so close

newsletter6.org

@@ -81,7 +81,7 @@ Unsurprisingly, no, his affiliate marketing site did not take off and make a bil
 
 I'm pretty sure this trend is going to die out the way "buy this NFT and soon you can sell it for infinite money" hype did. Maybe FoldingIdeas will even make a video on it.
 
-The trend I'm less confident is going to die out is "automate content creation with chatGPT". It's pretty big right now, from really low-effort entries like "faceless videos" on YouTube (...find link) to making automated blog posts---often for your affiliate marketing---people are arguing that soon you'll be able to automate away all that pesky writing you need to do in your life.
+The trend I'm less confident is going to die out is "automate content creation with chatGPT". It's pretty big right now, from really low-effort entries like "faceless videos" on YouTube (https://www.youtube.com/results?search_query=faceless+video+with+ai) to making automated blog posts---often for your affiliate marketing---people are arguing that soon you'll be able to automate away all that pesky writing you need to do in your life.
 
 If I can get a little polemical, I find most of this sorta fascinating because, well, I'm not sure if I understand the point in writing things that are simplistic enough you can automate the process.
 
@@ -93,9 +93,33 @@ So, yes, you can automate out huge chunks of writing if the writing was going to
 
 To make Wittgenstein wince via non-linear time: of that which one must automate, one must remain silent.
 
-And yet far from being silent we're seeing people literally bragging about using chatGPT to create the literary equivalent of shovelware:
- + [LINKS]
+And yet far from being silent we're seeing people literally bragging about using chatGPT to create the literary equivalent of shovelware: https://www.businessinsider.com/chatgpt-author-time-boucher-midjourney-ai-novels-books-2023-5
+
+Part of the danger here is that we're going to create an environment in which the web is further diluted in terms of real, careful, informative content. This is going to degrade search even more than it already has been and, I fear, we're going to start having the same companies providing the tools for flooding the internet will continue selling us the tools to fix the problems they're exacerbating.
+
+It's a bit of an aside but I think in a year or two we're going to see google and bing become essentially useless. We're going to need to use a mixture of personal bookmarking, webrings, rss feeds, and the like in order to actually manage our information rather than being able to retrieve it via search. The new internet is going to become the old internet, is what I'm saying.
+
+So all that---somewhat negative---news aside what are some of the good things that have been coming out of the LLM world?
+
+Well since my last newsletter gpt4 has become available as part of chatGPT (the original being a fine-tuned version of gpt3.5). I have to say in my own experiments gpt4 is incredibly impressive. It kinda wipes the floor with gpt3.5 in a lot of ways. This isn't just my impression, either. It does really well on a lot of various benchmarks and metrics. So much so that it's ignited a bit of a hype train: https://arxiv.org/abs/2303.12712. Now despite the fact that I don't think it's "agi" we're starting to see some kind of interesting emergent properties, like the fact that prompting tricks like the ones outlined in this video and the papers he links even work https://www.youtube.com/watch?v=wVzuvf9D9BU. If you don't have time to watch this I can summarize and say that we're discovering that a lot of tricks like "explain step-by-step", "reflect on your answer", "find what's wrong in the provided answer" &c. actually provide massively improved results when prompting. We don't really understand why this is true. Some people have feel like there's new abilities that show up in LLMs as they get larger, but it's actually hard to tell if that's real or if it's an accident of how we're evaluating the model's capabilities as can be seen in this paper: https://arxiv.org/abs/2304.15004.
+
+
+The big OpenAI news is not just that gpt4 is available but also that they've started adding extensions: gpt4 + plugins and gpt4 + browsing. The tl;dr of the former is that developers are starting to create integrations between the natural language interface to chatGPT with their products. They range from creating spotify playlists out of natural language prompts to searching for real estate to integration with pdf parsing tools so that you can load in pdfs as context for your queries. There are literally hundreds and hundreds of plugins for gpt4 now and it's only been available for development for a couple of months.
+
+You won't be surprised that I have a lot of concerns about the idea of creating an interface to your application with a natural language interface like an LLM. The flexibility that makes the LLM an easy tool for an interface also makes it hard to control. Simon Wilison has written a lot about this:
+ + https://simonwillison.net/2023/May/19/
+ + https://simonwillison.net/2023/May/15/
+ + https://simonwillison.net/2023/May/11/delimiters-wont-save-you/
+ + https://simonwillison.net/2022/Sep/16/prompt-injection-solutions/
+ + &c.
+
+But on the other hand I think things like the ability to query your own documents is pretty great. I guess you can summarize my feeling as "LLMs are great when they're used to process text, not to take actions based on that text". So while, yes, there's an instacart plugin that can let you ask about a recipe and then build a cart of all the ingredients for that recipe I think trying to do anything more complicated than that is---wait for it---a recipe for disaster.
+
+This does bring us to things like gpt4 + browsing, where with chatGPT you're now able to ask queries and also have it double check the query against a bing search. I'll be honest I can't tell if directly providing URLs to use in search actually affects how the bing search is done and the information is processed. It's fairly opaque to see what it's doing, which I find frustrating: two steps forward, one step back.
+
+Although I'm seguing to our last topic major topic which is that if you want to have more control and transparency over what's happening you need to have more local code and control rather than using an opaque platform such as chatGPT. That's not quite possible yet, but we're getting there.
+
+The big thing that happened in the "local LLM install" revolution was the release of the llama model weights
 
-Part of the danger here is that we're going to create an environment in which the web is further diluted in terms of real, careful, informative content. This is going to degrade search even more than it already has been and, I fear,