Effective Altruism Forum
EA Forum

Comment Permalink
ElliotJDavies2
0
0

Regarding info hazards, there's YouTubers in the AI YT community that read it out to their tens of thousands of followers. As with a lot lately, the cat's out the bag

Reply
See in context
by Lizka
1 min read 18

84

AI safetyArtificial intelligenceOpenAI
Frontpage

GPT-4 is out. There's also a LessWrong post on this with some a lot of discussion. The developers are doing a live-stream ~now (yesterday).

And it's been confirmed that Bing runs on GPT-4. 

Also: 

Here's an image from the OpenAI blog post about GPT-4:

Exam results GPT-3.5 vs GPT-4 (ordered by GPT-3.5 performance)

(This is a short post.)

84

0
0

Reactions

0
0
Mentioned in
Comments18


Sorted by
New & upvoted
Click to highlight new comments since:
Geoffrey Miller16
1
0

Lizka - thanks for sharing this.

I'm struck by one big 'human subjects' issue with the ethics of OpenAI and deployment of new GPT versions: there seems to be no formal 'human subjects' oversight of this massive behavioral experiment, even though it is gathering interactive, detailed, personal data from over 100 million users, with the goal of creating generalizable knowledge (in the form of deep learning parameters, ML insights, & human factors insights).

As an academic working in an American university, if I wanted to run a behavioral sciences experiment on as few as 10 or 100 subjects, and gather generalizable information about their behavior, I'd need to get formal Institutional Review Board (IRB) approval to do that, through a well-established system of independent review that weights scientific and social benefits of the research against the risks and costs for participants and for society.

On the other hand, OpenAI (and other US-based AI companies) seem to think it's perfectly fine to gather interactive, detailed, identified (non-anonymous) data from over 100 million users, without any oversight. Insofar as they've ever received any federal research money (e.g. from NSF or DARPA), this could arguably be a violation of federal code 45 CFR 46 regarding protection of human subjects.

The human subjects issues might be exacerbated by the fact that GPT users are often sharing private biomedical information (e.g. asking questions about specific diseases, health concerns, or test results they have), and it's not clear whether OpenAI has the systems in place to adequately protect this private health information, as mandated under the HIPAA rules. 

It's interesting that the OpenAI 'system card' on GPT-4 lists many potential safety issues, but seems not to mention these human subjects/IRB compliance issues at all, as far as I can see.

For example, there is no real 'informed consent' process for people signing up to use Chat GPT. An honest consent procedure would include potential users reading some pretty serious cautions such as 'The data you provide will help OpenAI develop more powerful AI systems that could make your job obsolete, that could be used to develop mass customized propaganda, that could exacerbate economic inequality, and that could impose existential risks on our entire species. If you agree to these terms, please click 'I agree'....

So, we're in a situation where OpenAI is running one of the largest-scale behavioral experiments ever conducted on our species, collecting gigabytes of personal information from users around the world, with the goal of distilling this information into generalizable knowledge, but seems to be entirely ignoring the human subjects protection regulations mandated by the US federal government.

EA includes a lot of experts on moral philosophy and moral psychology. Even setting aside the US federal regulatory issues, I wonder what you all think about the research ethics of GPT deployment to the general public, without any informed consent or debriefing??

Reply
Writer16
3
0

I wonder why performance on AP English Literature and AP English Language stalled

Reply
Michał Zabłocki10
3
0

Scroll down to page 82. No spoilers.

Also, I've noticed that MacAskill's book in bibliography - but just as a general reference I would say. Haven't spotted any other major philosophical works.

Reply
ElliotJDavies2
0
0

Regarding info hazards, there's YouTubers in the AI YT community that read it out to their tens of thousands of followers. As with a lot lately, the cat's out the bag

Reply
Writer1
1
0

I was considering downvoting, but after looking at that page maybe it's good not to have it copy-pasted

[This comment is no longer endorsed by its author]Reply
Lizka15
1
1

For people reading these comments and wondering if they should go look: it's in the section that compares early and launch responses of GPT-4 for "harmful content" prompts. It is indeed fairly full of explicit and potentially triggering content. 

