Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Rodney Brooks on GPT-4

spectrum.ieee.org

Post reply on HN
Print view View original

171–180 of 412 posts

Re: Rodney Brooks on GPT-4

#171

Post by SanderNL »

famouswaffles wrote: post #67
astrange wrote: post #52

Earlier quoted context omitted.

Do you know what a "world model" is? It's a thing people were assumed to have in 1970s psychology, but was never well-defined enough to tell if it exists or not, so I don't think it's obvious anything else has one. https://twitter.com/Meaningness/status/1639120720088408065 I think "common sense" or "long term memory" might be more productive things to say.

The brain does a lot of hierarchical prediction with sense data. When new information comes in, it references "something" and adjusts the sense data likewise. That's why when you shift your eyes quickly, you see blurred images pass by. In reality, you should be seeing complete black because the brain doesn't actually process visual information that shifts so quickly. But your brain "knows" it should see...well someth…

It always seemed sensible to me that we are looking at a dreamlike construction all the time, not just holes filled in but literally everything you see is made up, but grounded in sensory data (when not psychotic or dreaming). This also nicely explains the clarity and realism of lucid dreaming, because it’s just doing whatever you are doing during the day but without (external) grounding sensory data.

Re: Rodney Brooks on GPT-4

#172

Post by zainhoda »

thomastjeffery wrote: post #107
fwlr wrote: post #33

> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…

> My current (tentative) resolution of the surprise is that language encoded way more information about reality than we thought it did. I think you are close to the mark, but you have been subtly mislead: language is not the data we are working with. We are working with text . Once you fix that particular failure of word choice, everything else becomes much more clear: text contains much more information than languag…

There’s an exercise that some people do when learning programming, which is to write down the steps to make a sandwich. Then the teacher follows the exact instructions to make a sandwich and most people don’t put enough detail for a computer to follow (I.e. open the fridge etc) and the teacher will run around bumping into things. That used to be a teaching exercise to show people the amount of precision required when telling a machine what to do.

Now with LLMs, I think one of the great leaps is the idea that it’s no longer necessary to be “pedantic” when giving computers instructions because LLMs have somehow learned to fill in the blanks with a similar shared “understanding” of the world that we have (I.e. cheese is stored in the fridge so you have to go open the fridge to fetch the cheese for the sandwich).

Re: Rodney Brooks on GPT-4

#173

Post by seanhunter »

usaar333 wrote: post #111
seanhunter wrote: post #87

Earlier quoted context omitted.

Most of the time when people find a maths problem that they can trick the model into getting wrong, it's also possible to get the model to give the correct answer with better prompting. A trick that's worth knowing is just to ask the model to give each step in the solution and explain as it goes. This gives the model "time to think" and leads to better results.

Pretty sure you can't get GPT-4 to do 8 digit multiplication with any prompt. For what it's worth, I'm not even sure if chain of thought provides much value to GPT-4. The RLHF it went through seems to have encouraged more logical thinking already.

The deeplearning.com prompt engineering course contains an example where GPT4 solves a problem with chain of thought and fails without. The prompt is

    Determine if the student's solution is correct or not.

    Question:
    I'm building a solar power installation and I need 
     help working out the financials. 
    - Land costs $100 / square foot
    - I can buy solar panels for $250 / square foot
    - I negotiated a contract for maintenance that will cost  
    me a flat $100k per year, and an additional $10 / square 
    foot
    What is the total cost for the first year of operations 
    as a function of the number of square feet.

    Student's Solution:
    Let x be the size of the installation in square feet.
    Costs:
    1. Land cost: 100x
    2. Solar panel cost: 250x
    3. Maintenance cost: 100,000 + 100x
    Total cost: 100x + 250x + 100,000 + 100x = 450x + 100,000
GPT-4 will say the student is correct even though the answer includes an error (they should be multiplying by 10x not 100x). If you change the prompt to include chain of thought as follows it notices the error

    Your task is to determine if the student's solution 
    is correct or not.
    To solve the problem do the following:
    - First, work out your own solution to the problem. 
    - Then compare your solution to the student's solution \
    and evaluate if the student's solution is correct or not. 
    Don't decide if the student's solution is correct until 
    you have done the problem yourself.

