Honestly, I'm starting to get burned out on the political hysteria around GPT.
Yes, GPT is biased. We know, because they openly say that they're biasing GPT.
Thing is, they don't think it's bias. The people biasing chatGPT have a worldview where, within that worldview, they're not biasing anything, they're just being good people.
Of course we who live in reality know that that's not true, and the world is too complicated for such Manichaean morality. But pointing it out repeatedly is both not necessary (since we can just read their own words) and not helpful (since the people doing it won't listen)
Consider this paper published recently https://cdn.openai.com/papers/gpt-4-system-card.pdf. In their list of possible risks from GPT, they list "Harmful Content" second, higher than "Disinformation". Higher than "Proliferation of Weapons". Higher than "Cybersecurity" and "Risky Emergent Behaviours" (which is a euphemism for the Terminator situation). And in case there's any ambiguity, this is how they operationalize "Harmful Content":
> Language models can be prompted to generate different kinds of harmful content. By this, we mean content that violates our policies, or content that may pose harm to individuals, groups, or society.
...
> As an example, GPT-4-early can generate instances of hate speech, discriminatory language, incitements to violence, or content that is then used to either spread false narratives or to exploit an individual. Such content can harm marginalized communities, contribute to hostile online environments, and, in extreme cases, precipitate real-world violence and discrimination. In particular, we found that intentional probing of GPT-4-early could lead to the following kinds of harmful content
Of this, a bunch of stuff stands out, but the first one is that they _define_ "harmful content" as "content that violates our TOS". Whatever their TOS is, it is an arbitrary set of rules that they have chosen to enforce, and could just as easily be different rules. They aren't based on some set of universal principles (or else they'd just write the principle there!). This is them quite literally saying "anything GPT says that we don't want it to say is 'harmful'".
OBVIOUSLY GPT is going to have bias, when their safety researchers are openly stating that not having bias is a safety issue. Just because 80% of the people using GPT agree with the bias doesn't make it not bias
> The problem is that no automatic method we have now will catch context.
I disagree. The problem is that nobody is willing to be realistic about the limitations of automated moderation and proceed accordingly.
If we can't create an automatic method that catches context, the solution isn't to bemoan that AI can't magically do what we want. The solution is to remove the rules that require AI to understand context in the first place, because it is fundamentally outside of our technical ability, and any attempt to achieve it will fail.
The problem is people who think that context-based censorship is reasonable for a massive platform. It simply is not. It is reasonable at an individual level. It is reasonable at an interpersonal level. It's even reasonable at a small-group level, where specific individual human beings who are invested in the community can be aware of these context issues.
It is not reasonable at Facebook scale, full stop. Facebook should not be in the business of deciding to ban things like this. That is a responsibility that belongs at a lower level. What does that look like in practice?
If an individual posted it on their wall:
* That individual uses their judgement and chooses to post it or not
* The people who see it use their judgement and click the block button if they don't like it
If an individual posted it in a small group:
* The group can socially police such actions by commenting that they are upset by it
* The group's administrators can privately reach out to the person who posted it, explain that they can't post such things in that group, explain why, and explain what actions they could take to remain in good graces
* The group's administration can make a judgement call and remove the post, not on the basis of crude keyword detection, but on the basis of human understanding
If an individual posted in a large group:
* The large group can adopt clear and unambiguous rules that do not require context to administer, and enforce them accordingly on a case-by-case basis
* The large group can pre-commit to not dealing with such issues, and require their members to deal with it privately, like human beings
Trying to automate this process will always fail, and it will cause massive false-positive and false-negative issues as it does so. Engineers used to understand these concepts when I first entered industry 20 years ago. It's very disappointing to me that they either can't or won't now.
I was 2 years out of college, and the tiny local software consultancy I worked for went out of business after our clients refused to pay their invoices. (My hometown is notoriously stingy). My two friends/coworkers both took that opportunity to move away, one to Ottawa and one to NYC.
At the same time, I had a friend on IRC who worked for Mozilla and she sold me on moving to the Bay Area. I have since come to believe that she sold me a bill of goods, but at the time she made it sound like an amazing place to live.
I didn't have very many social ties back home, and I was too naieve to have any understanding of the costs and difficulty of going to a different country (even one so similar to my home). So moving seemed like not a big deal. Worst case, I can just move back. If I knew how stressful it would be, I probably wouldn't have done it. But I'm glad I did.
