Lone Pine

FLI open letter: Pause giant AI experiments

that's only an excuse to keep people with consumer GPUs from getting LLMs,

Is this really the reason why?

Is law (AI lawyer) safety critical?

Replying toChatbot convinces Belgian to commit suicide

Chatbot convinces Belgian to commit suicide

I think we can resolve this manifold market question and possibly this one too.

Also, apologies for the morbid humor, but I can't help but laugh imagining someone being talked into suicide by the OG ELIZA.

Replying toWhy no major LLMs with memory?

Lone PineMar 28, 2023

Why no major LLMs with memory?

There is an architecture called RWKV which claims to have an 'infinite' context window (since it is similar to an RNN). It claims to be competitive with GPT-3. I have no idea whether this is worth taking seriously or not.

Replying toGPT-4 Plugs In

GPT-4 Plugs In

The entire conversation is over 60,000 characters according to wc. OpenAI's tool won't even let me compute the tokens if I paste more than 50k (?) characters, but when I deleted some of it, it gave me a value of >18,000 tokens.

I'm not sure if/when ChatGPT starts to forgot part of the chat history (drops out of the context window) but it still seemed to remember the first file after long, winding discussion.

Replying toGPT-4 Plugs In

GPT-4 Plugs In

I'm pretty confident that I have been using the "Plugins" model with a very long context window. I was copy-pasting entire 500-line source files and asking questions about it. I assume that I'm getting the 32k context window.

To be honest, this argument makes me even more confident in short times. I feel like the focus on scaling and data requirements completely miss the point. GPT-4 is already much smarter than I am in the ways that it is smart. Adding more scale and data might continue to make it better, but it doesn't need to be better in that way to become transformative. The problem is the limitations -- limited context window, no continual learning, text encoding issues, no feedback loop REPL wrapper creating agency, expensive to run, robotics is lagging. These are not problems that will take decades to solve, they will take years, if not months.

Gary Marcus's... (read more)

But when will my Saturn-branded car drive me to Taco Bell?

GPT-2005: A conversation with ChatGPT (featuring semi-functional Wolfram Alpha plugin!)

every 4 to 25 months

Is that a typo? That's such a broad range that the statistic is completely useless. Halving every 4 months is over 32 times as significant as halving every 25 months. That's completely different worlds.

-1

Why not constrain wetlabs instead of AI?

Yesterday I read this blog post which speculated on the possibility of training GPT-2 in the year 2005, using a supercomputer. For context, please read that post. Today I had the following conversation, on the same topic, with ChatGPT (GPT-4) using the new Wolfram Alpha plugin. This new plugin can (attempt to) do math and even (attempt to) look up statistics, such as the FLOP/s of the most powerful supercomputers from the TOP500 list.

However, don't get too excited, as ChatGPT doesn't seem to know how to use its new toy. It got an error message almost every time it tried to use the plugin.

The most startling thing I saw in this conversation... (read 6309 more words →)

Google's PaLM-E: An Embodied Multimodal Language Model

Most of the object-level stories about how misaligned AI goes wrong involve either nanotechnology or bio-risk or both. Certainly I can (and have, and will again) tell a story about AI x-risk that doesn't involve anything at the molecular level. A sufficient amount of (macroscale) robotics would be enough to end humanity. But the typical story that we hear, particularly from EY, involves specifically nanotechnology. So let me ask a Robin Hanson-style question: Why not try to constrain wetlabs instead of AI? By "wetlabs" I mean any capability involving DNA, molecular biology or nanotechnology.

Some arguments:

Governments around the world are already in the business of regulating all kinds of chemistry, such as the

... (read more)

Why doesn't my device automatically warn me when they have that daylight savings time nonsense? (I refuse to learn whatever is the proper sentence structure for talking about DST.) I just want DST to die, but failing that, it would be nice if my device was more interested in me making meetings on time than notifying me about a school shooting or whatever non-news my device shoves in my face. (Okay rant over)

Replying toGoogle's PaLM-E: An Embodied Multimodal Language Model

Lone Pine3y*

I honestly now believe that AGI already exists. This model may not have been it, and we will debate for hundreds of years* about whether the threshold was transformers or MLPs or multimodal, and which first model was really the first, in the same way we still debate which electronic computer was truly the first. But I do believe that it is here.

We do not have human-level machine intelligence (HLMI) yet. These systems still have a lot of limitations, in particular the context window and lack of memory. They are very limited in some domains such as robotics. However, it seems unlikely to me that we are not already in the takeoff.

