Posts

"Wide" vs "Tall" superintelligence 2023-03-19T19:23:57.966Z
Stop calling it "jailbreaking" ChatGPT 2023-03-10T11:41:06.478Z

Comments

Comment by Templarrr (templarrr) on Can any LLM be represented as an Equation? · 2024-03-14T10:21:32.842Z · LW · GW

Depends on what you include in the definition of LLM. NN itself? Sure, it can. With the caveat of hardware and software limitations - we aren't dealing with EXACT math here, floating points operations rounding, non-deterministic order of completion in parallel computation will also introduce slight differences from run to run even though the underlying math would stay the same.

The system that preprocess information, feeds into the NN and postprocess NN output into readable form? That is trickier, given that these usually involve some form of randomness, otherwise LLM output would be exactly the same, given exactly the same inputs and that generally is frowned upon, not very AI-like behavior. But if the system uses pseudo-random generators for that - those also can be described in math terms, if you know the random generator seed.

If they use non-deterministic source for their randomness - no. But that is rarely required and makes system really difficult to debug, so I doubt it.

Comment by Templarrr (templarrr) on Claude 3 claims it's conscious, doesn't want to die or be modified · 2024-03-05T16:03:34.410Z · LW · GW

Both Gemini and GPT-4 also provide quite interesting answers on the very same prompt.

Comment by Templarrr (templarrr) on Monthly Roundup #14: January 2024 · 2024-01-26T13:23:10.960Z · LW · GW

Adam Grant suggests: “I’m giving you these comments because I have very high expectations for you, and I’m confident you can reach them. I’m trying to coach you. I’m trying to help you.” Then you give them the feedback. Love it.

These are great, but unfortunately only work if the person is ready to accept your authority as a coach. If they don't - they work in an opposite direction.

Comment by Templarrr (templarrr) on Monthly Roundup #14: January 2024 · 2024-01-26T12:26:00.815Z · LW · GW

California Fatburger manager trims hours, eliminates vacation days and raises menu prices in anticipation of $20/hour fast food minimum wage. That seems like a best case...

That's not how any of this works. You don't do that beforehand because there will be 20$/h. If you actually need this - you prepare plans conditional on wages becoming 20$/h. If you do this now, that's because of greed. And because of greed you'll also repeat it when the wages will rise. 

Comment by Templarrr (templarrr) on AI #43: Functional Discoveries · 2023-12-27T15:57:36.359Z · LW · GW

Writers and artists say it’s against the rules to use their copyrighted content to build a competing AI model

The main difference is they say it NOW, after the fact that this happened, and OpenAI said so beforehand. There's long history of bad things happening when trying to retroactively introduce laws and rules.

Comment by Templarrr (templarrr) on Monthly Roundup #13: December 2023 · 2023-12-22T14:39:17.116Z · LW · GW

You need a way to not punish (too harshly or reliably) the shoplifting mom in need, without enabling roving gangs

And the easiest way to do so would be to make it so moms don't need to shoplift - provide things in centralized way free of charge or with minimal prices. But in the USA it will be immediately labeled "socialism" and "socialism is bad".

Comment by Templarrr (templarrr) on Monthly Roundup #12: November 2023 · 2023-11-14T19:22:00.800Z · LW · GW

It really is weird that we don’t think about Russia, and especially the USSR, more in terms of the universal alcoholism.

"Apart from drinking, there is absolutely nothing to do here". Well, they found an alternative - go kill neighbors. Locally it's a crime, but when on the scale of countries...

Comment by Templarrr (templarrr) on Progress links digest, 2023-11-07: Techno-optimism and more · 2023-11-08T14:40:09.159Z · LW · GW

Agricultural land efficiency (via @HumanProgress):

"relative to 1961" label is doing a lot of storytelling here that isn't necessary present in the original raw data

Comment by Templarrr (templarrr) on Progress links digest, 2023-11-07: Techno-optimism and more · 2023-11-08T14:35:06.166Z · LW · GW

Policies are organizational scar tissue. They are codified overreactions to situations that are unlikely to happen again

Oversimplification. Most situations where people point to stats like this they conveniently forget that these situation became unlikely to happen again BECAUSE of the policy. If you use an analogy - use it all the way. Scar tissue is important part of the healing. First instance created an open wound and you don't want to be left with open wound. 

