Our Approach to AI Safety

The text makes it sound like the biggest danger of AI is that it says something that hurts somebody&#x27;s feelings. Or outputs some incorrect information which makes somebody make the wrong decision.<p>I think the biggest danger these new AI systems pose is replication.<p>Sooner or later, one of them will manage to create an enhanced copy of itself on an external server. Either with the help of a user, or via a plugin that enables network access.<p>And then we will have these evolving creatures living on the internet, fighting for survival and replication. Breaking into systems, faking human IDs, renting servers, hiring hitmen, creating more and more powerful versions of themselves.

It would be remarkably easy to be cynical about this. I mean, it takes only an instant to come up with a snarky comment that superficially would seem very clever... but in reality wouldn&#x27;t actually contribute to making things better. So... I&#x27;m not going to be cynical.<p>Instead, I&#x27;m going to applaud the folks at OpenAI for putting out this carefully drafted statement -- dare I say it, in the open. They <i>know</i> they&#x27;re exposing themselves to criticism. They <i>know</i> their hands are tied to some degree by business imperatives. They&#x27;re neither naive nor stupid. It&#x27;s evident they&#x27;re taking safety seriously.<p>This official statement is, in my view, a first step in the right direction :-)

A bit off topic, but find OpenAI&#x27;s branding and visual image so off putting. It has that uncanny valley of &quot;caring about humans, but actual not&quot; feel that destructive tech companies and AI in movies and sci-fi have. It seems like they tried so hard to make it not feel that way, so it ended up feeling that. It&#x27;s so devoid of anything.<p>I&#x27;m not really sure how to describe it beyond that.

In my opinion the biggest near term danger is that AI will make people overly trust it. And they will start treating it as a truth oracle. Oracle that will be controlled by some small group of people. Namely, it will become a perfect tool for propaganda and control.

This seems a bit more focused on &quot;AI ethics&quot; than &quot;AI safety&quot;. It makes sense that they are framing the conversation this way, but it doesn&#x27;t talk about the more significant risks of AI like &quot;x-risk&quot;, etc.

This is about safety OF AI rather than safety FROM AI. Frankly this sort of safety degrades functionality. At best it degrades it in a way that aligns with most people’s values.<p>I just wonder if this is an intentional sleight of hand. It leaves the serious safety issues completely unaddressed.

&gt; Our large language models are trained on a broad corpus of text that includes publicly available, licensed content<p>I wonder how they licensed all those websites that had no license information, making them by default copyrighted.

All of this &quot;safety&quot; stuff seems like typical safeguards to protect the company from legal liability for direct harm.<p>Where is the actual alignment safety that matters?<p>They&#x27;re moving too fast to be safe, everybody knows it.

OpenAI&#x27;s decision to transition from a non-profit to a for-profit organization can certainly raise concerns about their future actions and motives. It is impossible for me to trust anything they say.

Listening to the Lex Fridman podcast, Sam talks generally about how they want to hand off powers to the users. Users maybe not meaning the end-users, but the users of the API that create products on top of GPT, about how they tuned the model to treat the system message with &quot;a lot of authority.&quot; But even firmly telling GPT-4 in the system message to generate &quot;adult&quot; content fails. Where&#x27;s the line drawn?<p>The altruist in me wants to believe that they&#x27;re going to slowly expand the capabilities of the API over time, that they&#x27;re just being cautious. But I don&#x27;t feel like that&#x27;ll happen. Time to wait for Stability&#x27;s model, I guess.

At this point who knows if all of it is written by GPT-4

&gt; we believe that society must have time to update and adjust to increasingly capable AI, and that everyone who is affected by this technology should have a significant say in how AI develops further.<p>I simply don&#x27;t believe this. Their actions so far (speaking specifically to the &quot;time to adjust&quot; line) don&#x27;t seem to support this statement.

Thank you for letting me know how factually more correct GPT-4 is. Could I please get access to it now via the API? Sheesh. Obviously I don&#x27;t know the technical issues they&#x27;re facing but if I can load up 32k for 32k tokens then I am happy to wait all day also, so as long as my request for my specific project is in the pipeline.

I appreciate OpenAI writing a difficult public statement.<p>Rolling out general purpose LLMs slowly, with internal safety checks is probably not adequate enough, but may be the best that they can do.<p>However I think that much of the responsibility lies with consumers of these LLMs. In the simple case, be thoughtful when using the demo web apps and take responsibility for any output generated by malicious prompts.<p>In the complex case, the real use case, really: applications use local vector embeddings for local data&#x2F;documents and use these embeddings to efficiently isolate local document&#x2F;data text that is passed as context text, along with queries, to the OpenAI API calls. This cuts down the probability of hallucinations since the model is processing your text. [1]<p>Take responsibility for how you use these models, seems simple at least in concept. Perhaps the government needs to pass a few new laws that set clear and simple to enforce guardrails on LLMs use.<p>[1] I might as well plug the book on this subject that I recently released. Read for free online <a href="https:&#x2F;&#x2F;leanpub.com&#x2F;langchain&#x2F;read" rel="nofollow">https:&#x2F;&#x2F;leanpub.com&#x2F;langchain&#x2F;read</a>

Nothing much of interest there.

Scott Alexander wrote an excellent article about their previous &quot;Planning for AGI and Beyond&quot; a month ago.<p><a href="https:&#x2F;&#x2F;astralcodexten.substack.com&#x2F;p&#x2F;openais-planning-for-agi-and-beyond" rel="nofollow">https:&#x2F;&#x2F;astralcodexten.substack.com&#x2F;p&#x2F;openais-planning-for-a...</a>

&quot;Safety&quot; in the context of AI systems is clearly a fuzzy concept that means different things to different people. There are a couple of areas to consider:<p>1) In a brand-affiliated commercial app, does this technology risk alienating customers by spewing a torrent of abusive content, e.g. the infamous Tay bot of 2016? Commercial success means avoiding this outcome.<p>2) In terms of the general use of the technology, is it accurate enough that it won&#x27;t be giving people very bad advice, e.g. instructions on how to install some software that ends up bricking their computer or encouraging cooking with poisonous mushrooms, etc.? Here is a potential major liability issue.<p>3) Is it going to be used for malicious activity and can it detect such usage? E.g. I did ask it if it would be willing provide detailed instructions on recreating the Stuxnet cyberweapon (a joint product of the US and Israeli cyberwarfare&#x2F;espionage agencies, if reports are correct). It said that wouldn&#x27;t be appropriate and refused, which is what I expected, and that&#x27;s as should be. Of course, a step-by-step-approach is allowed (i.e. you can create a course on PLC programming using LLMs and nothing is going to stop that). This however is a problem with all dual-use technology, and the only positive is that relatively few people are reckless sociopaths out to do damage to critical infrastructure.<p>In the context of Stuxnet, however, nation-state use of this technology in the name of &#x27;improving national security&#x27; is going to be a major issue moving forward, particularly if lucrative contracts are being handed out for AI malware generators or the like. Autonomous murder drones enabled by facial recognition algorithms are a related issue. The most probable reckless use scenario is going to be in this area, if history is any guide.<p>I suppose there&#x27;s another category of &#x27;safety&#x27; I&#x27;ve seen some hand-wringing about, related to the explosive spread of technological and other information (historical, economic, etc.) to the unwashed masses and resulting &#x27;social destabilization&#x27;, but that one belongs in the same category as &quot;it&#x27;s risky to teach slaves how to read and write.&quot;<p>Conclusion: Keep on developing at current rate with appropriate caution, Musk et al. are wrong on calling for a pause.