* (assuming the debate doesn't get abruptly stopped)

Please Do Fight the Hypothetical

Scott Alexander and others in this part of the internet have been debating the Repugnant Conclusion recently. (I'll assume you're up-to-date on this discussion. If you aren't, follow the link, or just google "Repugnant Conclusion" or "Mere Addition Paradox".)

Many have noticed that there is a problem with the argument: the world-state that is described as 'repugnant' actually isn't so bad. Trillions of people just chilling -- living out their lives -- and not, on balance, suffering, is a much better world than the one we currently inhabit, where the rabbit runs for his life and the fox runs for his dinner. So if you force me to accept the hypothetical, then my... (read 879 more words →)

A Bias Against Altruism

What Diet Books Don't Teach: A book review and a request for more reading

I've noticed a phenomenon in our culture whereby altruistic actions face much more scrutiny than selfish actions. I think we should be aware of this bias effect, especially when discussing incentivizing people to e.g. work on AI alignment instead of ML research.

When I say 'our culture', I mean modern WEIRD culture, especially the English-speaking world. Here's what I notice: when I declare that I'm doing something selfishly and avowedly, I get praised. When I do something out of altruism, or do something that is coded as altruistic, my motives and true values get heavily scrutinized. The assumption is that I'm doing good in order to accrue praise and social status, which is... (read 470 more words →)

I'm happy to report that I'm now working on a project related to AI safety. The exact details will be announced sometime in the future. In totally unrelated news, I'm less optimistic about alignment than I was before (but still not pessimistic.) What was it that Mark Twain said about man's beliefs and paychecks and AI alignment?

[Test Post Please Ignore] Testing polling features

I’ve read a fair number of self-help books in my life. I’m long past the point where I expect a book to change my life, but they often are fun to pleasure-read. As Scott Alexander notes, these books all make claims they cannot fulfill. I don’t read so many diet books because they are the worst in the category. All diet books discuss what to eat but never how to get food. The underlying advice in all diet books, which is never elucidated beyond a sentence or two, is this: Cook for yourself, don’t eat at restaurants and don’t eat ‘processed’ food. All diet books divide food into good and bad, and... (read 1137 more words →)

Parable: The Bomb that doesn't Explode

I'm testing LW polling features. I haven't found a good guide online for the polling features. I tried googling, reading the FAQ and LW search. If anyone can point me in the right direction, I'd be very thankful.

[poll]{Option 1}{Option 2}{Option n}

[poll]{Red}{Green}{Blue}

[poll:Good..Bad]

[poll:Good...Bad]

[poll:Good....Bad] [poll:Good.Bad]

In 2003, a computer worm infected 90% of vulnerable hosts on the entire internet in less than 10 minutes. Wikipedia indentifies such fast spreading malware as a Warhol worm.

Forestalling Atmospheric Ignition

You're an engineer working for a military contractor. One day, your project manager comes to you and asks you to design a container for plastic explosive. It is to contain several kilograms of C4, enough to destroy a building. This is pretty dangerous, but you know that C4 is actually pretty safe. It can't be accidentally detonated by fire, impact or bullets. Only a detonator (another explosive) can trigger C4, so you don't include a detonator in your design, for safety.

Your PM reviews the design and gives you some feedback. "It would be a lot more useful if you put a blasting cap inside." You grimace. More useful, yes. But a lot... (read 347 more words →)

Conor Sullivan's Shortform

It's April, 1945 and you're a high ranking scientist in the Manhattan Project. You also happen to have the ear of President Truman, so your words and actions have a major impact on world outcomes. Everyone has been working hard on the bomb, determined to defeat the Nazis, except that the Nazis are presently being defeated and you're not nearly as concerned about the war in the Pacific. (You know Japan will lose eventually, regardless of the bomb.)

Another scientist points out to you a disturbing concept: It is possible that the forthcoming A-bomb test (scheduled for July) will cause the atmosphere itself to ignite in a self-sustaining fusion reaction that would consume... (read 153 more words →)

Two quick points about DOOM:

It is not necessary for the mad AI to kill every single human on the planet. The AI only needs to destroy our ability to respond. In order for human civilization to respond to such a threat we need access to technology, infrastructure and functional social structures (such as the military). The AI could defeat us without killing a single person, if it targets industry and infrastructure. (EY's melt-all-GPUs nanotech concept could easily be inverted: melt all rubber tires so no one can get to work, for example.)
It is entirely possible that we could see a full Kurzweilian singularity where everything is utopic, and then 20 years later

... (read more)