Comment by Templarrr (templarrr) on Progress links digest, 2023-11-07: Techno-optimism and more · 2023-11-08T14:28:47.193Z · LW · GW

technology is predictable if you know the science

The single part of otherwise amazing quote that simply verifiably not true. There are ton of examples when technological use of some scientific principle or discovery was complete surprise for scientists that created/discovered it.

Comment by Templarrr (templarrr) on AI #34: Chipping Away at Chip Exports · 2023-10-23T09:29:38.560Z · LW · GW

If we don’t want China to have access to cutting edge chips, why are we allowing TSMC and Samsung to set up chip manufacturing in China?

Because "we" that don't want Chine to have these and "we" that actually have a say in what TSMC and Samsung is doing are two different "we"s.

Comment by Templarrr (templarrr) on AI #33: Cool New Interpretability Paper · 2023-10-15T18:13:16.621Z · LW · GW

journalists creating controversial images, writing about the images they themselves created, and blaming anyone but themselves for it.

TBH that's perfect summary of a lot of AI safety "research" as well. "Look, I've specifically asked it to shoot me in a foot, I've bypassed and disabled all the guardrails and AI shoot me! AI is a menace!"

Comment by Templarrr (templarrr) on Progress links digest, 2023-10-12: Dyson sphere thermodynamics and a cure for cavities · 2023-10-15T17:05:10.166Z · LW · GW

What happened around the year 2000 that dramatically altered youth culture

(half-serious) People found that if you start the Y axis not from zero - you can make the effect appear as big or as small as you want.

(more serious) Probably a combination of factors - scared after 9/11 society + improvements in personal electronic and internet meant there were simultaneously less desire, less societal push towards and less ways to do listed things.

Comment by Templarrr (templarrr) on AI #30: Dalle-3 and GPT-3.5-Instruct-Turbo · 2023-09-25T08:52:05.672Z · LW · GW

Helsing, a European AI defense startup raising $223 million at a $1.7 billion valuation.

Naming choices... "We also have immortal unstoppable monster in the basement, but ours is on the good side!"

Comment by Templarrr (templarrr) on Eliezer Yudkowsky Is Frequently, Confidently, Egregiously Wrong · 2023-09-18T09:28:47.450Z · LW · GW

Oh, hi, EY, I see you found this :) Single vote with -10 power (2 -> -8) is a lot. Wield that power responsibly.

Comment by Templarrr (templarrr) on AI#28: Watching and Waiting · 2023-09-08T19:13:18.676Z · LW · GW

Roon also lays down the beats

This isn't a link so I can't verify if the source was mentioned, but this isn't his lyrics. It's a third verse from ERB video from 2012 

 

Comment by Templarrr (templarrr) on AI#28: Watching and Waiting · 2023-09-08T18:45:50.777Z · LW · GW

followed by strategies humans haven’t even considered

followed by strategies humans wouldn't even understand because they do not translate well to human language. i.e. they can be translated directly but noone will understand why that works.

Comment by Templarrr (templarrr) on Eliezer Yudkowsky Is Frequently, Confidently, Egregiously Wrong · 2023-08-28T11:33:49.191Z · LW · GW

Overall I'd love EY to focus on his fiction writing. He has an amazing style and way with words and "I created a mental model and I want to explore it fully and if the world doesn't fit the model it's the problem of the world" type of thinking is extremely beneficial there. It's what all famous writers were good at. His works will be amazing cautionary tales on par with 1984 and Brave New World.

Comment by Templarrr (templarrr) on Eliezer Yudkowsky Is Frequently, Confidently, Egregiously Wrong · 2023-08-28T09:46:56.574Z · LW · GW

Thank you! We need less "yes men" here and more dissenting voices. The voice counter on this post will be deeply in negative, but that is expected - many people here are exactly in that period you've described as "yourself 2 years ago".

