• Re: AI: a new hobby

    From David Entwistle@qnivq.ragjvfgyr@ogvagrearg.pbz to rec.puzzles on Mon Jun 16 07:26:38 2025
    From Newsgroup: rec.puzzles

    On Sat, 31 May 2025 03:38:12 +0100, Richard Heathfield wrote:

    Firefox has decided to install ChatGPT on my system. At some point I
    will have to find out how inquisitive it is, but for now I've been
    setting it puzzles.

    I've been running through a knapsack problem, which I hadn't heard of
    before, with it. It came up with the following solution:

    Total Value:

    Diamond Necklace: -u800
    Ancient Vases: -u1200
    Silver Cups: -u800
    Gold Coins: -u250

    Total Value = 800 + 1200 + 800 + 250 = -u4250

    I'm not sure what to make of that, but I wouldn't want that AI driving vehicles just yet.
    --
    David Entwistle
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From David Entwistle@qnivq.ragjvfgyr@ogvagrearg.pbz to rec.puzzles on Mon Jun 16 10:33:50 2025
    From Newsgroup: rec.puzzles

    On Mon, 16 Jun 2025 07:26:38 -0000 (UTC), David Entwistle wrote:

    I'm not sure what to make of that, but I wouldn't want that AI driving vehicles just yet.

    I suspect it has been reading sci.maths.
    --
    David Entwistle
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From Richard Heathfield@rjh@cpax.org.uk to rec.puzzles on Mon Jun 16 16:48:40 2025
    From Newsgroup: rec.puzzles

    On 16/06/2025 08:26, David Entwistle wrote:
    On Sat, 31 May 2025 03:38:12 +0100, Richard Heathfield wrote:

    Firefox has decided to install ChatGPT on my system. At some point I
    will have to find out how inquisitive it is, but for now I've been
    setting it puzzles.

    I've been running through a knapsack problem, which I hadn't heard of
    before, with it. It came up with the following solution:

    Total Value:

    Diamond Necklace: -u800
    Ancient Vases: -u1200
    Silver Cups: -u800
    Gold Coins: -u250

    Total Value = 800 + 1200 + 800 + 250 = -u4250

    I'm not sure what to make of that, but I wouldn't want that AI driving vehicles just yet.

    Yes, the three Rs are not its longest suits. It seems to do
    rather better at history and geography (and turning ASCII into
    UTF-8 after being specifically told not to).
    --
    Richard Heathfield
    Email: rjh at cpax dot org dot uk
    "Usenet is a strange place" - dmr 29 July 1999
    Sig line 4 vacant - apply within

    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From Carl G.@carlgnews@microprizes.com to rec.puzzles on Mon Jun 16 09:12:13 2025
    From Newsgroup: rec.puzzles

    On 6/16/2025 12:26 AM, David Entwistle wrote:
    On Sat, 31 May 2025 03:38:12 +0100, Richard Heathfield wrote:

    Firefox has decided to install ChatGPT on my system. At some point I
    will have to find out how inquisitive it is, but for now I've been
    setting it puzzles.

    I've been running through a knapsack problem, which I hadn't heard of
    before, with it. It came up with the following solution:

    Total Value:

    Diamond Necklace: -u800
    Ancient Vases: -u1200
    Silver Cups: -u800
    Gold Coins: -u250

    Total Value = 800 + 1200 + 800 + 250 = -u4250

    I'm not sure what to make of that, but I wouldn't want that AI driving vehicles just yet.

    The AI has figured out sales tax.
    --
    Carl G.


    --
    This email has been checked for viruses by AVG antivirus software.
    www.avg.com
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From David Entwistle@qnivq.ragjvfgyr@ogvagrearg.pbz to rec.puzzles on Mon Jun 16 18:29:53 2025
    From Newsgroup: rec.puzzles

    On Mon, 16 Jun 2025 16:48:40 +0100, Richard Heathfield wrote:

    Yes, the three Rs are not its longest suits. It seems to do rather
    better at history and geography (and turning ASCII into UTF-8 after
    being specifically told not to).

