• What is AI good for?
  
  A reader (I have readers?) wrote noting that my coverage of AI is almost entirely negative and wondering what AI is actually good for, presumably on the basis that private investors would not throw that many billions of dollars into something that didn't have at least some chance of making money, unlike, for example, the government.
  
  It's a good question.
  
  First we should probably note that there are two broad classes of AI being actively researched right now: Generative AI and Discriminative AI.
  
  Generative AI, driven by LLMs - large language models - is behind all the well-known AI instances worth untold billions of dollars. OpenAI's ChatGPT, Twitter's Grok, Anthropic's Claude, Google's Gemini, and Microsoft's Copilot; and open-source or nearly open-source solutions like Meta's LLaMA and Mistral's Mistral.
  
  The goal of generative AI is to ingest a huge amount of information in advance, and then, given a short and simple prompt, process that information in order to produce a response.
  
  Discriminative AI does the opposite. Given a data prompt of something in the real world - video, or sound, or an image - it uses a classifier to determine what it is examining. Is this apple ripe for the robotic apple-picking machine to pick it? Is it even an apple in the first place? What kind of spider is this that just bit me? Do I need to call an ambulance, or will it save me time to just lie down and die?
  
  It's no secret that Generative AI is getting all the attention. But is it worthy of that attention? The Verge asked that question yesterday and the answer turned out to be no.
  
  With Joe Biden's recent pardoning of his catspaw son Hunter, journalists were driven to defend him by digging up the little-known pardons of family members by former presidents, like George H. W. Bush's pardoning of his son Neil, or Woodrow Wilson's pardon for his brother-in-law.
  
  The problem is, these things never happened.

  Whatever happened in this case, there’s a running pattern of people relying on ChatGPT or other AI services to provide answers, only to get hallucinations in return. Perhaps you remember earlier this year when a trailer for Francis Ford Coppola’s Megalopolis was pulled because it contained fabricated quotes from critics. A generative AI, not identified, had made them up. In fact, ChatGPT is often "entirely wrong," according to the Columbia Journalism Review. Given 200 quotes and asked to identify the publisher that was the source of those quotes, ChatGPT was partially or entirely wrong more than three-quarters of the time.

  Journalists, being journalists, asked ChatGPT to do their research for them.
  
  ChatGPT, being ChatGPT, lied.
  
  LLMs are language models. They model language - well, sort of. They don't model the language itself, but construct an abstract model of the dataset fed into them.
  
  They don't understand facts. They don't actually have a notion of facts; nor do they have the contrary notion of falsehood. When they get information wrong, they are said to "hallucinate" rather than to have lied, because they have no basis for telling the difference between truth and falsehood.
  
  And that's intrinsic to the design of LLMs. Even before they enter "alignment" - a virtual lobotomisation that leaves AIs prone to crash when the wrong name is mentioned - they are fundamentally incapable of the kind of thought processes that most animals can do.
  
  This leaves us with sophisticated composite AIs like the virtual vtuber Neuro-sama, who can read every written language but is frequently unable to translate road signs, who has access to the sum total of human knowledge but insists that an anime figurine covered in glue is the perfect complement to your cookie recipe.
  
  Neuro is supposed to be like that, an impish hyperintelligent five-year-old, the perfect foil to her long-suffering father Vedal, because the main purpose there is entertainment. But you can't really expect to hand your job off to a five-year-old and not land with unexpected consequences.
  
  Or indeed entirely expected ones.
  
  So if it's useless at answering questions, what is AI good for?
  
  
  1. Image Generation
  
  If you use Grok on Twitter and ask it to generate an image of a Jaguar concept car, it will take a couple of seconds before producing something that would have any rational CEO looking to fire the entire design and advertising departments.
  
  
  
  Is it perfect?  If you look closely you'll see signs that the image generator has run into its bete noire, Euclid.  But I made no effort at all in selection here; I asked:
  
  generate an image of a jaguar concept convertible in british racing green
  
  And posted the first image that appeared.  And it took seconds.
  