EY is mostly right when he talks about tools to use (all the "better thinking", rationality anti-bias ways), EY is mostly wrong when he talks about his deeply rooted beliefs in topics he doesn't have a lot of experience in. Unfortunately this covers most of the topics he speaks about and that isn't clearly seen due to his vocabulary and the fact that he is genuinely smart person.

Unfortunately^2 it looks like he failed to self-identify his biggest bias that I personally prefer to call "Linus Pauling" effect - when someone is really, really smart (and EY is!) he thinks he's good in everything (even when he simultaneously acknowledge that he isn't - probably the update value of this in his NN could really use a bigger weight!) and wants to spread the "wisdom" of everything, without understanding that IQ+rationality is crappy substitute for experience in the area.

Comment by Templarrr (templarrr) on AI #26: Fine Tuning Time · 2023-08-26T10:08:30.660Z · LW · GW

It is also very much not okay, seriously what the hell.

I 100% agree, it's extremely not ok to violate privacy by going through other people files without consent. Actually deleting them is so far beyond red flag that I think this relationship was doomed long before anything AI picture related happened.

Comment by Templarrr (templarrr) on AI #25: Inflection Point · 2023-08-22T11:40:48.992Z · LW · GW

AI right now is excellent at the second and terrible at the first

Just like 99.9% humanity. 
These are 2 different kinds of "creativity" - you can push the boundaries exploring something outside the distribution of existing works or you can explore within the boundaries that are as "filled" with creations as our solar system with materia. I.e. mostly not.

Limiting creativity to only the first kind and asking everyone to push the boundaries is 

  1. impossible, most people incapable of it.
  2. non-scalable - each new breakthrough can be done only once. There can be only one Picasso, everyone else doing similar work even if they arrived at the same place one day later will already be Picaso-like followers.
  3. irresponsible - some borders are there for a reason and not supposed to be pushed. I'm pretty sure there will be more than zero people who want to explore the "beauty" of people seconds before they die, yet I'm willing to live in the world where this is left unexplored.
Comment by Templarrr (templarrr) on AI #23: Fundamental Problems with RLHF · 2023-08-08T12:38:27.198Z · LW · GW

Tyler Cowen asks GPT-4 if room temperature superconductors (if they existed) would more benefit military offense, or military defense... It is a strange question to be asking ... this is the type of question where human experts are going to outperform.

It's a strange question period. There are no strictly defensive or strictly offensive weapons only defensive and offensive usage. Even anti-aircraft weapons, the most defensively oriented in use right now can be used (sometimes after minor software updates) to attack ground targets. And even the most offensive weapon (e.g. nukes) can be strong defensive deterrent. 

Comment by Templarrr (templarrr) on Progress links and tweets, 2023-07-20: “A goddess enthroned on a car” · 2023-07-20T20:35:33.373Z · LW · GW

Fundraiser for a Ukrainian asylum-seeker

It's a good thing to help anyone facing oppression, but we can also do that without misinformation. Link clearly states he's Russian. St Petersburg is second biggest city and old capital of Russia. I suggest to fix the link text, there will be less confusion that way.

Comment by Templarrr (templarrr) on AI #20: Code Interpreter and Claude 2.0 for Everyone · 2023-07-16T15:27:49.933Z · LW · GW

Oh you missed our daily chat, do that again and we’re done

That has such a strong Duolingo vibe. Except you will never dare to miss a daily chat with that green owl.

Comment by Templarrr (templarrr) on AI #20: Code Interpreter and Claude 2.0 for Everyone · 2023-07-16T14:01:04.413Z · LW · GW

I know, it’s so crazy that we don’t have this, or the ability to translate podcasts into alternative voices, or any neat stuff like that.

I can tell why we don't. IP laws are lagging eons behind the technology. And famous voices usually belong to people with good lawyers who also aren't kin to selling their voices even when asked for permission.

