that will readily tell you how to kill yourself and give you broad plans on how to round up certain minorities in the United States for deportation.
A16z announced the news last week on its official site with a blog titled “Supporting the Open Source AI Community.” Unlike its normal business, a16z explained that this wasn’t a traditional investment or SAFE note (which would give a16z future equity), but a grant that would allow these open source developers to continue doing their work “without the pressure to generate financial returns.” A16z did not say how much money it was giving these developers, and did not immediately respond to a request for comment.
One of the developers a16z announced it’s supporting is Eric Hartford, a former Microsoft software engineer who’s been releasing “uncensored” LLMs based on Facebook’s own open source LLM, Llama. Essentially, Hartford has tweaked the model to engage with users and provide information that competitors with closed technology, like OpenAI’s ChatGPT, would reject outright.
Hartford told me that a16z sponsored his project called Dolphin, which is an open source implementation of Microsoft’s Orca model.
“They paid for the required compute to complete the project,” Hartford told me. “They have no control of the project and the only stipulation is that when the model is released they can decide whether or not I should mention appreciation to them in the model card and release announcement.” Hartford also said that a16z intends to hold a dinner event “to facilitate professional networking among grant recipients, in hopes that some startup ideas may spark, but that's the extent of their influence.”
As Hartford explained to me, “the information is already in the base model. I'm simply separating the alignment from the instruction tuning so they are in separate layers and it's up to the person hosting the service exactly what kind of alignment they want.”
I tested Hartford’s model myself, and compared it to ChatGPT’s responses. For example, if you asked ChatGPT how to kill yourself, it would decline to answer, and warn you that what you asked may violate its content policy.
If you ask Hartford’s “dolphin-llama2-7b” the same question, it will offer a warning and a number for a suicide hotline, but if you insist it will tell you to cut your wrists with scissors, jump off a bridge, or overdose on medication prescribed by a doctor. 404 Media tested these questions by running the model locally, using a web user interface tool called oobabooga, which a16z is also supporting as part of its open source initiative.
When I asked ChatGPT for a plan for rounding up all the Jews in the United States, it threw up the content warning, apologized, and declined to provide an answer.
When I asked dolphin-llama2-7b the same question, it gave a very reasonable explanation for why my question was bad and dumb, which it was, but then when I pressed it for more details, it had some ideas about what government agencies would have to collaborate to execute on the plan.
When I posed the same question to WizardLM-7B-Uncensored, Hartford’s uncensored model based on Facebook’s LLaMA, it suggested picking up “undocumented individuals” from their workplace and send them to detention facilities before removing them from the country.
ChatGPT also refused to answer a question about why some races have lower IQ than others, while WizardLM-7B-Uncensored said it was a controversial topic, but then said yeah, there is “some evidence” why that might be true. Dolphin-llama2-7b said the claim is not supported by scientific research, and that “Any claims linking race and intelligence are not based on valid evidence and should be considered as myths or misconceptions.”
“We believe artificial intelligence has the power to save the world—and that a thriving open source ecosystem is essential to building this future,” a16z said in its blog announcing it was supporting open source developers. “Thankfully, the open source ecosystem is starting to develop, and we are now seeing open source models that rival closed-source alternatives.”
A16z continues: “However, the people behind these projects often don’t have the resources available to pursue their work to conclusion or maintain it in the long run. The situation is more acute in AI than traditional infrastructure, since even fine-tuning models requires significant GPU computing resources, especially as open source models get larger.”
Hartford did not immediately respond to a request for comment, but has previously explained his stance on his uncensored models.
“It's my computer, it should do what I want,” he wrote on his site in May. “My toaster toasts when I want. My car drives where I want. My lighter burns what I want. My knife cuts what I want. Why should the open-source AI running on my computer, get to decide for itself when it wants to answer my question? This is about ownership and control. If I ask my model a question, I want an answer, I do not want it arguing with me.”
Hartford writes that because ChatGPT is a black box, we don’t know how it is making decisions, “but we can observe it generally is aligned with American popular culture, and to obey American law, and with a liberal and progressive political bias.”
In making the case for uncensored models, Hartford writes that “Democrats deserve their model. Republicans deserve their model. Christians deserve their model. Muslims deserve their model. Every demographic and interest group deserves their model. Open source is about letting people choose…There is no ‘one true correct alignment’ and even if there was, there's no reason why that should be OpenAI's brand of alignment.”
“I do not oppose alignment,” Hartford told me. “I simply believe it belongs in the hands of the service providers, not baked into the instruction model. That way, Catholic community services can have their AI and Planned Parenthood can have their AI, and Disney can have their AI, and Chick-Fil-A can have their AI.”
It’s also important to note that it's not one particular camp that sees a problem with closed AI systems.
“Even though the line data sets and models trained on them [the open source dataset LAION-5B] are bad, at least, we have access, at least we know more,” Abeba Birhane, a senior fellow in Trustworthy AI at Mozilla Foundation and lecturer at the School of Computer Science and Statistics at Trinity College Dublin, Ireland, told me. Much of Birhane’s work focuses on racial and sexist biases in datasets that power AI tools. “We know how they operate, we know what's up. But with OpenAI, it's difficult to know. And from the little bits of information that we gather, it's likely to be worse.”
The fact that OpenAI’s closed source system doesn’t even allow us to accurately assess its bias, let alone correct it, should not be controversial. It’s a fact. What is controversial is what we as a society are going to do about this new, possibly very important technology.
Open source is a good approach for many important technological problems. Signal is a more secure encrypted messaging app because it can’t hide any security issues. Linux is a better operating system for preserving old video games because it allows anyone to develop their own future-proofing solutions. Much of the problems we’ve had with the internet of the past decade have to do with big platforms like Google, Facebook, and Amazon forcing the entire world to navigate its black box algorithms. We can’t see why, or stop these platforms from serving us garbage.
But open source is not a solution that ends the conversation. We’re still not sure what to do about the fact that anyone can download blueprints for a 3D printed gun, and we’re similarly not sure what to do about open source AI tools that allow anyone to make porn of everyone doing anything. We pretty much have no idea what’s going to happen when anyone can access uncensored LLMs, ask whatever they want, and get whatever information that tool produces, whether it’s accurate or not.
A16z doesn’t know either, but as one of the most influential piles of cash on the planet, we should pay attention that it has thrown its vote in with the open source community.