LLM Applications I Want To See

🌈 Abstract

The article discusses the author's perspective on the current state of large language models (LLMs) and their applications. The author argues that the AI community is overly focused on general-purpose performance benchmarks rather than exploring useful or creative applications of LLMs.

🙋 Q&A

[01] Critique of the current focus on general-purpose LLMs

1. What are the author's key criticisms of the current focus on general-purpose LLM performance?

• The author believes the AI community is overwhelmingly focused on general-purpose "performance" of LLMs at the expense of exploring useful or fun applications.
• The author notes that most of the "energy" in LLMs is on general-purpose models competing on benchmarks, rather than specialized applications.
• The author argues that the most interesting potential applications of LLMs go beyond simply automating tasks humans can already do, but rather enabling new capabilities that humans cannot do at comparable scale.

2. What examples does the author provide of more specialized or creative LLM applications?

• The author provides examples such as "predicts a writer's OCEAN personality scores based on a text sample" or "uses abliteration to produce a wholly uncensored chatbot that will indeed tell you how to make a pipe bomb."
• However, the author notes that these more specialized applications are quite rare compared to the focus on general-purpose models.

3. What are the author's concerns about certain LLM applications, such as chatbots for children or bots writing emails?

• The author expresses discomfort with the idea of little kids talking to chatbot toys or having a chatbot one can talk to in public, citing concerns about noise pollution.
• The author is also skeptical of the idea of bots writing emails, finding it creepy and undesirable.

[02] Proposals for novel LLM applications

1. What are the author's ideas for using LLMs to enable more customized content filtering?

• The author proposes a browser extension that allows users to train an LLM to identify and hide "unwanted" content based on the user's preferences.
• This would go beyond simple muting/blocking of specific people or keywords, and instead allow users to define a more nuanced "gestalt" of content they want to avoid.
• Users could also share their custom filters, potentially leading to popular community-curated filters.

2. What other novel LLM application ideas does the author suggest?

• Automatically generating plain language or lower reading level versions of text to aid those with limited literacy or cognitive disabilities.
• Using LLMs as a neutral "oracle" to help resolve interpersonal disputes, similar to how clergy historically played that role.
• Developing LLM-powered bots that can analyze group conversations and provide insights on conversational dynamics, potentially even moderating problematic behavior.

3. How does the author view the potential for LLMs to enable new types of digital interactions and experiences?

• The author believes LLMs can "restart" the diversity of possibilities for what can be done with computers, beyond just the basic "types of things you could do" that became consolidated in the 2000s and 2010s.
• The author suggests LLMs could enable novel applications, like using them to match dating profiles, that were technically possible before but didn't occur to people to try.