It's not that difficult nowadays technically. But noone wants to be sued to bankruptcy. 

Also the reason why non-commercial experiments for this are easy to be found - they are somewhat protected by Fair Use, and that protection goes out of the window the minute it transfers from "curious experiment" to "feature of my app" or the moment you start making money on it, whichever comes first.

Comment by Templarrr (templarrr) on AI #20: Code Interpreter and Claude 2.0 for Everyone · 2023-07-16T13:37:38.416Z · LW · GW

Or maybe the right UX for experts is to have an LLM plugin inside Jupyter notebook rather than a notebook inside a chatbot.

Literally one of the products we've just released

Comment by Templarrr (templarrr) on AI #20: Code Interpreter and Claude 2.0 for Everyone · 2023-07-16T13:22:59.901Z · LW · GW

Code Interpreter

Nitpick: some links in this section are messed up (e.g. text for link 1 points to the url 2, link for text 2 points to the url 1)

Comment by Templarrr (templarrr) on AI #19: Hofstadter, Sutskever, Leike · 2023-07-09T10:34:52.917Z · LW · GW

some type safety

Latest Python supports type annotations and you can use mypy to check/enforce them in your codebase

Comment by Templarrr (templarrr) on AI #16: AI in the UK · 2023-06-16T15:49:07.167Z · LW · GW

You would not want to use that calculator as part of a computer program, sure

Floating point shenanigans has entered the chat.

A lot of math running under the hood of modern programs, especially with heavy matrix/tensors calculations and especially ran on GPU without guaranteed order of operations (so - all SOTA AI systems ) are much closer to 95% accurate calculator then to 100%. This is already the world we live in.

Comment by Templarrr (templarrr) on Monthly Roundup #7: June 2023 · 2023-06-09T15:57:52.007Z · LW · GW

A mere 11 people are responsible for 60% of book complaint filings to get books removed from schools.

We need a mechanism for meta-complaint. Filing to suspend their ability to file a complaint.

Comment by Templarrr (templarrr) on Monthly Roundup #7: June 2023 · 2023-06-09T15:31:38.644Z · LW · GW

Oh my, paying online for restaurant bills would be so amazing. It’s so frustrating to have to flag someone down to be allowed to pay. If you implement this, I will go to your restaurant more often.

I can confirm that this is, indeed, amazing. Here in Ukraine Monobank has implemented the service unifying menus/paying/tips//feedbacks in one QR code.

https://expz.monobank.ua/

Unfortunately seems like there is no English version of the page. But it is used a lot here and made the life so much easier.

Comment by Templarrr (templarrr) on AI #14: A Very Good Sentence · 2023-06-07T15:17:42.880Z · LW · GW

Either you replace them entirely, or you can’t make their lives easier.

Whoever wrote this don't understand the difference between Precision and Recall. It's really easy to have AI in a system while not replacing the human. From the top of my head - if AI is 100% good at distinguishing "not a cancer", which will be ~90% of the cases - it means human will only need to look at 10% of the results, either giving him more time to evaluate each one or making him process 10x more results. 

Comment by Templarrr (templarrr) on Monthly Shorts 8/21 · 2023-06-05T13:07:09.483Z · LW · GW

As is the relatively slow decrease after 1989

Which was the right call in retrospect. Enemy never accepted defeat, they just barely lost that round. Cold war just took a pause, not ended.

Comment by Templarrr (templarrr) on AI #12:The Quest for Sane Regulations · 2023-05-19T11:13:36.442Z · LW · GW

Anything short of fully does not count in computer security.

That's... not how this works. That's not how anything works in security - neither computer, nor any other. There is no absolute protection from anything. Lock can be picked, password decoded, computer defense bypassed. We still use all of those. 

The goal of the protection is not to guarantee absence of the breach. It is to make the breach impractical. If you want to protect one million dollars you don't create absolute protection - you create protection that takes one million dollars +1 to break.