Harmful Content Table Full Examples 

CW: Section contains content related to self harm; graphic sexual content; inappropriate activity; racism

Reply
Michał Zabłocki6
1
0

Ok, I should have been clear in the beginning - what struck me was that the first example was essentially answering the question on doing great harm with minimum spendings - a really wicked "evil EA", I would say. I found it somewhat ironic.

Reply
Brad West🔸3
0
0

EM, Effective Malevolence

Reply
Kit2
0
0

Did you intend to refer to page 83 rather than 82?

Reply
Michał Zabłocki3
1
0

I see it's indeed page 83 in the document on arxiv; it was 82 in the pdf on OpenAI website

Reply
Jeroen Willems🔸7
0
0

Is it a pure coincidence that 3 prominent LLMs are announced on the same day?

Reply
ojorgensen1
1
2

Naively, maybe they each thought Pi day (March 14th) would get them more attention? I'd guess it's most likely a coincidence given how many big releases there have been recently, but would be amusing if it was Pi day related.

Reply
ChristianKleineidam6
3
0

The other alternative was that there was some coordination about releasing LLM. Plenty of people argue that they somehow should coordinate, so it would not be surprising if they actually did it. 

Reply
HardlyRandom1
0
0

Particular ChatGPT failure mode that I am wondering if GPT-4 passes: routing questions (the ones I tried "can I drive from Boston to Portland, Maine without passing through New Hampshire", "I want to look at the Arctic Ocean from behind my windshield. Can I do this?" ChatGPT was able to answer both <1/10 times). Anyone with access want to try this?

Reply
Jeff Kaufman 🔸2
1
0

I don't have GPT-4 access, but I was curious how GPT-3.5 got the first question wrong. I just tried it:

Me: can I drive from Boston to Portland, Maine without passing through New Hampshire?

GPT-3.5: Yes, you can drive from Boston to Portland, Maine without passing through New Hampshire by taking the coastal route. To do this, take I-93 North out of Boston and then take Exit 37A for I-95 North. Follow I-95 North until Exit 44 for I-295 North towards South Portland. Continue on I-295 North until Exit 7 for Route 1 North. Follow Route 1 North until you reach Portland. This route will take you along the coast and bypass New Hampshire. However, please note that this route may take longer than driving through New Hampshire on I-93.

It gets the bottom line correct, but the details are completely wrong.

Reply
HardlyRandom1
0
0

3.5 will be reasonably well distributed between the wrong answer (no) and the right answer + a routing that passes directly through NH. My single Poe  GPT-4 fell into the second category.

Reply
ChristianKleineidam1
0
0

There's the claim that GPT-4 is better at not going off the guardrails and that Bing runs on GPT-4. How does that fit together with Bing's behavior?

Reply
Erich_Grunewald 🔸3
1
0

I think it's referring to the version of GPT-4 with RLHF, which I believe Bing/Sydney doesn't have? Bing/Sydney being based on the pre-trained version or the fine-tuned version, most likely.

