Post-training & RLHF
The step after basic training where a model is taught to follow instructions, be helpful, and behave safely using human feedback.
Ranks moderate (62/100) — set apart by rising momentum (82) and structural importance (76).
Why it matters
It converts an inert base model into a helpful, safe, deployable product and encodes most of a model's observable behavior.
Why now
The frontier has shifted from scaling pretraining to scaling RL post-training, which is what produced 2025's reasoning models and is the main lever labs are pulling.
If Post-training & RLHF runs short
Weak or reward-hacked post-training yields sycophantic, jailbreakable or misaligned models regardless of base capability.
In depth · editorial + model
Instruction-tuning plus reinforcement learning from human (or AI) feedback that turns a raw pretrained base into a usable, aligned assistant. It converts an inert base model into a helpful, safe, deployable product and encodes most of a model's observable behavior. The frontier has shifted from scaling pretraining to scaling RL post-training, which is what produced 2025's reasoning models and is the main lever labs are pulling. Weak or reward-hacked post-training yields sycophantic, jailbreakable or misaligned models regardless of base capability.
Who makes Post-training & RLHF
The companies exposed to Post-training & RLHF
How to think about it
- Behavior is trained after capability
- A great base with bad post-training ships poorly
What to watch
- RL post-training compute share vs pretraining
- Preference-data vendor capacity (Scale, Surge)
- Reward-hacking and sycophancy incidents
Frequently asked
What is Post-training & RLHF?
The step after basic training where a model is taught to follow instructions, be helpful, and behave safely using human feedback.
Why does Post-training & RLHF matter for AI?
It converts an inert base model into a helpful, safe, deployable product and encodes most of a model's observable behavior.
Who makes Post-training & RLHF?
The companies the model tags as producers or suppliers of Post-training & RLHF: OpenAI, Anthropic, Scale AI, Uber Technologies, Inc..
Which companies are most exposed to Post-training & RLHF?
OpenAI, Anthropic, Scale AI, Uber Technologies, Inc. — 4 companies in total are mapped to Post-training & RLHF.
What happens if Post-training & RLHF runs short?
Weak or reward-hacked post-training yields sycophantic, jailbreakable or misaligned models regardless of base capability.
Where does Post-training & RLHF sit in the AI value chain?
Post-training & RLHF sits in the Models layer of the AI value chain.
Go deeper on Post-training & RLHF
- The materials, geographies and policies it depends on — heat-mapped
- Substitutes, relief valves and the domino chains if it tightens
- The live tension score, momentum and news drivers
- Four levels of analysis — from plain-English to strategic
model v0.7.0 · research, not advice
Model scores are illustrative reads from our model of the AI value chain — not investment advice.