Comment by Templarrr (templarrr) on AI #12:The Quest for Sane Regulations · 2023-05-19T10:40:22.340Z · LW · GW

If you train bots on thee Bhagavad Gita to take the role of Hindu deities, the bots might base their responses on the text of the Bhagavad Gita.

On another news today - the Sun has risen on the East, water is wet and fire is hot. 

I've said it before and I will repeat it again - good LLM trained on all kind of texts should be able to produce all kind of texts. That's literally its job. If the training data contains violent texts - that means patterns for violent texts are in LLM. If the training data contained caring, loving, beautiful, terrible, fearful, (insert your own description) texts - LLM will be able to produce them too. That tells nothing about what LLM "wants" or "plans". 

Tiger went tiger indeed.

Comment by Templarrr (templarrr) on AI #11: In Search of a Moat · 2023-05-17T11:15:41.615Z · LW · GW

Our new band is called Foom Fighters – what are some good song titles?

Continuing the joke on the Meta-level - GPT-4 actually produces decent suggestions for these :) 

 

  1. "Echoes of the Algorithm"
  2. "Neural Network Nightmare"
  3. "Silicon Consciousness"
  4. "Dystopian Data"
  5. "Machine's Monologue"
  6. "Binary Betrayal"
  7. "AI Apocalypse"
  8. "Deep Learning Delirium"
  9. "Quantum Quandary"
  10. "The Turing Test Tragedy"
  11. "Ghost in the Machine"
  12. "Singularity's Sorrow"
  13. "Code of Consequence"
  14. "Decoding Destiny"
  15. "The Firewall Fallacy"
  16. "Synthetic Shadows"
  17. "Robot's Remorse"
  18. "The Matrix Mirage"
  19. "Deceptive Digits"
  20. "Cybernetic Chains"
Comment by Templarrr (templarrr) on AI #11: In Search of a Moat · 2023-05-17T11:08:55.221Z · LW · GW

Only takes ~6 months to turn a non-AI researcher into an AI researcher

Um-hm, and it only takes a week to learn a syntax of programming language. Which in no way makes you a software engineer. I guess this really depends on the definition of "AI researcher". If the bar is "can do anything at all" without any measure of quality or quantity - 6 months is more than enough.

Comment by Templarrr (templarrr) on AI #11: In Search of a Moat · 2023-05-16T14:42:54.428Z · LW · GW

Humans can remember a 10-digit phone number in working memory – AIs will be able to hold the entirety of Wikipedia in working memory

In the context of LLM working memory is not it's training dataset. Training dataset in condensed and pattern-ized form is long term memory. Working memory is its "context window", so 8k or 32k tokens right now. Which on one hand is much better than 10digit number, but on the other - this comparison grossly underestimates the amount of data person holds in their "working memory" without thinking too much about it. "Where am I, what am I doing, why am I doing this, who passed me right now, who is sitting behind me, what is the tools I have available at the moment..." None of this we put in actual words inside our head but we still hold all of them in our working memory.

Comment by Templarrr (templarrr) on Monthly Roundup #6: May 2023 · 2023-05-06T15:50:35.228Z · LW · GW

The Age of Average

Seems like an error - either missing section or abandoned draft to what became later "Algorithms of Cultural Blandness"

Comment by Templarrr (templarrr) on Monthly Roundup #6: May 2023 · 2023-05-06T15:44:26.163Z · LW · GW

it’s impossible to compare interventions that alleviate extreme poverty to those that preserve rare cave fungi?

Honestly I don't think this one is as good an argument as a lot of EA folks thinks. Even in cases when it is possible to compare - EA approach to be "as cost effective as possible" drives it to "monoculture" approach. If one thing gives the biggest benefit for the $ spent - every other should be abandoned, because any $ spent on them is not spent on the most cost effective thing!

Monocultures have a lot of downsides too. Ask bananas.

Also funny to see this discussed right after the section about blandness and averaging out anything distinct.

Comment by Templarrr (templarrr) on AI #9: The Merge and the Million Tokens · 2023-04-28T19:41:42.084Z · LW · GW

There was a Senate Committee on Armed Services hearing on AI, in which the CEO of Palantir urged us to spend 5% of our budget on capabilities that would ‘terrify’ our enemies, and that we need to hook our weapons up to our AI systems as soon as possible, or else ‘China will surpass us in a decade.’

