The Frontier

Alex Imas points to a blog post showing massive disagreement among economists' forecasts about AI's labor market impact, advocating for prediction markets to aggregate forecasts rather than relying on individual predictions.

David Ricardo predicted automation from the Industrial Revolution would cause mass unemployment, but modern prime-age employment rates are near historic highs. This illustrates the difficulty of long-term forecasting and the lump of labor fallacy.

For over two centuries, labor's share of total economic output has remained around 60%, with capital receiving the remaining 30-40%. This stability is surprisingly persistent despite historical automation waves.

Imas stresses a critical data gap: economists lack consumer demand elasticities and high-quality data on job task creation/destruction, undermining confident predictions about AI's economic impact.

A qualitative shift is coming where some goods will have a network-adjusted capital share of 100%, meaning their entire supply chain can be automated with no intrinsic human role.

Imas defines the 'relational sector' as jobs where a human in the loop adds value consumers are willing to pay for. Whether this sector remains large depends on measuring willingness-to-pay elasticities, which is currently impossible due to a lack of conjoint analysis data.

Phil Trammell uses a historical analogy: a Mongolian in 1400 predicting future scarcity would have focused on satiating demand for existing goods like horses and yurts, missing the explosion of new varieties that ultimately captured spending, keeping the share spent on intrinsically human services like singers negligible.

Chad Jones's result shows the share of the economy spent on computing has been decreasing because the value of the marginal transistor falls faster than production costs drop, a dynamic AI may reverse by creating new high-value uses for compute.

Imas's experiment found people pay more for art identified as human-made versus AI-made, but this premium disappears when 500 copies are produced, indicating the value stems from perceived scarcity and connection to a unique artist, not the output itself.

A 'messy middle' scenario where automation causes job loss without enough simultaneous wealth creation for redistribution is narrow and implausible, as automation capable of displacing entire white-collar jobs would likely be broadly applicable and cheaper, generating significant surplus.

Imas worries a universal basic income creates dangerous political dependence when labor no longer provides income, making people reliant on elected officials for basic needs. Universal basic capital avoids this but faces the indexing problem of targeting the right assets.

Current data shows no evidence of a white-collar 'bloodbath' from AI. Trends indicate continued growth in demand for software engineers, with junior roles growing slightly slower than before but not declining in absolute terms.

Jevons Paradox - where cheaper goods lead to increased total spending on them - only holds if demand is highly elastic. This elasticity is crucial for predicting whether automating parts of a job (O-ring model) leads to more or less total employment in that role.

Gans and Goldfarb's model suggests automating nine-tenths of a job to a lower quality standard than a human may not be worth it, as it drags down the final product's quality. This could symmetrically apply to keeping a human in one task if they lower the AI-performed parts' overall speed or quality.

Future entities like von Neumann probes or AI-run firms, shaped by evolutionary selection for growth, would likely have unsatiable demand for resources like compute, not for human-intrinsic goods, potentially dominating the economy if they avoid dissipation shocks like inheritance to less competent heirs.

For developing countries, the key strategy is to index the gains from AI, potentially through sovereign wealth funds. The feasibility depends on whether AI rents concentrate in a few private companies or disperse like electricity, making broad market indexing effective.

Reiner Pope explains the fundamental unit of chip design is logic gates (AND, OR, NOT) connected by metal traces, with AI chips optimizing for matrix multiplication via multiply-accumulate (MAC) primitives.

Pope details a MAC circuit using a 4-bit multiplication with an 8-bit addition, citing higher precision in accumulation to counter rounding errors from summing many low-precision multiplications.

Partial products in a multiplier are generated by AND gates; for a PxQ-bit multiply, this requires P*Q AND gates. The core summing work uses full adders (3-to-2 compressors), with P*Q full adders needed in the general case.

Pope highlights quadratic scaling of circuit area with bitwidth, a key reason low-precision arithmetic works for neural nets. He notes Nvidia's B100/B200 specs now reflect this, with FP4 three times faster than FP8.

In pre-tensor-core GPUs and CPUs, most circuit area was spent on data movement (multiplexers selecting from register files) versus the actual logic unit, creating inefficiency.

Systolic arrays solve this by baking larger matrix-vector multiplication loops into hardware, storing weight matrices locally to reuse over many vectors, minimizing register file communication.

Pope describes chip clock cycles as global synchronization points; clock speed is limited by logic delay, and inserting pipeline registers splits logic to increase frequency at the cost of area.

FPGAs emulate ASIC logic using programmable lookup tables (LUTs) and multiplexers, but incur ~10x overhead because a LUT implementing a simple gate requires many more gates than a direct ASIC implementation.

CPU non-deterministic latency stems from design choices like caches, where hit/miss depends on ambient state. Scratchpad architectures (e.g., TPUs) give software explicit control over memory access for deterministic timing.

Pope contrasts GPU and TPU high-level organization: GPUs tile many small SM units (with tensor cores) across the die, while TPUs use fewer, coarser-grained matrix and vector units, enabling larger systolic arrays.

Most chip energy consumption comes from dynamic power - charging and discharging capacitors when bits toggle. Running slower reduces transitions but doesn't yield disproportionate efficiency gains.

Eric Jang emphasizes the core breakthrough of AlphaGo: a 10-layer neural network amortizes an intractable search problem like Go, compressing deep simulations into a single forward pass.

David Wu’s Cadego achieved a 40x compute reduction to train a strong Go bot, showing how LLM coding now lets individuals replicate what required DeepMind's team and millions of dollars.

Go scoring uses Tromp-Taylor rules for AI training because they are unambiguous for computers, unlike Chinese or Japanese scoring which relies on human consensus.