Skip to content
Slow Variables
Ask

Methods · Training input

Public text data exhaustion year (Epoch projection)

Epoch's central projection of the year in which language-model training fully utilises the effective stock of public human-generated text, from the 2024 paper.

Reason through this

The reading

A projection, not a reading; the 80% interval runs 2026 to 2032, and the estimate moved four years later between Epoch's 2022 and 2024 papers.

Every readingeach dot links to its record
year
public text · 2028 · 2024-06-06 · estimate
a readingnormal rangefast range
estimate

1 reading drawn from epoch_blog.public_text.exhaustion_year.pt; 1 observation in the evidence.

Source: Epoch AI research posts. Epoch AI (CC BY 4.0).

The numbers
as ofreadingvaluerecord
2024-06-06public text2028estimate

Why it matters

Bottleneck 88 (data exhaustion) as a dated estimate: when the readily available data is spent, continued scaling becomes a business decision about synthetic and licensed data.

What cuts against it

One research group's estimate; synthetic data, licensed corpora and multimodal data all extend the stock; the estimate has already been revised once by four years.

How it is measured
  • series epoch_blog.public_text.exhaustion_year.pt
  • source Epoch AI research posts · default tier 6 · CC BY 4.0
  • proxy benchmark · cadence annual · valve invention to product

Normal = exhaustion in 2028 or later (the paper's compute-optimal case); fast = 2026 or earlier (the overtrained case, which is what the AI 2027 scenario's compute path implies). Between is emerging.

Applied to epoch_blog.public_text.exhaustion_year.pt.

Latest readings
  • 2024-06-06 public_text · 2028as of 2024-06-06
Why this status, and its history

Epoch's 2024 projection puts full utilisation of public human-generated text in 2028 for compute-optimal training (80% interval 2026-2032), inside the normal band (2028 or later); the 2022 paper had said 2024. One research group's estimate, tier 6 and single-source, so the status is capped at emerging; refreshed when Epoch republishes. Initial seed.

  1. 2026-09-10unmeasured to emergingconf — → 40 · the evaluator (automatic)

    Epoch's 2024 projection puts full utilisation of public human-generated text in 2028 for compute-optimal training (80% interval 2026-2032), inside the normal band (2028 or later); the 2022 paper had said 2024. One research group's estimate, tier 6 and single-source, so the status is capped at emerging; refreshed when Epoch republishes. Initial seed.

    source1
Why this confidence

Rests on 1 observation from Epoch AI research posts; best evidence tier 6, grade C; current.

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.

Predictions that lean on this