Episode Overview

Podcast: Azeem Azhar's Exponential View
Episode: The Science of Making Truthful AI
Date: February 7, 2024
Host: Azeem Azhar
Guest: Richard Socher, AI researcher, entrepreneur, founder/CEO of You.com, former Chief Scientist at Salesforce

Theme:
Azeem Azhar and Richard Socher explore the nature of intelligence, the historical roots and current frontiers of artificial intelligence, the challenges of building truthful AI, and the future of AI architectures and applications. Through candid discussion, the episode cuts through AI hype, addresses the realities and misconceptions of large language models (LLMs), and unpacks the implications for the future of technology, industry, and society.

Key Discussion Points & Insights

1. What is Intelligence?

• Historical Context: Socher discusses the roots of modern AI, referencing influential work on neural nets for speech recognition and the watershed moment of the ImageNet dataset, which enabled significant breakthroughs in computer vision.
  • "There's one event that happened even before ImageNet and that was George Dahl and Geoff Hinton actually working on speech recognition and neural nets... Speech recognition actually is best done with a neural network." (03:53, Socher)
• Multifaceted Intelligence: Intelligence is not just one thing—it includes motor, visual, language, and higher-level reasoning capabilities. Language and the ability to transmit knowledge across generations is highlighted as a unique human feature.
  • "Language enables you to have collective intelligence, too, and historical intelligence and memory." (06:42, Socher)
• Engineering Challenge: From an engineering perspective, the challenge is that intelligence is not well-defined, making it difficult to build.
  • "It's kind of a disaster from an engineering perspective though, isn't it? ... There's any amount of terrible enterprise software that has been built from that kind of specification." (09:39, Azhar)

2. From Replicating Human Skills to Building New Capabilities

• Beyond Human Intelligence: AI need not strictly mimic humans; it can exceed in areas humans did not evolve for, like protein folding or analyzing massive datasets.
  • "AI can do things that no human has ever evolved to be able to do..." (10:13, Socher)
• Shift in Developer Mindset: LLMs and modern AI models rely less on human-crafted rules and more on learning directly from labeled data.
  • "As a developer, you don't think about your own skills and logic that much anymore. You think about what does the neural network need? How do I clean my data..." (13:03, Socher)

3. Limitations and Capabilities of Current AI

• Zero-shot Capability & Biological Analogy: Some animal intelligence is genetically 'pre-trained', a useful analogy for the design of neural architectures.
  • "Biology figured out a way to store that learning in a genetic sequence such that when that brain gets instantiated and evolves in the womb, then it already has a set of knowledge..." (15:34, Socher)
• Surprises in LLM Progress: Progress in LLMs is impressive but not wholly unexpected; abstraction and generalization have predecessors in word embeddings and earlier work on prompt engineering.
  • "My dream had always been to build a single model for all of natural language processing. And so in 2018 we invented DECA NLP. ... we invented prompt engineering." (17:49, Socher)

4. Do LLMs Invent or Simply Interpolate?

• Mathematical Analogy & Extrapolation: Socher clarifies misconceptions—LLMs can extrapolate within a conceptual 'hypercube' and combine new concepts, not just interpolate between existing data.
  • "A big misconception people have is that these models can only interpolate... But that's actually not true." (21:55, Socher)
  • "I have images of black cats and I have images of yellow cars. Now the model ... will eventually be able to... create a yellow cat, even though it has never seen a yellow cat in the training data per se." (22:12, Socher)

5. How Far Can LLMs Go? The Limits of Scaling

• Scale vs. New Architectures: The most significant upcoming leap may be in having LLMs write and execute code—enabling them to solve problems more reliably. Further scaling (adding more data/parameters) may not be as fruitful as architectural advances.
  • "The biggest change for large language models will be their ability to program... You can actually force [an LLM] ... to translate [a problem] into code, run it, and give me an answer." (24:50, Socher)

6. Building Truthful AI: Engineering, Limitations, and Citations

• LLMs as CPUs & Orchestration Analogy: Socher likens LLMs to CPUs—powerful general engines which need to interact with memory, storage, the internet, and other specialized components.
  • "It's useful ... to think of the LM as the CPU of a computer. ... But ... it needs RAM ... a hard drive ... an Internet connection ..." (27:43, Socher)
• The Challenge of Truthfulness: Hallucination (making up information) is unavoidable; combining LLMs with real-time search, up-to-date data, and reliable citation is critical.
  • "Generative AI is only useful if the artifacts it produces are quick to verify, but would take you a long time to create yourself." (31:53, Socher)
  • "In large language models ... you can't verify if that answer is actually correct or not, and you don't know where the answer came from ... You need that verification." (33:15, Socher)
• Engineering Citations: It’s nontrivial for an AI to cite sources accurately; not just a software challenge, but an AI problem itself.
  • "The citation logic itself is also a hard AI problem. When do you use which resource for your facts?" (34:12, Socher)

7. The Future Industry Structure: Commoditization and Differentiation

• Commoditization of LLMs & AI: Open source models (like Mixtral) may outperform proprietary LLMs. The differentiators for companies will increasingly be design, engineering, and user experience—standard startup advantages.
  • "There's a good chance that large language models and maybe even AI will be commoditized, will be not the big differentiator. ... The main differentiator is in standard company startup tech stuff like marketing design, engineering..." (36:40, Socher)
• Industry Analogy: Azhar compares LLMs to traditional databases or engines—important but only part of a broader, composable stack, with much of the value in application, not just the core tech.

