Notes

Why the name "Everything is Everything" matters

The show's title comes from Bruce Springsteen's "You're Missing" from *The Rising*, written after 9/11. The song describes someone whose loved one has died, listing ordinary objects still present in their absence. The refrain "Everything is everything, but you're missing" captures both the interconnectedness of all things and the profound impact of absence.

Amit explains why this resonates beyond the song itself:

"Both of us are people who don't want to look at the world through one frame. We are interested in many different subjects and we want to look at them through many different frames. And we want to emphasise the interconnectedness of everything, that we contain multitudes, the world contains multitudes and every piece of us is joined to every piece of everything else."

Ajay adds that the phrase suggests a calm, thoughtful approach to complex problems rather than rage or emotional reaction. This reflects their belief that understanding the world requires sustained, interdisciplinary thinking rather than narrow specialisation.

The statistical model argument against AI terminology

Ajay objects to calling current systems "artificial intelligence," arguing they're statistical models processing data rather than exhibiting genuine intelligence. He emphasises his background working with these systems:

"Everything that passes for AI in this world are really statistical models that are processing data. And you know, I love statistics. I love data analysis. That's what I've done for a living my entire life. And it's a great tool, but it is a highly limited tool."

The concern isn't with the technology's usefulness but with the terminology creating false expectations. Statistical models can flag X-rays for TB or predict words effectively, but this differs fundamentally from human consciousness, creativity, and purpose.

Ajay illustrates the limitation with a thought experiment about Einstein's special relativity:

"Let's imagine you're standing in 1905 and you have every text in the world in front of you. No LLM word correction under any prompting scheme will come up with Einstein's special theory of relativity."

The case that humans are also statistical models

Amit challenges the human exceptionalism in Ajay's argument, suggesting that humans may be sophisticated biological statistical models. He points to how much human communication relies on predictable patterns and cliches.

The observation that humans are "trained on LLMs that are a fraction of the size of modern LLMs" and have "a fraction of the processing power" leads to a provocative conclusion about AI potential. Regarding creativity, Amit argues:

"We ascribe a mysticality to that act of creation only because we don't understand it, right? We don't understand exactly how our neurons are firing and what they are drawing upon to create what we are creating."

This suggests the appearance of mystical creativity stems from ignorance about brain mechanisms rather than evidence of something beyond computational processes.

Where current AI systems prove genuinely useful

Despite his skepticism about the "intelligence" label, Ajay sees significant practical value in current AI tools, particularly for India's development context. He describes research showing these tools help convert second-quartile workers into third-quartile performers with appropriate oversight.

For India specifically, this addresses a crucial bottleneck:

"This is a tool made in heaven for the Indian production environment. We have very little high-end talent. The Indian top end capability is vanishingly small. There are very large numbers of weak people."

However, he emphasizes these remain "tools for masters"—requiring expert oversight to catch errors and verify output quality. The systems work best when someone with deep expertise can evaluate and refine their output.

The concern about learning pathways to mastery

Ajay worries that relying on AI tools too early might undermine the development of fundamental skills. Using writing as an example, he questions whether the craft of clear thinking and expression can be learned while depending on automated text generation.

The analogy to calculators provides a counterpoint—we stopped teaching manual multiplication without apparent harm. But Ajay leans toward writing and programming being different:

"I feel that people fixing up text manufactured by ChatGPT may permanently lose the craft of writing."

This touches on deeper questions about which human capabilities matter to preserve and develop versus which can be safely automated away.

AlphaZero and the excitement of genuine AI breakthroughs

Moving beyond language models, both speakers express genuine excitement about AI systems that discover new knowledge. AlphaZero's chess achievements exemplify this—not just playing stronger chess, but developing novel strategic concepts that changed how human masters think about the game.

Amit describes the revolutionary impact:

"Alpha Zero discovered the Berlin defense. The same evolutionary path and then going further beyond where humans are and then doing things that we don't understand."

Magnus Carlsen's coach captured the significance: wondering how aliens with superior intelligence would play chess, "and now I know" after seeing AlphaZero games.

The system didn't just optimize existing approaches—it developed genuinely new strategic insights about material versus initiative, flank attacks, and long-term sacrificial play that human players then adopted.

Algorithmic breakthroughs that solve fundamental problems

Ajay highlights AI achievements in algorithm optimization that represent genuine advances in human knowledge. The breakthrough in matrix multiplication algorithms particularly impressed him:

For 50 whole years, no human had ever been able to figure out how to do it better. And what the scientists were able to do was again to gamify it, to represent the very algorithm used for multiplying matrices as a parameter vector, and then optimize.

Similar advances in sorting algorithms demonstrate machines helping improve the fundamental tools that power all computing. This creates a feedback loop where better algorithms enable more powerful machines, which can then discover even better algorithms.

This connects to classic science fiction themes about machines helping build better machines, potentially leading to recursive improvement cycles.

The mundane reality of AI safety concerns

Rather than worrying about existential threats, both speakers focus on practical safety challenges comparable to any powerful technology. Ajay draws parallels to automobile safety:

"When humans built the automobile, there were accidents and people died and you don't describe agency to the automobile, you describe agency to the driver."

Medical diagnosis AI will make mistakes and people will die. Military robots will kill civilians. These represent engineering challenges about error rates, human oversight, and appropriate deployment contexts rather than consciousness or malevolent intent.

The real work involves designing systems that appropriately balance different types of errors while maintaining human responsibility for consequential decisions.

Supplementary Resources

1. The Hidden Dangers of LLMs, Big Ideas, Episode 55 (video)
2. Bender et al "On the Dangers of Stochastic Parrots" (paper)
3. Yann LeCun "A Path Towards Autonomous Machine Intelligence" (paper)
4. LLMs don’t do formal reasoning - and that is a HUGE problem (article)

The complete transcript file is available to download below.