    Although I know little about the subject of Artificial Intelligence, I'd
    have though the basics of arithmetic and the physical laws would be
    embedded in to any system, in an immutable way, before it began training
    on other, more questionable, material.

    I would hope so, at least.
    --
    David Entwistle
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From Richard Heathfield@rjh@cpax.org.uk to rec.puzzles on Mon Jun 16 20:10:16 2025
    From Newsgroup: rec.puzzles

    On 16/06/2025 19:29, David Entwistle wrote:
    On Mon, 16 Jun 2025 16:48:40 +0100, Richard Heathfield wrote:

    Yes, the three Rs are not its longest suits. It seems to do rather
    better at history and geography (and turning ASCII into UTF-8 after
    being specifically told not to).

    Although I know little about the subject of Artificial Intelligence, I'd
    have though the basics of arithmetic and the physical laws would be
    embedded in to any system, in an immutable way, before it began training
    on other, more questionable, material.

    I would hope so, at least.

    All hope abandon!

    Go to brainbashers.com, open the puzzle of the day, and note the
    URL, which contains a date. You can hack it and go back about a year.

    Many of the puzzles can be copy-pasted directly into ChatGPT. It
    generally catches on pretty quick to what it's supposed to do,
    and *sometimes* it gets it very right very fast, but often it
    gets its knickers in a twist, and it's frankly rather
    embarrassing when it tries to count the letters in a word, and
    /fails/.
    --
    Richard Heathfield
    Email: rjh at cpax dot org dot uk
    "Usenet is a strange place" - dmr 29 July 1999
    Sig line 4 vacant - apply within

    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From David Entwistle@qnivq.ragjvfgyr@ogvagrearg.pbz to rec.puzzles on Tue Jun 17 08:06:50 2025
    From Newsgroup: rec.puzzles

    On Mon, 16 Jun 2025 20:10:16 +0100, Richard Heathfield wrote:


    All hope abandon!

    Go to brainbashers.com, open the puzzle of the day, and note the URL,
    which contains a date. You can hack it and go back about a year.

    Many of the puzzles can be copy-pasted directly into ChatGPT. It
    generally catches on pretty quick to what it's supposed to do,
    and *sometimes* it gets it very right very fast, but often it gets its knickers in a twist, and it's frankly rather embarrassing when it tries
    to count the letters in a word, and /fails/.

    Chat GPT 4.0 mini did appear to have the concept of a knapsack problem. I asked it to write a problem and it initially suggested I had broken into a jewellery shop and had the option, amongst other things, to steal a watch weighing 4kg...

    A later puzzle iteration was a bit more practical, which I'll post later.
    Chat GPT couldn't solve it and when shown options for small improvements,
    had forgotten those improvements when next asked for an optimal solution.
    By that time it had forgotten how to add up.
    --
    David Entwistle
    --- Synchronet 3.21a-Linux NewsLink 1.2
  • From richard@richard@cogsci.ed.ac.uk (Richard Tobin) to rec.puzzles on Tue Jun 17 10:51:41 2025
    From Newsgroup: rec.puzzles

    In article <102pnr1$1q3dn$1@dont-email.me>,
    David Entwistle <qnivq.ragjvfgyr@ogvagrearg.pbz> wrote:

    Although I know little about the subject of Artificial Intelligence, I'd >have though the basics of arithmetic and the physical laws would be
    embedded in to any system, in an immutable way, before it began training
    on other, more questionable, material.

    No. These are large *language* models. An LLM can only do arithmetic
    to the extent that generating sentences similar to the ones it was
    trained on happens to give correct answers. ("Similar" is doing a lot
    of work there.)

    It is possible to add specific skills such as arithmetic to an
    LLM-based system.

    -- Richard
    --- Synchronet 3.21a-Linux NewsLink 1.2