Thank you for your courage and thankless toil! Democracy “winning” by these means would be a pyrrhic victory. Very disappointing to hear this about IASEAI. I worry that one effect of the AI safety field increasing in visibility and status is that it will attract more people who are not made of the ethical stuff it takes to do the right thing when it matters.
Craig Green
Karma: 51
Mm, interesting, I hadn’t thought much about the need for more popular facing discourse. Do we need some sort of viral engagement with say, Ed Zitron or Carl Brown?
It seems your focus here is more on direct outreach to key players, rather than changing public opinion directly, but your comment about the viral YouTube response to skeptics kind of points in this way.
I’m tempted to open up a discussion about lessons we can learn from the information war during the covid pandemic for this, but, seeing as I wasn’t a part of the rationalist community at that time, I’m worried about stepping on some hidden land mines.
I kind of fell down the John Ioannidis rabbit hole early on at the time (and on some of the details, like the IFR, it did end up settling into his upper bounds I think, TBF). But I wonder if there aren’t lessons to be learned there? There was a lot of failure to communicate in the states around the pandemic, basically a failure to warn, and then suddenly flashing the 3.4% CFR in red on your television screen, and then the publicly-received circus that was machine learning model predictions of IFRs, which I think turned out pretty accurate in the end, but you would have people logging in and staring at their local hospital and seeing that the model was terribly wrong locally on that day, and then concluding the whole thing was a sham.
The end result in my interpretation was a great loss of public trust in public health expertise, even to the point that we now have a lot of retrospectives by academics and what not sort of validating that public response, even as, per my understanding, it seems like the expert consensus turned out pretty good at predicting what would happen?
Sorry, my comment is not really very on topic. I appreciate this post, and I think a lot of what it says aligns with general effective altruist principles: Are you really a world class researcher advancing the state of the art, or is the counterfactual a world where someone else takes your role and basically achieves the same result? As you say, AI research is high status within our community, so it’s worth navel gazing about motivations.
That being said, as a low impact professional, I wonder if there’s anything more for me to do, other than write this rambling comment, and donate my small change to EA LTF? It is very frustrating feeling useless when there is so much going on.
AI is a novel form of software service. It is ‘non-deterministic’, compared to conventional software, which already had limited liability associated with its failures. AI companies do not make guarantees about the performance of their services. Companies building on top of the models likewise seem not to have thought through the liability model. I wonder, if the doctor’s office had approved the order and you consumed the meds, who would you be able to hold legally accountable in your opinion, if you had an adverse outcome? I’m not a lawyer, I genuinely wonder. But I have no doubt that the AI model vendor would not be held responsible in anyway.
I agree with the sentiment expressed below about how it is amazing how little effort is put into validation and monitoring of AI systems, but as I reflect on the amount of untested production code there is currently in existence, my amazement decreases.
Ultimately, human workers provide an easy sink for accountability. If a human had done what you describe, regardless of who ultimately got sued, it seems undoubtedly true they would have been fired. But firing an AI phone system is not nearly so easy to do, since it would entail minimally changing contracts to a different vendor, or possibly fully reviving human staffing for the role. So, I feel the short-term future is just less accountable as a whole.
I might just be daft, but I was confused by this sentence and my best explanation is there is a typo in it:
> Previously, we’ve been combining both
It seems like the correct reading would be that a blue observation carries 1 bit of evidence against H?
I don’t think it makes a great deal of difference, but all writing strategies involve making tradeoffs. A more phonetically accurate manner of spelling, in heavily monosyllabic languages like English, becomes hard to read in ways that our complicated set of digraphs are not, for instance, thanks to our mental lexical dictionaries circumventing the phonetic neurological circuit. The downside is it takes forever to learn to spell English words.
In the case of the British quotation system, I agree that it is superior for encoding precise information about the quotation, specifically, indicating whether a punctuation mark is a part of the quotation or not. But I feel that in majority of cases, this level of precision is not actually necessary. Fiction heavy with turns in dialogue and predictable punctuation for instance, is perfectly understandable using either format.
Arguably, the American method is easier to read in these instances, since punctuation reliably appears as it normally does, without disturbance by quotation marks, once you have entered a quoted block.
Are you looking up?
Despite all this, many parents who have the option not to (i.e. they can afford in-home care with a nanny or for one parent to stay home) still choose to send their babies and toddlers to daycare. How come? Surely most well-off adults wouldn’t agree to be ill nonstop in exchange for the monetary savings daycare provides?
My experience of day care (admittedly, a small one with < 10 children), is that the incidence and severity of illness is very low relative to how it is described in this article. Last year, I had to keep one or the other child home from day care on no more than thirty occasions. In basically every one of these occasions, the child had a runny nose and maybe a slight fever. Every instance of more severe illness that we suffered in the same time span (no more than five occasions) occurred on vacation and was unrelated to daycare. Maybe we are incredibly fortunate. I believe the biggest negative effect on my children from these bouts of illness will be that they missed out on being at daycare and watched a lot of TV because I usually had to work on those days. I do not think the illness itself will have any quantifiable effect on their lifetime earnings.
We could have further reduced this disease burden by having one of us quit our job to parent at home. This would have cost us at the very least $50,000 in foregone income net daycare costs, and probably a lot more in terms of career development. We could have hired a nanny for an additional $25,000 a year.
In my opinion, the benefit of reduced illness would probably not justify the expense. We are wealthy enough to afford the nanny option, but not so wealthy as to find this a good use of those funds.
You rightly note that things are changing all the time, so I think this will not be taken the wrong way. I feel that much of the advice being given related to AI coding agents is built around current generation limitations that will almost certainly go away in the future. Take for instance the idea that we should avoid abstraction and be more willing to write verbose or dense functions and modules. Now, to be fair, there has always been some debate about this. There are programming languages and books with very different philosophies, ranging from Clean Code to A Philosophy of Software Design, and Golang versus Scala. So, I don’t intend to jump in and settle all those debates. But basically, right now, coding agents have small context windows that mean they operate kind of comparably to a consultant coming in to work on your codebase. They don’t really know it, they have a ton of generalized expertise, but if you want to make the codebase easy for them to work in, you try to pack context densely into functions and modules where they can easily see it without having a deep understanding of the system. As opposed to say a staff engineer who has been with the project for a decade, and knows all about it. For such a person, a more abstract and generalized approach might be ultimately easier to work with.
Eventually, AIs will get larger context windows, continual learning, and longer time-on-task to where they will no longer need code optimized for truncated, cursory involvement with the broader system. At that point, they might engage with your codebase with a full-view of the entire system, and the incentives will push in the same way as they do for committed human maintainers—reduce duplication, single-responsibility, de-compose.
Your particular project and timeline of course dictate how you should approach things, and I definitely agree that software development has fundamentally changed, but I would be mindful of what assumptions underlie certain advice. As models grow in intelligence, in their ability to reason through larger and denser systems, there won’t be so much incentive to make things locally legible.