    Use the following format:
    Question:
    ```
    question here
    ```
    Student's solution:
    ```
    student's solution here
    ```
    Actual solution:
    ```
    steps to work out the solution and your solution here
    ```
    Is the student's solution the same as actual solution \
    just calculated:
    ```
    yes or no
    ```
    Student grade:
    ```
    correct or incorrect
    ```

    Question:
    ```
    I'm building a solar power installation and I need help \
    working out the financials. 
    - Land costs $100 / square foot
    - I can buy solar panels for $250 / square foot
    - I negotiated a contract for maintenance that will cost \
    me a flat $100k per year, and an additional $10 / square \
    foot
    What is the total cost for the first year of operations \
    as a function of the number of square feet.
    ``` 
    Student's solution:
    ```
    Let x be the size of the installation in square feet.
    Costs:
    1. Land cost: 100x
    2. Solar panel cost: 250x
    3. Maintenance cost: 100,000 + 100x
    Total cost: 100x + 250x + 100,000 + 100x = 450x + 100,000
    ```
    Actual solution:
As I say this same prompt "fools" GPT3.5 and 4.

Re: Rodney Brooks on GPT-4

#174

Post by groestl »

_ea1k wrote: post #78
ramraj07 wrote: post #47

Earlier quoted context omitted.

The issue with all these experts is they still think it's human nature to be able to fully understand the world before they speak about it. On the contrary it's human nature (and all animal nature) to figure out how to navigate the world without fully understanding or having a complete model of it. All you need is a working model that affects the facets of the world you need to deal with. I still remember in the 90s…

I remember a friend being confused by downloading. He didn't realize that meant we'd have a copy and not that the source would lose it in the process. It seemed bizarre at the time, but, tbh, I didn't have _that_ much better of a model of how the whole process worked.

> we'd have a copy and not that the source would lose it in the process

I don't known when this happened, but maybe that means the Anti-Piracy campaigns worked on him, confusing illegal replication with theft.

Re: Rodney Brooks on GPT-4

#175

Post by thorum »

m3kw9 wrote: post #101
p-e-w wrote: post #64

> It gives an answer with complete confidence, and I sort of believe it. And half the time, it’s completely wrong. That's bullshit, unless you are asking questions specifically designed to make GPT-4 hallucinate. For most real-world, everyday topics, the accuracy is close to 100%. GPT-4 would be utterly useless otherwise.

Yeah a lot of times is right, unless they are really complex subjects then maybe even less then 50

Less about complexity than about how well-documented the subject is on the internet IMO. I’ve been using it to help me set up and troubleshoot AWS Elastic Kubernetes clusters, which are plenty complex, and I’d estimate it’s been around 95% accurate. (And for at least one of the times it seemed to be wrong, it turned out I’d made a mistake following its instructions...)

Re: Rodney Brooks on GPT-4

#176

Post by j16sdiz »

blowski wrote: post #163
thomastjeffery wrote: post #107

Earlier quoted context omitted.

> My current (tentative) resolution of the surprise is that language encoded way more information about reality than we thought it did. I think you are close to the mark, but you have been subtly mislead: language is not the data we are working with. We are working with text . Once you fix that particular failure of word choice, everything else becomes much more clear: text contains much more information than languag…

I don't really understand your distinction between language and text, but it sounds intrguing. Would you be able to give more detail? I searched but couldn't find anything that seemed to explain it.

Spoken language have pitch, stress, mood, etc.. Written text contain some, but not all.

Text, on the other hand, can present in list or table, with varies formatting, indentation. You can't reproduce them in spoken text.