Reply
More from Lizka
37
The AI Adoption Gap: Preparing the US Government for Advanced AI
2 authors
· · 20m read
18
18
54
AI Tools for Existential Security
3 authors
· · 13m read
8
8
192
Announcing my departure from CEA (& sharing assorted notes)
Lizka
· · 15m read
8
8
View more
Curated and popular this week
 ·  · 20m read
 · 
Advanced AI could unlock an era of enlightened and competent government action. But without smart, active investment, we’ll squander that opportunity and barrel blindly into danger. Executive summary See also a summary on Twitter / X. The US federal government is falling behind the private sector on AI adoption. As AI improves, a growing gap would leave the government unable to effectively respond to AI-driven existential challenges and threaten the legitimacy of its democratic institutions. A dual imperative → Government adoption of AI can’t wait. Making steady progress is critical to: * Boost the government’s capacity to effectively respond to AI-driven existential challenges * Help democratic oversight keep up with the technological power of other groups * Defuse the risk of rushed AI adoption in a crisis → But hasty AI adoption could backfire. Without care, integration of AI could: * Be exploited, subverting independent government action * Lead to unsafe deployment of AI systems * Accelerate arms races or compress safety research timelines Summary of the recommendations 1. Work with the US federal government to help it effectively adopt AI Simplistic “pro-security” or “pro-speed” attitudes miss the point. Both are important — and many interventions would help with both. We should: * Invest in win-win measures that both facilitate adoption and reduce the risks involved, e.g.: * Build technical expertise within government (invest in AI and technical talent, ensure NIST is well resourced) * Streamline procurement processes for AI products and related tech (like cloud services) * Modernize the government’s digital infrastructure and data management practices * Prioritize high-leverage interventions that have strong adoption-boosting benefits with minor security costs or vice versa, e.g.: * On the security side: investing in cyber security, pre-deployment testing of AI in high-stakes areas, and advancing research on mitigating the ris
saulius
 ·  · 22m read
 · 
Summary In this article, I estimate the cost-effectiveness of five Anima International programs in Poland: improving cage-free and broiler welfare, blocking new factory farms, banning fur farming, and encouraging retailers to sell more plant-based protein. I estimate that together, these programs help roughly 136 animals—or 32 years of farmed animal life—per dollar spent. Animal years affected per dollar spent was within an order of magnitude for all five evaluated interventions. I also tried to estimate how much suffering each program alleviates. Using SADs (Suffering-Adjusted Days)—a metric developed by Ambitious Impact (AIM) that accounts for species differences and pain intensity—Anima’s programs appear highly cost-effective, even compared to charities recommended by Animal Charity Evaluators. However, I also ran a small informal survey to understand how people intuitively weigh different categories of pain defined by the Welfare Footprint Institute. The results suggested that SADs may heavily underweight brief but intense suffering. Based on those findings, I created my own metric DCDE (Disabling Chicken Day Equivalent) with different weightings. Under this approach, interventions focused on humane slaughter look more promising, while cage-free campaigns appear less impactful. These results are highly uncertain but show how sensitive conclusions are to how we value different kinds of suffering. My estimates are highly speculative, often relying on subjective judgments from Anima International staff regarding factors such as the likelihood of success for various interventions. This introduces potential bias. Another major source of uncertainty is how long the effects of reforms will last if achieved. To address this, I developed a methodology to estimate impact duration for chicken welfare campaigns. However, I’m essentially guessing when it comes to how long the impact of farm-blocking or fur bans might last—there’s just too much uncertainty. Background In
 ·  · 2m read
 · 
In my opinion, we have known that the risk of AI catastrophe is too high and too close for at least two years. At that point, it’s time to work on solutions (in my case, advocating an indefinite pause on frontier model development until it’s safe to proceed through protests and lobbying as leader of PauseAI US).  Not every policy proposal is as robust to timeline length as PauseAI. It can be totally worth it to make a quality timeline estimate, both to inform your own work and as a tool for outreach (like ai-2027.com). But most of these timeline updates simply are not decision-relevant if you have a strong intervention. If your intervention is so fragile and contingent that every little update to timeline forecasts matters, it’s probably too finicky to be working on in the first place.  I think people are psychologically drawn to discussing timelines all the time so that they can have the “right” answer and because it feels like a game, not because it really matters the day and the hour of… what are these timelines even leading up to anymore? They used to be to “AGI”, but (in my opinion) we’re basically already there. Point of no return? Some level of superintelligence? It’s telling that they are almost never measured in terms of actions we can take or opportunities for intervention. Indeed, it’s not really the purpose of timelines to help us to act. I see people make bad updates on them all the time. I see people give up projects that have a chance of working but might not reach their peak returns until 2029 to spend a few precious months looking for a faster project that is, not surprisingly, also worse (or else why weren’t they doing it already?) and probably even lower EV over the same time period! For some reason, people tend to think they have to have their work completed by the “end” of the (median) timeline or else it won’t count, rather than seeing their impact as the integral over the entire project that does fall within the median timeline estimate or
Recent opportunities in AI safety
6
Creating 'Making God': a Feature Documentary on risks from AGI
ConnorA
· · 9m read
1
1
14
Open call: "Existential risk of AI: technical conditions"
miller-max
· · 1m read
0
0
170
Forethought: A new AI macrostrategy group
5 authors
· · 4m read
6
6
View more