LECTURE · FOR MY AI BROTHERS IN ARMS
Scaling laws · agent horizons · inference economics · quantization
Antonín Woff · woffx.eu
5×
annual growth in training compute
5.2 mo
doubling time for training compute
2.2×
annual growth in power draw; runs now exceed 100 MW
Schematic · light: frontier · grey: training runs · dashed: floor E
L(N, D) = E + A·N⁻ᵅ + B·D⁻ᵝ · L ∝ C⁻⁰·⁰⁵: 10× compute buys ~11 % lower loss
I · BREAKTHROUGH — Epoch AI (2024–25); Kaplan et al. (2020); Hoffmann et al., Chinchilla (2022)
02 / 09
Schematic of the method · — newer model · — older model
P(success | t) = σ(a − b·log₂ t) · time horizon h₅₀ = t at which P = 0.5
~7 mo
doubling time of the 50 % time horizon
<3 mo
doubling time on SWE-bench Verified
170
software, cyber-security and ML tasks, 2019–2025
I · BREAKTHROUGH — METR, Kwa et al. (2025), Measuring AI Ability to Complete Long Tasks
03 / 09
Light band: cost of running the frontier · grey band: price of a fixed capability
cost / task = price / token × tokens / task · reasoning inflates the second term
280×
cheaper at GPT-3.5 level, Nov 2022 → Oct 2024
50× / yr
median annual price decline at fixed capability
3–18× / yr
annual rise in the cost of running the frontier
II · DIFFUSION — Stanford AI Index 2025; Epoch AI (2025); Gundlach et al., MIT (2025)
04 / 09
Algorithmic progress halves the compute needed for a given result roughly every eight months.
Public human text, ≈ 300T tokens, is likely exhausted between 2026 and 2032. Schematic.
Memory = parameters × bytes per weight. At 4 bits, a 32B model fits one consumer GPU.
II · DIFFUSION — Ho et al., NeurIPS 2024; Villalobos et al., Epoch AI 2024; Dettmers et al., QLoRA 2023
05 / 09
Measured: Llama 3 70B, GGUF quants, wikitext-2 perplexity. Q4_K_M highlighted.
≈ 70 %
smaller than FP16 at Q4_K_M (131 → 40 GiB)
+4.8 %
perplexity cost at Q4_K_M, ~4.8 bits per weight
+249 %
at IQ1_S, ~1.7 bits per weight: coherence collapses
II · DIFFUSION — llama.cpp perplexity measurements, Llama 3 70B (dranger003); llama.cpp Llama 2 70B scoreboard
06 / 09
758 BCG consultants · GPT-4 · 18 realistic tasks · randomised field experiment
Inside the frontier, gains skew to the bottom: below-average performers +43 %, above-average +17 %
III · IMPACT & PRACTICE — Dell'Acqua et al. (2023), Harvard Business School & BCG
07 / 09
TIER
COST DYNAMICS
CAPABILITY
BEST FOR
running cost up 3–18× a year
state of the art
novel, long-horizon, high-stakes tasks
same score 5–10× cheaper each year
roughly last year's frontier
high-volume, routine workloads
hardware only, no per-token fee
~6–12 mo behind; weights ≤28–40B params fit one GPU
sensitive data, air-gapped agents
Pattern: route to the cheapest tier that clears the bar, escalate on low confidence · n8n · Ollama · NeMo Guardrails
III · IMPACT & PRACTICE — Gundlach et al., MIT (2025); Epoch AI (2025–2026)
08 / 09
01
If a fixed capability grows ~50× cheaper each year, what still justifies frontier prices?
02
How would we even notice our own agents straying beyond the jagged edge?
03
With open weights trailing the frontier by mere months, can capability-based regulation still bite?
STAY CURIOUS · STAY CRITICAL · THANK YOU
09 / 09