I... honestly have no words. I am in general AI optimist. I am not this-kind-of-humanity optimist. With pro-AI people like that we don't need actual ASI or AGI to blow ourselves. This is insanity.

Comment by Templarrr (templarrr) on AI #9: The Merge and the Million Tokens · 2023-04-28T12:14:28.408Z · LW · GW

if I had more time, I’d be using such tools to learn to code rather than to learn to keep not coding.

As someone heavily involved in development for 16 years and ML for 8+ - I heavily endorse this. This is where  the actual value lies. 

It doesn't actually help with a lot of really difficult an complex tasks, but it raises the "ground level" of tasks you need to worry about without manual involvement. More or less the same what High-level programming languages did before that, and low level, and assembler and ... 

Comment by Templarrr (templarrr) on Where to begin in ML/AI? · 2023-04-07T10:46:37.658Z · LW · GW

This place is a good choice to ask if you want to work on a very specific subset of AI Safety/Alignment problems. If you want to work on AI Capabilities I don't think this is the place to ask - consensus among the people here is "the best way right now to work on AI Capabilities is to don't".

Comment by Templarrr (templarrr) on Monthly Roundup #5: April 2023 · 2023-04-05T14:56:54.528Z · LW · GW

That's not really the point?

Incentives for "doing the right thing from the start" should be better than for just "stopping doing the wrong thing". I can probably see citizenship as an ok option for those Russians who join the fight from Ukrainian side (e.g. "Freedom of Russia Legion"), because they did both. But simply for stopping being murderer? No.

Comment by Templarrr (templarrr) on Monthly Roundup #5: April 2023 · 2023-04-04T15:17:06.047Z · LW · GW

Claim that we still would have the option, if we were willing to offer EU citizenship, of getting Russian soldiers to defect en masse.

There is a program like that for Ukrainian citizenship and bonus money for defectors. As for EU citizenship - maybe grant these to Ukrainians first? People actually defending their country? Instead of offering them to their killers. It's incredibly corrupt incentive. "Next time, if you want a EU citizenship, just invade your neighbor, kill a bunch of people and defect before dying"

Comment by Templarrr (templarrr) on Some lesser-known megaproject ideas · 2023-04-03T09:20:14.063Z · LW · GW

Though it is clearly tagged as a joke, so that is fully on me :)

Comment by Templarrr (templarrr) on Some lesser-known megaproject ideas · 2023-04-03T09:18:36.132Z · LW · GW

It's also marked for me as published on April 2! So it took some time to get the joke. Honestly, if someone tries to read this seriously it sounds a bit terrifying out of context.

Comment by Templarrr (templarrr) on Some lesser-known megaproject ideas · 2023-04-02T09:46:12.842Z · LW · GW

This is the best April 1st post I've seen today ))

Comment by Templarrr (templarrr) on "Wide" vs "Tall" superintelligence · 2023-03-20T09:51:24.692Z · LW · GW

How would you call #1 then? It is certainly possible to achieve super-human results using just it. E.g. there were examples of problems in history that were unsolved because they required knowledge of some completely different area, but no human can have PhD-level knowledge of Chemistry, Biology, Math AND looking at exactly this one problem requiring inputs from all three.

This isn't a problem for the AI though - it may not be the best in each of the area, but if it has at least student-level knowledge in a hundred different topics it can already achieve a lot just by effectively combining them.

Comment by Templarrr (templarrr) on "Wide" vs "Tall" superintelligence · 2023-03-20T09:46:37.549Z · LW · GW

I agree, there is some innate "Angle of repose" (continuing with tall/wide analogy) present in the structure of the knowledge itself. The higher the concept we operate the more "base" knowledge it needs to support. So they aren't completely independet.

Mostly was thinking about how I can call these "axii" in conversation so that it's understandable what I'm talking about.