1. Introduction

1.1 Summary of key claims

Even without further breakthroughs in AI, language models will have big impacts in the coming years, as people start sorting out proper applications
- The early important applications will be automation of expert advisors, management, and perhaps software development
- The more transformative but harder prizes are automation of research and automation of executive capacity
In their most straightforward form (“foundation models”), language models are a technology which naturally scales to something in the vicinity of human-level (because it's about emulating human outputs), not one that naturally shoots way past human-level performance
- i.e. it is a mistake-in-principle to imagine projecting out the GPT-2—GPT-3—GPT-4 capability trend into the far-superhuman range
- Although they’re likely to be augmented by things which accelerate progress, this still increases the likelihood of a relatively slow takeoff — several years (rather than weeks or [...]

---

Outline:

(00:12) 1.1 Summary of key claims

(03:02) 1.2 Meta

(04:25) 2. What type of thing are language models?

(04:29) 2.1 Emulating civilization, not individual people

(06:45) 2.2 An extremely crude picture of how language models work

(08:12) 2.3 What are foundation models approximating?

(13:13) 3. Techniques for getting value from language models

(13:37) 3.1 Prompt engineering

(14:26) 3.2 Scaffolding

(15:13) 3.3 Finetuning

(16:04) 3.4 Combining these

(16:56) 4. Natural limits of language models

(17:29) 4.1 Approximating human capabilities, not superhuman capabilities

(19:29) 4.2 Limited cognition per forward pass

(22:41) 4.3 Missing cognitive moves?

(23:34) Multimodal models

(24:31) 5. Early major impacts of language models?

(24:36) 5.1 Principles for thinking about this

(28:13) 5.2 Important early areas for automation

(31:46) 5.3 The big two applications

(35:20) 6. Timelines and takeoffs

(35:24) 6.1 How quickly is all of this likely to happen?

(38:02) 6.2 Scaling language models towards superintelligence

(38:33) 6.2.1 Finetuning for superhuman task performance

(40:09) 6.2.2 Scaffolding for amplification via reflection

(41:13) 6.3 Recursive improvement and takeoff

(41:29) 6.3.1 Reflection-based takeoff

(45:03) 6.3.2 Scaffolding-based takeoff

(46:28) 6.3.3 Finetuning-based takeoff

(46:46) 6.3.4 Mixed takeoffs

(47:40) 6.3.5 Systems not built on language models

(48:43) 7. Language model agents and transparency

(48:47) 7.1 Where does agency come from?

(51:58) 7.2 Strategy: avoid selection pressure for agency

(54:03) 7.3 The bitter lesson?

(56:50) 8. Risks and strategies

(56:54) 8.1 A rough taxonomy of risks

(01:02:18) 8.2 Example strategies for mitigating the different risks

(01:04:45) 8.3 Thoughts on tactical implications

---

First published:
December 20th, 2023

Source:
https://www.lesswrong.com/posts/pnaZajFFWuXAnetpz/on-the-future-of-language-models

---

Narrated by TYPE III AUDIO.

“On the future of language models” by owencb

LessWrong (30+ Karma)

“On the future of language models” by owencb