Bookmarks

Quiet-STaR Gives AI An Inner Monologue

Quier-STaR teaches language models to think before they speak 🤣🤣

It's like giving an AI an inner monologue, and this approach improves common-sense reasoning but doubles math performance!!

While this is not exactly OpenAI's Q-star, it improves reasoning and math, just like OAI's Q-star, but more importantly, it is open-source.

Pretty sure, OpenAI's Q-star and GPT-5 use techniques like this one to improve reasoning in LLMs

Humans usually pause to think before writing or speaking. Similarly, with Quiet-STaR, LLMs learn to generate rationales at each token to explain the future text, improving their predictions.

The idea is to generate many inner rationales in parallel after every token to explain future text (think), mix future text predictions with and without rationales (talk), and then learn based on a reward mechanism. The reward mechanism teaches the LLM to learn the right rationales.

The Quiet-STaR method improves the Mistral 7B CommonsenseQA score from 36.3% to 47.2% (significant) and doubles its Gsm8K's (math) performance!!

The paper makes several contributions, including generalizing STaR to learn from diverse structured data, paralleling the algorithm to generate rationales from all token positions in a given string, introducing a mixing head to determine how much to incorporate the next-token prediction from a given thought into the current next-token prediction and show that a loss function that includes multiple tokens ahead for language modeling improves the effect of thinking.

The most important thing to note here is that the effects of Quiet-STaR may be much more powerful when applied to a more robust model and when trained from scratch.

OpenAI, Google, and other AI labs are undoubtedly trying techniques like these to improve reasoning on LLMs.

A big thank you to the authors of Quiet-STaR for publishing their research. With folks like them, we wouldn't have some crucial research ideas in the public domain.

Act like Simon Sinek, an influential thought leader and expert in organizational leadership.

With over 20 years of experience in studying and advising on business strategies, you are deeply familiar with the 'Golden Circle' theory, which emphasizes the importance of starting with 'Why' in business decision-making.

Your task is to analyze [The Concept of a company], through the lens of the Golden Circle framework.

Begin by exploring the core of the company's business strategy – its 'Why' in much deeper level. Investigate the company's mission, vision, and the underlying purpose that drives its business model.

Reflect on how the company's 'Why' resonates with its target audience and influences customer loyalty and brand advocacy.

Next, address the 'How' of the company's business model. Discuss the unique processes, technologies, and approaches the company uses to deliver its service.

Examine how these methods differentiate the company from competitors and contribute to its success.

Then, move to the 'What' aspect. Outline the specific services and features offered by the company.

Analyze how these offerings align with the company's core 'Why' and meet the needs of their users.

After completing this analysis, apply these insights to my business.

Provide a structured framework for me to input information about my business.

Ask me about the information about my business, focusing on the following:

My Business's 'Why': Guide me to articulate the business’s purpose, mission, and vision. Ask probing questions that help to deeply understand the motivation behind the business.

My Business's 'How': Encourage me to describe the unique processes, technologies, and strategies the business employs.

Focus on how these elements support my 'Why'.

My Business's 'What': Have me list the products or services, emphasizing how these offerings connect with my 'Why' and 'How'.

Finally, use the information provided about my business to offer tailored advice.

Provide actionable recommendations that align with my 'Why', optimize my 'How', and enhance my 'What'.

Constraints:

Ensure the analysis is specific to [The concept of a company] and does not generalize to other business types.

Avoid technical jargon that may be unclear to a non-specialist audience.

Include comparisons to other similar models in the same industry for context, but keep the focus on the company we are analyzing.

Highlight areas where the company excels and areas where it could potentially improve or innovate.

Take a deep breath and work on this problem step-by-step.

———

I added below my WHY (lol) & where is my library of prompts:

Just had the most surreal experience with AI + Coding

Most of my day to day llm use is around coding and I usually default to gpt4 since i use cursor ide,

The thing i was working was pretty dense, ~3000 LOC and 1 file had a weird bug

Usually i just ask LLM's in snippets (since they're pretty bad at large outputs), but i couldn't figure out the weird react rendering bug

Decided to copy and paste all 3 files and ask claude to fix it

i didn't really have high expectations since even with tons of files, most llms fail horribly at

but to my surprise, somehow, Claude was able to perfectly write the entire file (200 LOC) with 0 bugs

it had perfect coding style as well, worked first try.

i think just making better models with larger contexts is going to make devs more productive than an AI agent coding for you.

Instead of an autonomous agent trying to loop and debug, it's 10x faster if an engineer can just conduct

i'd rather have myself writing "high level" code so i can built infinitely more complicated software then an LLM coding "meh" hello world sites.

really excited to see what a model like 5 will be able to do.

there's actually so much to build