I have written a lot over the past three years that I did not publish. Now is a good time to list some of what remains after finding a middle ground for my use of language models, development assistants, and delegates12.
We sent our manufacturing to China, chip fabrication to Taiwan, and now we’re struggling to bring it back. Will we loose important intellectual faculties to language models?
Let me offer some simplistic framing: language models do text search not key-word search, sentences and paragraphs, not terms and links.
Is synthesizing text the same as synthesizing concepts?
Are language models capable of first principles thinking, something to that effect, or do they have a recency bias of a sort which restricts them to the present, the past, and incremental advancements?
Will offline training/learning, as opposed to online training/learning, prove to be a problem long term?
Will peripherals such a memory, text search, and code execution remedy the above?
What question of efficacy am I trying to answer anyway?
Ambiguity in natural language does not seem to be as big of an issue as we’d made out; Why is that? Does this speak hand in hand to commodity and stock code and incremental improvement through iteration?
How great is it really when people “build their own” CMS, CRM, or internal tool with the help of a language model? After all, they can do that with Notion or Zapier today or a CASE tool 30+ years ago, more systematically minus the chat interface. You still need a subscription, several in fact, and it’s not cheaper. It’s as if we’re really capitalizing on the habitual –if not addictive– precedent that Messenger and then Slack had set in our personal and professional lives.
On the nature of their output (and consequently their use), they are locally consistent, but globally perplexing. It is easy to be led astray because of this. The early demo of DALL-E tiling around the Girl with The Pearl Earring is a good example because it makes the problem obvious. Her surroundings morph from one room to another and the perspective is distorted. It is locally consistent (tile edges match up). It is globally perplexing (no such arrangement is physically possible). We have improved this since but my point is that cohesion is not the same as coherence. Cohesion is about pieces fitting together which is easier. Coherence is about properties of the whole which is harder.
Are the long-term deficiencies we see in language models deficiencies in intelligence more broadly speaking or simply deficiencies colored by our own perspectives and use?
How is a stroke of insight different from a creative leap?
Are they better referred to as modal text models than they are to as large language models? That is to characterize them by the properties of their output, not their internals/implementation, or indeed their input.
My response to “[language models] are only going to get better” is that cars have gotten better for over a century, and they continue to get better, but cars have not morphed into space craft. They will not take us to the stars but they did replace the horse-drawn carriage. In doing so they made some things better but brought their own problems.
Could GOFAI give us the AI we see in movies if we’d dedicated as much of today’s resources to it?
How is it that language models –not people– motivated integration/development APIs for such things as Web Servers and programming language runtimes, etc., via MCP?
Can we use language models to improve our state of affairs as opposed to to accelerating them?
How much of the code we use language models, development assistants, and delegates to write could we code away once and for all instead of again and again?
How can we complement and supplement our use of language models in the short term, with codification and abstraction in the medium term?
Reading, writing, comprehension, reasoning, understanding, creativity, sentience, and consciousness are confused a lot.
Is training and “Web use” more like trawling, not crawling the Web?
How are development assistants better than (any of) the juniors I have worked with?
Too often, reading and listening to online discourse, it feels like we are being tempted by evil of biblical scale.
Are we more likely to injure or kill each other through irresponsible deployment before we are replaced (courtesy P. Yiasemides)?