GenAI Diary

Last update: Wednesday 1/1/25

Diary of a computer savvy generative AI blogger

This page records the editor's experience as a computer savvy user of generative AI technology who only interacts with LLMs via chatbots, e.g., ChatGPT, Claude, Gemini, etc. He uses genAI whenever he thinks it will enable him to increase his productivity by doing things better and/or faster and/or with less effort. These notes reflect his successes and failures.

All of his notes are dated, and will be modified or deleted if/when his subsequent experience suggests that the issues he identified have been improved or fully resolved. The dates of his newest notes will be displayed in a "red" font that will fade to "black" whenever a newer topic is addressed.

A. Apple | B. ChatGPT | C. Claude |  D. Gemini | E. Perplexity F. Copilot

The most recent topics ➡ 1)  What to do if speech-to-text stops working, 2) Apple Intelligence improves speech-to-text transcriptions, and 3) Search, i.e., which chatbots can or cannot search and how effectively.

A. Apple

• Speech-to-text (1/1/25) ... What to do if it stops working.
  

  Apple Intelligence has the potential to greatly expand the primary functions of iPhones from hubs for communication, information, and entertainment to also becoming handheld workstations. As handheld workstations, one of their most important functions will be the composition of notes, memos, and reports via dictation. Dictation requires reliable and highly accurate speech-to-text transcription.. 

  
  

  Unfortunately, in the few days since the editor called his readers’ attention to their sudden acquisition of high accuracy, he discovered that iPhones’ speech-to-text sometimes stops working … unexpectedly.

  
  

  When he queried ChatGPT as to why this happened and how to prevent it from happening, the chatbot suggested that Apple Intelligence tries to run everything on the iPhone, but may encounter insufficient available memory from time to time. It therefore suggested a couple of remedies: closing all other apps that were running in the background and/or restarting the iPhone. Sometimes these remedies work, but sometimes they don’t. 

  
  

  If these remedies are unsuccessful, the chatbot suggested using ChatGPT itself as a workspace for dictating rough first drafts. Given that ChatGPT does all of its processing in a cloud, it never encounters memory shortage problems.

  
  

  Here’s how to use ChatGPT as a workspace for dictating rough first drafts:

  1. Open the ChatGPT app on your iPhone.
     
     
  2. Click the small microphone icon on the right side of the prompt box. (Do NOT click the larger icon in the lower right hand corner of the app window because that microphone sends your audio to Apple Intelligence first.)
     
     
  3. The first line you dictate should say something like “Do not process the following text. Just save it.”
     
     
  4. Now dictate your rough first draft. You may want to dictate a few words, phrases, or sentences at a time. If you pause, click the checkmark icon on the right side of the prompt box to convert your audio so far into text. If you want to add more dictation, click the small microphone in the prompt box again.
     
     
  5. When your dictation is done, click the up arrow on the right side of the prompt box to save your dictated rough draft.
     
     
  6. Copy the draft to the clipboard; then paste it into Notes, Word, or whatever application you will use to revise your draft into a final version. 
  
  
• Speech-to-text (12/30/24) ... A sudden dramatic improvement occurred today on iPhones.
  In iOS 18.1 and 18.2, Apple's speech-to-text conversions were error prone; they were way, way, way below the accuracy of speech-to-text conversions by ChatGPT and other apps. But suddenly, Apple's speech-to-text on iPhones in Messenger, Notes, and other Apple programs is almost as good as ChatGPT's. 
  
  In the past, if you clicked the microphone icon in the lower right hand corner of your iPhone, your speech audio went to Apple Intelligence. But apps like ChatGPT and Claude positioned a second smaller microphone icon in the side of their text input boxes. Clicking that icon sent your speech audio directly to OpenAI or Anthropic's cloud, which rendered far more accurate speech-to-text conversions.
  
  But today, when you click the microphone icon in the lower right corner of your window, a second smaller microphone appears that floats near the text inout box. This floating icon disappears when you stop speaking.
  
  1/1/25 … On 12/30/24 the editor incorrectly conjectured that ChatGPT was now handling speech-to-text for Apple Intelligence. But a more recent diary note reports that Apple’s speech-to-text sometimes stops working due to insufficient available memory, something that is highly unlikely to happen to ChatGPT's cloud based processing.

BackToTop

B. ChatGPT (OpenAI)

• Search (12/29/24) ... ChatGPT can search for (almost) anything, anywhere.
  -- Given prompt "Neoskeptics blog podcasts" ➡ it provided a brief description of this blog, plus summaries or three of its recent podcasts
  
  -- Given prompt "Cade Metz, NY Times, quantum computing, Google, 2024 ... cite link to Google announcement" ➡ it provided a concise, but multi-paragraph summary of Cade's extensive multi-page piece about Willow in the NY Times.
  page 
  
  -- Given prompt "VentureBeat agents 2024" ➡ It provided concise summaries of five recent articles about agents in VentureBeat.
  
  -- Given prompt "Best vegetarian restaurants in Bethesda MD" ➡ It provided five plausible recommendations, plus a map that showed where each restaurant was located. ... On the evening of 12/29/24 the editor examined the menus of the "vegetarian" restaurants more closely. They are not "vegetarian"; all of them are regular high quality restaurants that have some vegetarian meals on their menus.

BackToTop

C. Claude (Anthropic)

• Search (12/29/24) ... Claude cannot search
  -- Given prompt "Neoskeptics blog podcasts" ➡ Claude replied "I cannot perform web searches or access external URLs directly."

BackToTop

D. Gemini (Google)

• Search  (12/29/24) ... Gemini grossly misinterpreted the prompt 
  -- Given prompt "Neoskeptics blog podcasts" ➡ Gemini returned a lengthy response that began "It seems you're interested in the intersection of blogs, podcasts, and neoskeoticism" ... Its lengthy hallucination continued with brief discussions of what neoskepticism was, noted key figures in this political philosophy, recommended related blogs and podcasts, and concluded with referrals to additional resources ... :-)

BackToTop

E. Perplexity

• Search  (12/29/24) ... Perplexity can only find famous things.
  -- Given prompt "Neoskeptics blog podcasts" ➡ Perplexity responded "There is no specific information about a podcast or blog titled “Neoskeptics” in the provided search results. If you are looking for a particular podcast or blog with that name, it may not be widely indexed or available in these sources. You might try searching directly on platforms like Listen Notes, Podsearch, or YouTube for more precise results." ... Alas, this blog is not widely indexed, i.e., It is not famous ... :-(

BackToTop

F. Copilot (Microsoft)

• Search  (12/29/24)  ... Maybe Copilot does not have a ChatGPT Plus subscription yet.
  -- Given prompt "Neoskeptics blog podcasts" ➡ Copilot responded "You can check out the [Neoskeptics Blog] for their latest posts and podcasts. They tackle a variety of topics with a scientific skepticism lens. It's an insightful read if you're into critical thinking and debunking myths!" ... Flattering, but it doesn't describe out podcasts.
  
  The link to this blog works, sort of. It takes you to an intermediary site that provides a link to the blog. This is classic "Old School" search. Provide prmpt; get links or links to links. But ChatGPT and Perplexity are proposing a new paradigm: Ask quesstions (or prompts); get answers that are supported by links. Why isn't Copilot using this technology? Surely Microsoft's $11 billion investment in OpenAI provides access to it.

BackToTop