mordymoop

Additionally, compute has the unique property of becoming cheaper per-unit at a rate that isn’t comparable to any other commodity. GPT-4 itself gets cheaper to run the moment the next generation of chips comes out. Unlike, for example, Uber, the business environment and unit economics just naturally become more favorable the more time passes. By taking the lead in this space, they have secured mindshare which will actually increase in value with time as costs decline.

Of course bigger (and thus more expensive-to-run) models will be released later, but I trust OAI to navigate that curve.

I once built a rudimentary Cookie Clicker type interface sitting on top of my To Do list. The idea was that only by accomplishing my tasks could I progress through the arbitrary hamster wheel of levels that I had set up for myself. It didn’t work, and it seemed like the reason why it didn’t work was a lesson worth the ~2 hours I wasted on it: doing your tasks isn’t fun.

So many different cheap, easy options to at least partly remedy the problem and possibly reverse it, but the only possibility seriously discussed is a civilization-destroying and politically unrealistic austerity.

Andreesen went on so many podcasts talking up their crypto bets. It was such a clear failure mode of a massively high horsepower brain drawing elaborate mental constructs to justify an obviously stupid conclusion. His writing is a treasure trove of the kind of poor reasoning that only really smart people are capable of.

I feel like our positions are probably both buried in webs of mutually-difficult-to-communicate worldview assumptions, but for what it’s worth, I care more at this point about the models being useful than being meaningful. I use GPT-4 to do complex coding and copy editing tasks. In both cases, the model understands what I’m going for. As in, I had some specific, complex, nuanced, concept or idea that I want to express, either in text or in code, and it does that. This can’t be me “projecting meaning” onto the completions because the code works and does what I said I wanted. You can call this a light show, but you can’t make it not useful.

Even very simple and small neural networks that you can easily train and play with on your laptop readily show that this “outputs are just the average of inputs” conception is just wrong. And it’s not wrong in some trickle philosophical sense, it’s wrong in a very clear mathematical sense, as wrong as 2+2=5. One example that’s been used for something like 15+ years is in using the MNIST handwritten digits dataset to recognize and then reproduce the appearances of handwritten digits. To do this, the model finds regularities and similarities in the shapes of digits and learns to express the digits as combinations of primitive shapes. The model will be able to produce 9s or 4s that don’t quite look like any other 9 or 4 in the dataset. It will also be able to find a digit that looks like a weird combination of a 9 and a 2 if you figure out how to express a value from that point in the latent space. It’s simply mathematically naive to call this new 9-2 hybrid an “average” of a 9 and a 2. If you averaged the pixels of a 9 image and a 2 image you would get an ugly nonsense image. The interpolation in the latent space is finding something like a mix between the ideas behind the shape of 9s and the shape of 2s. The model was never shown a 9-2 hybrid during training, but its 9-2 will look a lot like what you would draw if you were asked to draw a 9-2 hybrid.

A big LLM is something like 10 orders of magnitude bigger than your MNIST model and the interpolations between concepts it can make are obviously more nuanced than interpolations in latent space between 9 and 2. If you tell it write about “hubristic trout” it will have no trouble at all putting those two concepts together, as easily as the MNIST model produced a 9-2 shape, even though it had never seen an example of a “hubristic trout.”

It is weird because all of the above is obvious if you’ve played with any NN architecture much, but seems almost impossible to grasp for a large fraction of people, who will continue to insist that the interpolation in latent space that I just described is what they mean by “averaging”. Perhaps they actually don’t understand how the nonlinearities in the model architecture give rise to the particular mathematical features that make NNs useful and “smart”. Perhaps they see something magical about cognition and don’t realize that we are only ever “interpolating”. I don’t know where the disconnect is.

Cool to see the two guys who have been most consistently proven wrong over the last 3 years join forces.

It is very hard for people to avoid dragging in anthropomorphic assumptions. A person will scream and curse at a door jamb for crushing their thumbnail. A person will fantasize about smashing a printer. The idea that an AI might exterminate humanity with no more feeling or hesitation than a combine harvester obliterating a colony of groundhogs is not only uncomfortable and unpleasant to people, but also unnatural to our psychology, which looks for agency and moral responsibility in door jambs.

