Minixhofer and colleagues converted subword language models to read raw bytes using 49.1 billion training tokens, under 1% of a typical pretraining budget. Byte-level models have long trailed subword systems, and starting from a trained model is how this work closes most of that gap.
Reality
- Evidence60
- Adoption5
- Hype gap+15
- Incentives
- Insufficient
- Confidence55
buildOne report1 publisher Clockwork.io raised $31 million as LinkedIn runs its LinkPass failover across its AI infrastructure and Together AI sells TorchPass job recovery. Restart-free recovery now runs on paying GPU clusters, though the savings figures all come from customers.
Reality
- Evidence45
- Adoption50
- Hype gap+20
- Incentives70
- Confidence55
Clockwork Systems raised $31 million and launched TorchSnap, a tool that snapshots distributed AI jobs so they can resume after hardware failures. The customer deployments it cites predate the new feature, so its value depends on how long a buyer's own restarts take.
Reality
- Evidence40
- Adoption35
- Hype gap+30
- Incentives70
- Confidence45
buildOne report1 publisher That ratio sets the per-token bill and how much of a document fits in context, and a vocabulary swap can nearly erase it. The accuracy gap behind Irish exam scores is a different problem and does not move with the tokenizer.
Reality
- Evidence52
- Adoption
- Insufficient
- Hype gap+22
- Incentives30
- Confidence42
buildOne report1 publisher The argument is that GPU failures are the expected case at scale, so recovery speed sets goodput and therefore spend. Two subsystems most teams write once and forget decide how much idle time each failure costs.
Reality
- Evidence40
- Adoption28
- Hype gap+18
- Incentives80
- Confidence55
buildOne report1 publisher Micron's HBM design fellow says the memory's roughly threefold wafer-area penalty against DDR5 grows with every generation. The pricing that follows does not mean-revert on its own.
Reality
- Evidence66
- Adoption74
- Hype gap+14
- Incentives68
- Confidence61