8. Regulation, Open Source, and the 'P(Doom)' Debate

• Regulatory Capture & Cynicism: Azhar suggests the narrative of dangerous AI may be a rational industry strategy to shape regulation and maintain first-mover advantage—a point Socher acknowledges (somewhat wryly).
  • "You might want to say your technology is so powerful that it's dangerous ... knowing that the government doesn't have the capability to assess that question..." (38:16, Azhar)
  • "We call it regulatory capture in Silicon Valley ... but it seems like the world will not adhere to that. Open source models are out." (39:03, Socher)
• AI Existential Risk ('P(Doom)'): Socher is firmly skeptical of apocalyptic AI scenarios, dismissing them as speculative science fiction not grounded in reality or real-world data. He instead favors pragmatic focus on real, already existing issues (bias, fairness, real-world applications), while supporting reasonable, targeted regulation where truly needed.
  • "It's a lot of cool sci fi scenarios, they're fun. I would probably watch the action movie ... but we got to keep it real and we can look at real problems, right? AI does have real problems. It will pick up biases ..." (43:18, Socher)

Notable Quotes and Memorable Moments

• Azhar on software engineering ambiguity:
  "It's kind of a disaster from an engineering perspective though, isn't it? ... Any amount of terrible enterprise software has been built from that kind of specification." (09:39)

• Socher on AI and evolution:
  "AI can do things that no human has ever evolved to be able to do..." (10:13)

• Socher on prompt engineering and LLM evolution:
  "My dream had always been to build a single model for all of natural language processing... And we invented prompt engineering." (17:49)

• Socher on LLMs and conceptual novelty:
  "The models can extrapolate a little bit, but not too far out...The most exciting stuff is where humans can't extrapolate anything because we're not evolved to look at, for instance, protein sequences or millions and millions of weather samples..." (22:12)

• Socher on programs as the next leap:
  "The biggest change for large language models will be their ability to program... That I think will give them so much more fuel for the next...few years." (24:50)

• Socher on citations and truthfulness:
  "The citation logic itself is also a hard AI problem. When do you use which resource for your facts?" (34:12)

• Socher on the future of AI as commodity:
  "There's a good chance that large language models...will be commoditized, will be not the big differentiator. ...The main differentiator is ... marketing design, engineering..." (36:40)

• Socher on AI existential risk hysteria:
  "A lot of cool sci fi scenarios, they're fun. I would probably watch the action movie that comes out...But we got to keep it real and we can look at real problems, right? AI does have real problems. It will pick up biases..." (43:18)

Timestamps for Key Segments

| Timestamp | Segment | |-----------|---------| | 02:15 | Defining Intelligence | | 03:53 | The historical roots of deep learning | | 06:42 | What is intelligence—from a brain perspective? | | 10:13 | Human vs. Artificial intelligence and developer mindset shift | | 13:03 | Sentiment analysis, feature engineering vs. neural networks | | 15:34 | Biological pre-training, architecture in animals and AI | | 17:49 | The surprising progress of LLMs, prompt engineering | | 21:55 | Limits and capabilities: do LLMs only interpolate? | | 24:50 | The biggest near-term leap: LLMs writing and executing code | | 27:43 | Architecture analogy: LLMs as CPUs, orchestration & engineering challenges | | 31:53 | Making generative AI truthful and verifiable; citation engineering | | 36:40 | Will LLMs be commoditized? Open source vs. proprietary future | | 38:16 | Regulatory capture, strategic narratives, open source's unstoppable rise | | 39:52 | AI existential risk ('P(Doom)'), science fiction vs. pragmatic issues | | 43:18 | Real-world AI risks; skepticism of sci-fi doom scenarios |

Conclusion

This episode provides a candid, accessible yet deeply informed tour through pressing questions in AI: what intelligence is, how models really work, whether LLMs are truly creative or just statistical parrots, the engineering realities behind 'truthful AI,' and the likely shape of the emerging AI industry. Richard Socher draws on his rare dual experience as both leading academic and entrepreneur to demystify the field, foregrounding scientific realism and product pragmatism over hype or fearmongering.
For listeners seeking to understand both foundational AI concepts and their real-world applications (and limits), this episode is indispensable.