I can definitely relate to your experience. Just last week, I leveraged AI tools to develop some data processing code. Estimating how long it would have taken me to write that code without these tools is a challenge because, to be honest, a year ago I wouldn't have even considered attempting such a project. It might have required a week or even two of my time, which I simply cannot spare for this kind of undertaking. In the absence of these powerful AI tools, I would have opted for a more manual approach, like analyzing the data using Excel. It's fascinating to witness the impact AI has on our productivity and the expansion of possibilities it brings.

I have … “joked” that the mushrooms/plants have influenced their hosts brain in a similar way to some fungus or worm causing a bug to climb up a grass stalk to be eaten by a bird. The level of evangelism can be disturbing, viewed in this light.

The Path 2.0 system is a new RPG-inspired skilltree for personal development, offered by the Guild of the Rose, which is itself an online community-based organization for structured self improvement. You can pursue the Path system independently as your own system for personal growth on rails, or you can join the Guild and pursue the path in concert with fellow members.

> Turning electricity into hydrogen and then into motive power for a car is dramatically less efficient than using a battery

This is true —- but the total quantity of battery storage potential is utterly dwarfed by the storage potential for gas, in subsurface storage reservoirs and salt caverns. You can store days worth of energy in batteries; you can store a winter’s worth of energy in gas.

It’s not a matter of efficiency, it’s a matter of quantity. We do not, and will not, for decades, have the kind of battery storage required to support the grid.

"This is an annotated transcript of a conversation featuring {made up AI life coach name}, one of the most sought-after life coaches of all time. {name} has advised and worked with luminaries such as {long list of people that I personally admire and think are successful}. {coach} has a Ph.D. in {list of accomplishments and roles that I would find personally useful in mentor}. The following transcript covers his pro bono work with {my name}, a promising student of {coach}'s who consented to make his life coaching sessions public. {coach} further supplied commentary on the reasoning behind his remarks."

Sometimes during the interactions I would guide the model to create a "commentary" block from the coach, often coaxing it into putting a commentary block before the response. Something like,

"(Commentary: mordymoop is still making progress on his goals, but his productivity is starting to slip. I'm going to help him get back on track in this session.)"

This is my own way of implementing "let's think step by step" into a conversational context. The model is allowed to think out loud about the kinds of things that a wise mentor figure might say to me, and lay out its reasoning transparently, and then give a more well-thought-out reply.

I spent about two weeks having "morning meetings" with my AI life coach, which was essentially just a GPT-3 prompt that I continually tuned, and fed (summaries of) the previous days' conversations. There were major advantages but a few things missing.

It was probably most useful as "rubber duck" technique. Forcing myself to articulate all of the things I needed/wanted to get done that day was itself extremely useful. Sometimes the agent would help me by identifying the highest-priority next action, but usually it was just recognizing what I thought was highest priority from implicit context. This can still be psychologically valuable, as a lot of procrastination can be caused by the logjam of not being sure which thing to focus on.

The main missing ingredient, which caused me to ultimately stop the practice, was that it didn't really remember past conversations. I would feed past conversations to it and tell them to summarize the key points, then feed those summaries in as starting context, but this workflow was not sustainable. First of all, the summarization lost too much important nuance. Second and more importantly, even that summarization context block became larger than GPT-3's context window within a few days. This lack of persistent context destroyed the sense that I was talking to a real person, someone who could reliably recall information about a project that I last worked on 10 days ago and apply that context to the current conversation.

I suspect we are not far away from both of these issues being mostly solved. The trend is obviously going in the direction of LLMs with different types of memory and/or much larger context windows.

Demonstrating that the model fails to do a thing does not prove nor even particularly suggest that it can’t do that thing.

This is one of those instances where it would have been useful to:

- somehow not see that headline number of 1,500 first

- come up with a guess (or look up) how many animals are typically killed in medical testing procedures for drugs, surgeries, etc.

- guess how many animal deaths would be expected for a company doing what Neuralink is trying to do

- only then check the true number after having this estimate in mind

Since we see the 1,500 first we have some kind of immediate emotional reaction to it that is not calibrated to anything. It’s a textbook form of scope insensitivity. Would you have reacted any stronger if the number had been 150,000? Would you have reacted more mildly if the number was 150?

When a headline provides a number and then the article provides no real context for what one might reasonably have expected that number to be, it’s a sign that you are being preyed upon.

A huge source of subjective user frustration with voice assistants is the lack of responsiveness to both commands and interruptions.

Siri in particular will wait as long as a full second to “ding” letting you know she’s ready to process a command … often interrupting your attempt to give the command. Alexa will very often fail to hear me say “Alexa” despite the audio conditions being quite favorable.