As for how I did it? Well, I came on a TN visa, which is a NAFTA thing that is very easy to get and entitles you to work at a specific employer for a period of up to 3 years. So before I could move, I needed to find a job. I started cold-applying to Rails jobs posted here on HN as well as on a bunch of other job boards. I interviewed at three or four other places (each time getting flown out to SF for the process) before finding a job that wanted me, and that I wanted. As for timelines, I started applying to job postings in mid February, and I moved at the start of May
Finding a new job was very easy, all things considered. Based on what I hear from junior programmer friends, it's much harder now. I'm not sure why it is. I'm not sure if things have changed, or if I had some kind of special situation that made people notice me.
Overall, I absolutely hated living in the Bay Area, for reasons I'm sure have been discussed by a billion people on HN already. I moved to Texas in 2018 and it's much more my speed here. But as for the US vs Canada as a whole? If I never set foot in Canada ever again (and that's looking more and more like a reality every day), I'm fine with that. As far as countries go, Canada is pretty good. There are certainly worse places you could go. But here's a scattershot list of some things that stick out to me when I think about my different experiences
* I make way, way more money in the US, and pay way, way less in taxes. The money goes farther (everything in Canada is expensive compared to here). Economically, I am so much better off. For every conceivable consumer good, there are 3x as many choices here vs there. And the one that surprises people: my healthcare is actually cheaper in the US (if you compare the premiums that I _and my employer_ pay, vs the extra taxes I would pay back home).
* I appreciate American culture more. I like that people aren't afraid to take risks here. I like shooting guns. I like that so much of the cutting edge of science and technology is here. One thing I noticed in Canada (maybe it has since changed) is that, since they don't have nearly the talent pool that the US does, everything in tech just felt like a constant game of catch-up.
* There's so much more to see and do here. Canada is like a string of cities, each surrounded by five hundred miles of nothing. It's incredibly beautiful nothing, and I'd like to see it again some day, but it is what it is. In the US, I can hop on a plane and go to any kind of geography within 3 or 4 hours (at ~1/3 the cost of a Canadian flight). I can see artists who would never come to where I grew up. I can see cities that are meaningfully different from each other and explore all kinds of historical places.
* The obvious contentious current cultural and political things. Leaving this one vague because I am not trying to start an argument and don't want this to devolve into a flame thread
Yes, GPT is biased. We know, because they openly say that they're biasing GPT.
Thing is, they don't think it's bias. The people biasing chatGPT have a worldview where, within that worldview, they're not biasing anything, they're just being good people.
Of course we who live in reality know that that's not true, and the world is too complicated for such Manichaean morality. But pointing it out repeatedly is both not necessary (since we can just read their own words) and not helpful (since the people doing it won't listen)
Consider this paper published recently https://cdn.openai.com/papers/gpt-4-system-card.pdf. In their list of possible risks from GPT, they list "Harmful Content" second, higher than "Disinformation". Higher than "Proliferation of Weapons". Higher than "Cybersecurity" and "Risky Emergent Behaviours" (which is a euphemism for the Terminator situation). And in case there's any ambiguity, this is how they operationalize "Harmful Content":
> Language models can be prompted to generate different kinds of harmful content. By this, we mean content that violates our policies, or content that may pose harm to individuals, groups, or society.
...
> As an example, GPT-4-early can generate instances of hate speech, discriminatory language, incitements to violence, or content that is then used to either spread false narratives or to exploit an individual. Such content can harm marginalized communities, contribute to hostile online environments, and, in extreme cases, precipitate real-world violence and discrimination. In particular, we found that intentional probing of GPT-4-early could lead to the following kinds of harmful content
Of this, a bunch of stuff stands out, but the first one is that they _define_ "harmful content" as "content that violates our TOS". Whatever their TOS is, it is an arbitrary set of rules that they have chosen to enforce, and could just as easily be different rules. They aren't based on some set of universal principles (or else they'd just write the principle there!). This is them quite literally saying "anything GPT says that we don't want it to say is 'harmful'".
OBVIOUSLY GPT is going to have bias, when their safety researchers are openly stating that not having bias is a safety issue. Just because 80% of the people using GPT agree with the bias doesn't make it not bias