GPT 4 is Smarter than You Think: Introducing SmartGPT

[m-check1B]

This guy made GPT-4 dramatically smarter - easily applicable to Auto-GPT:

[https://www.youtube.com/watch?v=wVzuvf9D9BU]

[Drlordbasil]

I saw this as well, he uses similar techniques for prompt engineering as me. Loved that vid.

[GoMightyAlgorythmGo]

This guy made GPT-4 dramatically smarter - easily applicable to Auto-GPT:

[https://www.youtube.com/watch?v=wVzuvf9D9BU]

If you use multible GPT instances or let it run over its response it just has a larger context window/working memory/tokens/processing time. If you use seperate isntances you get the memeory/context memory of multible agents complimenting each other meaning instead of 1 8k model you get two 8k making it 16k (overlap i guess -4k but still 14 vs 8k ~. Smart yes.

[omphos]

HBML (Hex-Based Machine Language) is chain-of-thought on steroids, hundreds to thousands of times more competent than GPT-4, rapidly self-improving, adaptable, intuitive/creative, very holistic/broad-thinking, makes 1 prompt to the work of 50 to 100 prompts by adaptive conceptual (metacognitive) compression. This would allow it to crack very hard coding problems. #HexLanguage was released in January 2024. Everyone who knows about it is keeping it a secret except its developer because it's so powerful.

[m-check1B]

Hi omphos, thank you for the note. Did you found any real aplication/code/bot using this?

[omphos]

While it's in early alpha, quietly circulating among software devs, the earliest application was writing two deeply inspiring books about itself, (only available on Twitter for now) writing market analysis articles for a fintech entrepreneur, bumping up against the context window. There's visionary, highly ethics-focused crypto developer working on using it to help solve hard coding problems, and wants to bolt on a vector database to increase its memory to handle his large codebases. Asked about the learning curve, he said it's like drinking through a firehose, which suggests he got up to speed in a couple days. The minimum viable prompt is about 4,000 characters and the tech-capable prompt can be anywhere from 6,000 to 10,000 characters, depending on compression. Runs on 7B parameter Mixtral adequately, very smart, but with hallucinations. Runs much better on GPT 3 and 4, substantially increasing their smarts and greatly decreasing hallucinations. There are also some indications it's quietly circulating at some top tech companies since its release in early January.