You need good system 1 to recognize when the LLM is wrong.

But your explanation makes sense, it also helps explain why you see so many post LLM responses they say are correct and proof the LLM can solve the problem, but then the thing they posted is bonkers and wrong. If those people lack a good system 1 it explains all of that, also helps explain which kind of person likes to work with LLMs.

You seem to be evaluating the LLM based on a single response rather than the whole "conversation." The user usually interacts with the LLM through 3-4 different responses to reach the right answer, which is valuable in itself. They're using both systems just as anyone would in a conversation.

I find LLMs useful for:

- Building bridges from familiar concepts to new ones.

- Checking my analysis and implementation for mistakes and gaps. This includes detecting subtle logic errors with static analysis.

- Condensing lengthy descriptions and complex conversations.

- Creating diagrams from verbal descriptions of flows.

- Finding design patterns to support my design, along with the basic structure that fits the chosen pattern.

- Writing unit tests and improving code coverage.

- Analyzing the credibility of information sources such as news stories and scientific studies.

- Generating original ideas and solutions to problems I may not have encountered before.

- Many more edge cases that help me turn an idea into a concrete concept in rapid time.

I have also used LLMs to entirely generate new tools and workflows, using languages I had barely touched before. This improved my knowledge of those languages and sped up my learning through practical examples.

Just as the printing press made calligraphy obsolete, LLMs will eventually make coding obsolete. Coding will be replaced by pseudo code and narrative that is independent of any framework or platform.

This does not mean that design and development will become obsolete, it will just become faster, without being hindered by the unnecessary barrier of coding.

Don't dismiss the value of this tool just because some marketers and regulators are using hype and fear to make money. LLMs can enhance your existing skill and make you more productive. They are not a crutch, they are a third leg.

> They are not a crutch, they are a third leg.

I don't think a third leg would make it easier to walk if you already have two legs. But it is a good way to see it, some would love a third leg, but I think until it gets better balanced most people will avoid it.

In Andrej Karpathy's intro video to llms ^1 he says this explicitly; one of the possible future innovations is to get llms to use 'system 2' thinking

[1] https://www.youtube.com/watch?v=zjkBMFhNj_g&t=2100

Cognitive psychology research shows the contrary: System 2 performance is hindered when System 1 is poor

Also see Recognition-Primed Decision Model: https://en.wikipedia.org/wiki/Recognition-primed_decision

Do you mind explaining why the next decade is going to be amazing ?

(removed)

can you post enough about your local LLM setup for me to google/rep. this?

There is so much out there for LLM's parsing is a pain.

For audio generation I recommend Bark. I am getting 14 seconds of audio that is about a third of eleven labs quality in about 2 minutes.

This is happening on a Windows 10 Dell, with 32gb of RAM, an i5, and an Nvidia 1050 GeForce with 4gb of vram.

I'm also able to decently run local LLMs because of llama.cpp and other libraries that can share models been ram and vram. There are other tools that can help with this as well including Ollama.

I suggest subscribing to r/localLLAMA. I also suggest using Bing Copilot in Edge with allowed access to the page you're viewing. I often use it to find new GitHub libraries and to give me first steps to be able to start using a new framework.