BACK TO THE FUTURE: LISP IN THE NEW AGE OF AI - European Lisp Symposium

Summary

Anurag Mendhekar argues that Lisp’s historical strengths—especially its ability to represent and manipulate programs as data—could make it valuable in an AI-driven future. He sees Lisp not merely as another programming-language choice, but as a way to teach people to understand entire computing systems.

Lisp and the first era of AI

Early AI was shaped by the belief that intelligence could be built from knowledge and reasoning:

• Epistemology described what a system knows.
• Heuristics described the methods it uses to act on that knowledge.

Lisp suited this approach because a small set of constructs—especially lists, quotation, and cons—could represent complex symbolic knowledge. Its interactive environment also made it possible to inspect, change, and evolve programs while working.

Mendhekar describes Lisp as a major force in early AI and commercial computing. He cites expert systems, Lisp machines, and the Connection Machine as examples of the era’s ambitious applications and hardware.

In his account, the approach ran into trouble as knowledge bases became too complex and difficult to maintain. As confidence and investment fell, hardware markets collapsed, followed by the AI winter.

From symbolic knowledge to statistical AI

The talk traces a shift from logic and symbolic representations toward statistics and neural networks. Developments such as backpropagation, convolutional neural networks, numerical libraries, GPUs, CUDA, and AlexNet helped make modern deep learning practical.

In statistical AI, knowledge is encoded in model parameters rather than in explicit, inspectable symbolic structures. The work relies on training and inference, often using specialized libraries and hardware.

Mendhekar argues that this makes models harder to inspect and explain, and can lead to errors such as hallucinations. In this setting, Lisp is not inherently required: it is one language option among many for connecting to models, libraries, and APIs.

Lisp’s influence and opportunity

Mendhekar says Lisp is relatively uncommon in today’s AI repositories, but its ideas have influenced many widely used languages and practices. He points to interactive development, garbage collection, recursive data structures, and functional programming, as well as Lisp’s influence on languages including JavaScript, Python, and Ruby.

He identifies Lisp’s distinctive contribution as treating programs themselves as manipulable knowledge structures. Macros, evaluation, compilation, and interactive inspection let programmers work on a program’s representation, not just run it.

He uses Symbolics Genera as an example of this philosophy taken to an extreme: much of the system could be inspected and modified as objects, with the ability to follow definitions and behavior down through layers of the system.

Mendhekar’s prediction: fewer coders, more system-level thinkers

Mendhekar presents a chart showing a decline in software hiring at major technology companies after late 2022, and suggests that AI is changing demand for developers. During the Q&A, he acknowledges that pandemic-era overhiring and other factors also contributed; he does not claim AI explains the entire decline.

His forecast is that AI will increasingly generate code, while people specify what to build, evaluate it, and guide its development. Small teams may be able to create and maintain systems that once required many specialized developers.

He argues that this shifts the need from narrowly skilled coders toward “deep system visionaries”: people who can understand users and their needs, design products, reason about algorithms and architecture, and diagnose problems throughout the software stack—even when AI wrote much of the code.

Such people still need technical depth. AI can make mistakes, and fixing them requires understanding underlying principles, not merely familiarity with a language or tool.

Educational proposal: teach computing through Lisp

Mendhekar’s central recommendation is to teach more of computer science through Lisp, emphasizing first principles rather than the details of particular tools and languages. He proposes that education should:

• Use a common notation across different areas of computing, as mathematics uses shared foundations and notation.
• Teach the fundamental principles of each field while allowing AI to handle more routine implementation work.
• Help students understand how systems work at multiple levels, so they can inspect, debug, and reason about components throughout the stack.
• Reduce the burden of learning what he calls “arcania”—specialized syntax, conventions, and tooling that can distract from core concepts.

He urges universities to reconsider curricula built primarily around preparing students for conventional coding jobs.

Discussion and qualifications

• In response to concerns about AI-generated errors, Mendhekar says first-principles knowledge is essential for debugging; he does not argue that technical understanding becomes unnecessary.
• He expects AI-assisted development to speed up experimentation and potentially increase innovation, rather than cause stagnation.
• He says language choice may become more flexible as AI can generate bindings, translate code, or help build tools. In his view, this could make ecosystem size and available talent less decisive, leaving programmers freer to choose languages they prefer.
• He stresses that large, reliable systems still require people who understand how the whole stack works.
• On code quality and readability, he says AI can be guided with explicit rules and can sometimes detect and correct its own mistakes, though it can still hallucinate.

Speakers and sources featured

• Anurag Mendhekar — presenter.
• Unnamed audience members — asked questions about hiring trends, AI-generated errors, Lisp education, GitHub statistics, hardware and software design, and the future role of programmers.
• Unnamed moderator or conference host — introduced the Q&A, relayed an online question, and closed the session.
• Online audience member — asked whether AI-assisted development might cause technological stagnation; the moderator relayed the question.

People and sources cited in the talk: Dan Friedman and The Little Learner; Gregor Kiczales and John Lamping; John McCarthy; Peter Lee; Peter Norvig; James Gosling and Guy Steele; Brendan Eich; Guido van Rossum; Danny Hillis and Thinking Machines; a Reddit comment about inspecting a Lisp system “all the way down”; GitHub repository statistics presented by Mendhekar; and statements attributed to companies including Facebook, Microsoft, and Y Combinator.