• Behavioural scientists behaving badly: The work of a top researcher at Harvard Business School has come into question after accusations that she is - and I quote* - just making shit up.  (New York Times)
  
  This is particularly striking because her area of expertise - such as it is - is honesty.
  
  Only half of published medial research holds up when other researchers try to reproduce the results, and for psychology the number is closer to one third.  (The Atlantic)
  
  Part of that is a form of selection bias: Research that doesn't find a result often doesn't get published in the first place.
  
  But it can't help if the effect was never real in the first place because the research wasn't real.
  
  * Quotes may settle in shipping.

Tech News

• AI's bigger-is-better approach is running out of road.  (The Economist)
  
  OpenAI's GPT-3 cost nearly $5 million to train in 2020.  GPT-4 just over two years later cost more than $100 million.  Is OpenAI prepared to spend $2 billion on GPT-5?  Even if they are, is there enough high-quality data that they can spend that much with it automatically going to waste?
  
  The article suggests that AI companies will be forced to work smarter, not expensiver.  But even if they do that will mean instead of spending exponentially more money for incrementally better results, they'll need to work exponentially smarter for incrementally better results.
  
  That's an even worse tradeoff.  It's the Technological Nothingularity, where even with AI helping train new generations of AI, progress slows to a crawl indistinguishable from a dead stop, where the technology of tomorrow can be safely predicted by assuming that nothing ever changes.
  
  
  
• ChatGPT can't program in INTERCAL.  (Muppet Labs)
  
  That's okay.  Neither can anyone else.
  
  
  
• Midjourney 5.2 is here and seems to be pretty good.  (Ars Technica)
  
  It may not matter if your progress stalls, so long as you get to good before stalling it.  If you run out of fuel after arriving safely at your destination. meh, you can deal with it later.
  
  It was hard to get good results out of Midjourney 2.  It was vague not only on how many fingers people had and where they should be located, but hands and heads as well.  They latest version appears to produce much more coherent images.

Posted by: Rick C at Monday, June 26 2023 01:33 AM (BMUHC)

Hide Comments | Add Comment