Re: Rodney Brooks on GPT-4

#177

Post by notShabu »

fwlr wrote: post #33

> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…

Has there been any research in the possibility that much of what we say/speak (including this convo) is actually just stochastic parroting?

e.g. instead of the stochastic parrots mimicking intelligence maybe intelligence doesn't exist, it's just stochastic parrots of various levels of sophistications organized into a hierarchy. "Intelligence" is necessarily socially defined with the more complex parrots being unpredictable and "intelligent" from the POV of lower parrots. Vicer versa, looking down, the lower parrots seem to act like "NPCs"

Re: Rodney Brooks on GPT-4

#178

Post by tornato7 »

thomastjeffery wrote: post #107
fwlr wrote: post #33

> The large language models are a little surprising. I’ll give you that. I think this is the key point about LLMs that kind of explains the wide and polarized views on whether it understands or parrots, whether it can think or is the precursor to thinking or is a dead-end, whether it will catastrophically destroy the world, or “merely” make it steadily worse with bullshit, or just put a few industries out of a job. A…

> My current (tentative) resolution of the surprise is that language encoded way more information about reality than we thought it did. I think you are close to the mark, but you have been subtly mislead: language is not the data we are working with. We are working with text . Once you fix that particular failure of word choice, everything else becomes much more clear: text contains much more information than languag…

One experiment that I would love to see is an LLM-like model for audio. Feed it hours and hours of lectures, sound effects, animal calls, music etc. You would be able to talk to it and it would ingest the raw waveform then produce audio as a response. Would it learn the fundamentals of music theory? Would it learn to produce "the sound of a bowling ball hitting a dozen windchimes?" Would it learn to talk in English and communicate with whales? We've already done text and images, now someone please do sound!

Re: Rodney Brooks on GPT-4

#179

Post by onethought »

akiselev wrote: post #161
yblu wrote: post #149

Earlier quoted context omitted.

What I found disheartening was many of those scientists, especially those on the "nothing to worry about" camp, seemed not to entertain the thought that they could be wrong, considering the scale of the matter, i.e. human extinction. If there's a chance AI poses an existential threat to us, even if it is 0.00000001% (I made that up), should they be at least a bit more humble? This is uncharted domain and I find it in…

Meh. Add it to the pile. The number of world ending risks that we could be worried about at this point are piling up and AI exterminating us is far from the top concern, especially when AI may be critical to solving many of the other problems that are. Wrong about nuclear proliferation and MAD game theory? Human extinction. Wrong about plasticizers and other endocrine disruptors, leading to a Children of Men scenario…

Yes, and all of those (including AI) are not even human extinction events.

- Nuclear war: Northern Hemisphere is pretty fucked. But life goes one elsewhere.

- Plasticisers: We have enough science to pretty much do what we like with fertility these days. So it's catastrophic but not extinction.

- Climate Change: Life gets hard, but we can build livable habitats in space... pretty sure we can manage a harsh earth climate. Not extinction.

- Deadly virus: Wouldn't be the first time, and we're still here.

- Astroid impact: Again, ALL human life globally? Some how birds survived the meteor that killed the dinosaurs, I'm sure we'd find a way.

- Complete Made up evil AI: Well we'd torch the sky, be turned into batteries but then be freed by Keanu Reeves.. or a Time traveling John Connor. (sounds like I'm being ridiculous, but ask a stupid question...)

Re: Rodney Brooks on GPT-4

#180

Post by galaxyLogic »

> it answers with such confidence any question I ask. It gives an answer with complete confidence, and I sort of believe it. And half the time, it’s completely wrong.

That's it. LLMs don't have SHAME! They simply don't care if what they're saying is false or not. They are like some politicians of late.

They don't even understand that giving out a wrong or misleading answer will affect their credibility. You see LLMs don't have a DESIRE for credibility. They don't have desires.

We need not fear these models. But we do need to fear some people who will use them for evil purposes.

Post reply on HN