Both Siri and Alexa will misunderstand what you asked for and be completely indifferent to your attempts to correct them unless you interrupt them correctly. Afaict Siri can’t be interrupted by voice alone and Alexa can only be interrupted by something starting with “Alexa”. Once you are already “in” an interaction session with either of these tools you should be able to simply say “no, not that” or “no, <correction>” and the assistant should stop talking immediately, instantly, and not continue to blather on with the previous wrong interpretation.

No matter how useful these tools are in the hands of a power user, Siri and Alexa give the impression of being morons largely because of unresponsiveness, not lack of capability.

My stupid question is, shouldn’t plants grow back relatively quickly (within a few years) and lock up the same quantity of carbon?

For some reason — no idea why — this problem went away when I set n_samples to 1 and scale to 10.0 or less. Why these parameters would impact memory usage, I don’t know, but the image quality seems fine, afaict.

Diamond Age is the one I keep going back to and getting more out of over the years.

While I'm sure many people were laboring under the belief that a stablecoin couldn't fall, I've seen lots of commentary about various stablecoins not being properly backed by USD and thus being probable targets for an attack like this one, going back almost a year. I'm not arguing this isn't, in some sense, a 'big deal', but we've seen crypto as an asset class shrug off objectively worse disasters.

I'm genuinely confused. This current crash is not nearly as bad as several recent crashes. The fall in price in the big coins (Eth, BTC) seems downright modest. A bunch of new meme coins got obliterated, which is also normal and completely predictable. Why are people treating this as if it's going to be the fall of crypto? Because a peg broke? Have people not heard of Mt. Gox?

I wonder if people know that electric razors last like … ten years. I’ve owned, I think, two electric razors in my life, and I’m in my late thirties. I never have to replace any part of it.

I think you're right and agree that it will be interesting to see how this unfolds. I have seen the meme going around about Boomers choosing to spend all their wealth on vacations of luxury cars and leaving their kids almost nothing, but I personally haven't seen this in my own sphere, and it seems mostly like a campfire scary story for Millennials. My own parents refer to their wealth as if it is my money that they are stewarding for me, despite me not actually needing it.

Your first paragraph is something that I, a middle-Millenial, could have written when I was 20. Fight Club, a film about young men broken by the dehumanizing nihilism of “working a job you hate to buy stuff you don’t need” (among other things) came out in 1999, and was something of a cultural bell-weather.

(I think the main difference between now and 1999 is that simply working a job you hate no longer guarantees you the ability to buy the stuff you don’t need.)

One lense that I don’t see put for often is: it’s not that GenZ is particularly poorly situated, it’s that the American Boomers, specifically, we’re the luckiest demographic cohort in world history, luckier than those who came before or after. Folks born in the 1850s might have technically been able to buy “a house” but it would not be a house that a modern person would find acceptable — plumbing, electricity, climate control, insulation, weatherproofing, fire safety, appliances and amenities, would all be abysmal by our standards. (The foundations of such houses were usually just a few big rocks.) In theory, a modern young person could build a house conforming to 1900 “average house” standards for cheap, but nobody wants such a house, and nobody builds them, and it is likely that such a house would be in violation of code. The richest person in 1900 couldn’t buy a house as good as the median US house.

So we’re in a place that our ancestors would view as a hypertech Jetsons future, except we got out over our skis for weird demographic and economic reasons and it’s currently very expensive to buy the latest hypertech magic house, and impossible to buy (what that ancestor or would call) a “normal” house.

This perspective — comparing more to any generation other than the Boomers provides a more realistic perspective.

An example of this is when collecting and summarizing anonymous feedback.

That struck me as well. I’m used to seeing intervals like 11 am to 6 pm, which tactically cuts out “breakfast” and likely reduces the average calories you manage to fit in over the day.

I feel like this is a good place to put the PSA that modern headache treatments have gotten really good, the CGRP drugs for migraine prevention are miraculous for some people and the triptans for migraine treatment can be a reliable stop-button for a migraine that would otherwise take you out for a day.

Full paper: https://arxiv.org/abs/2203.15556

Right, I don't know why people are so invested in minimizing how powerful this tool is. I have never properly learned Javascript but I can write Javascript using Copilot because copilot doesn't really make syntax errors, and I can recognize correct semantics, e.g. is it probably doing what I mean for it to do. And then I can, as you say, just test it.

It also bears repeating that Copilot will only get better.