  AI image generators have come a long way in a short time, mostly because they just have to look good, not produce a correct answer.  The tendency to produce human figures with hands attached at the elbows has been sharply reduced (though not yet banished entirely).  Now you more commonly see doors with hinges adjacent to the handles, or furniture that could only exist with access to Buckaroo Banzai's eighth dimension.
  
  Or cats.  Don't talk to me about AI cats.
  
  
  2. Software Testing
  
  If you write public-facing software, as I do daily, it's critical that the software be able to defend itself from both generic nonsense that is the core competency of the internet, and malicious nonsense that comes from a certain corner of the internet.
  
  When you've already tested all the known cases, there's a concept known as fuzzing that combines randomness and algorithms to generate horrible data to throw at your software to make sure that nothing falls apart in unexpected ways.  You are permitted to fail, but you are not permitted to break.
  
  Generative AI is perfect for fuzzing.  While it can't really understand your code, it can generate test patterns that reflect its analysis of your code and directly test potential flaws.  And it can do so nearly instantly, when writing an exhaustive test suite can take longer than writing the code in the first place.
  
  
  3. Discriminative AI
  
  Most of the flaws I listed arise from Generative AI.  Discriminative AI is much more useful, and consequently is much harder and receives much less attention and much less funding.
  
  
  And...  That's about it.  If you want mediocrity and are unconcerned with correctness, AI can fill you in with a poem or a song.  It's terrible at movies because it has the attention span of a frog in a blender, it's usually wrong but never uncertain, and it can't consistently count the number of letters in the word "the", but it is easy to use.
  
  
  

Tech News

• The "Hawk Tuah" girl launched a meme cryptocurrency.  It went exactly as you would expect.  (Web3 Is Going Great)
  
  Whether this was a rug-pull by the creators or by experienced investors, her individual followers lost all their money but her "advisors" made millions.
  
  
  
• Maxsun, best known for its anime-theme video cards, has announced three models based on Intel's new B580 chip.  (Tom's Hardware)
  
  Two of those are boring, but the third adds two M.2 slots to the video card.  Since these are low-end cards and only use eight lanes of PCIe but occupy a sixteen lane slot, it's fairly simple to hand the eight spare lanes off to two M.2 slots.
  
  If the CPU supports it, which AMD does but Intel sometimes doesn't.
  
  
  
• Need a UUID?  Here they are.  (Every UUID)
  
  All of them.
  
  
  
• The DC Circuit Court has declared the communists can go suck a lemon, handing a win to...  The other communists.  (CNN)
  
  TikTok filed a suit against the law requiring it to either sell or shut down.  It lost.
  
  TikTok is based in China and also banned in China, providing sufficient reason to wonder why any country should permit it to operate within its territory, without even pausing to consider the innumerable other scandals TikTok has been caught up in over the past week.
  
  
  
• Generation X is called "generation lead" by psychiatrists who have consider self-reliance and independence to be signs of severe mental illness.  (USA Today)

  "I tend to think of Generation X as 'generation lead,'" said Aaron Reuben, a study co-author and assistant professor of clinical neuropsychology at the University of Virginia. "We know they were exposed to it more and we're estimating they have gone on to have higher rates of internalizing conditions like anxiety, depression and symptoms of attention deficit hyperactivity disorder. So why the fuck aren't they all miserable like I am?"

  Too busy, Aaron.
  
  
  
• Orico has announced the MiniMate, a Thunderbolt storage device design to precisely match the new Mac Mini and add up to 8TB of SSD connected at 40Gbps.  (Notebook Check)
  
  Like the Mac Mini itself, it is not upgradeable in the slightest.  You can't add to or replace the storage, and it doesn't daisy-chain to connect a second device to the first.
  
  Thanks Orico.
  
  
  

Milet Video of the Day

Posted by: PatBuckman at Sunday, December 08 2024 03:51 AM (rcPLc)

Posted by: cxt217 at Tuesday, December 10 2024 12:52 PM (ZLF73)

Posted by: J Greely at Wednesday, December 11 2024 07:26 AM (oJgNG)

Posted by: PatBuckman at Wednesday, December 11 2024 09:47 AM (rcPLc)

Hide Comments | Add Comment