Daily Tech News 18 June 2023

• The Rot Economy: Big tech and the venture capital ecosystem that supports it have become a swirling cesspool of self-reinforcing mental illness. (Where's Your Ed At)
  
  Well, I mean, yes. Lurching from one fad to the next while squandering hundreds of billions of dollars.
  
  And it was only ten years ago that many of these companies actually did something, or at least tried to.
  
  
  
• Maybe they can both lose: Major subreddits are coming back to life now - with nothing but pictures of John Oliver. (The Verge)
  
  A timely reminder that while Reddit may be run by vapid radical left narcissists, so are most of the communities built on the platform.
  
  
  

• Another nail in OpenAI's coffin: Falcon LLM has been open-sourced. (TII)
  
  It comes in two versions: Falcon-7B which needs 16GB of RAM to run and is suitable for hobbyists, and Falcon-40B which needs 100GB of RAM and is aimed at academic researchers and startups. Both are available under the Apache open source license, making them free for individual and commercial use.
  
  Yes, these LLMs are still wildly overhyped, but they're not entirely useless, not when they're free and can run on a laptop - and when they're not intentionally crippled by ultra-woke marketing teams.
  
  The developer notes make it clear that Falcon was trained on wide set of data from the public internet, making it into a very well-informed digital schizophrenic.
  
  
  
• AI does not help programmers. (CACM)
  
  Or more specifically, it helps bad programmers become mediocre programmers, but it can't go beyond that because Large Language Models do not distinguish right answers from wrong.
  
  In the example provided, ChatGPT corrects a bug in the sample code, but introduces a new bug. When this is pointed out, ChatGPT fixes the specific case but not the general case.
  
  And when this is pointed out, ChatGPT coughs up what appears to be a hastily rewritten version of the author's own work, because the author turns out to be Betrand Meyer, creator of the Eiffel programming language and a leading expert in formal verification of computer programs.
  
  To be fair, most programmers aren't Betrand Meyer, but my opinion of the current state of AI tools in programming mirrors his.
  
  
  
• GPT-4 can pass MIT's Electrical Engineering and Computer Science curriculum with a perfect score. (HuggingFace)
  
  A remarkable result.
  
  
  
• No it fucking can't. (Notion)
  
  The second paper highlights two problems with the first.
  
  One, 4% of the problems in the test set cannot be solved with the information provided, or in some cases, at all:

  Below you are given the delays for the different gates you were permitted to use in part D above. Compute the propagation delay of your circuit from D.

  That's the entire question. There is no part D above, and yet the claim is that GPT-4 answered this question correctly. There are many questions like this in the test set - this second paper links to a spreadsheet with the full list of questions, good and bad.
  
  Two, the answers provided by GPT-4 are scored by GPT-4. If GPT-4 tells GPT-4 that GPT-4 got the question wrong, GPT-4 gets to try again, indefinitely.
  
  Supposedly the answers were verified manually, but if so, they did a pretty poor job because they missed all the wrong questions.
  
  Three - not included in the paper, but posted today on Twitter - the original code used to run the tests leaks the answers used for verification by GPT-4 to the GPT-4 instance answering the questions.
  
  Oops.
  
  
  
• Why EVs won't crash the electric grid. (Washington Post / MSN)
  
  Yes, EVs will require massive upgrades to the electric grid, but the grid has grown at that rate before.
  
  Well, it hasn't, but it almost did.
  
  Once. Fifty years ago.
  
  For a while.
  
  We're doomed.
  
  
  
• Putting a stick shift in an EV because... (The Verge)
  
  There is no because.

posted by Pixy Misa at 04:00 AM

| Access Comments