In [6]: # specs.ipynb

Specialisation Premiumspremiums against pre-LLM NLP baseline

Specialisation is the steepest single contributor to base-salary variance after tier and level. Premiums are stacked: a frontier-lab LLM engineer at L5 sees the LLM premium on top of the T1 multiplier.

5.1Specialisation cards

8 specialisations

LLM / foundation-model

+15 to +35%

$230k - $320k

Pre-training, fine-tuning, scaling-laws work. Concentrated at T1 frontier labs.

Tooling stack

·Distributed training (FSDP, megatron)
·Scaling-laws empirical work
·RLHF / RLAIF post-training
·vLLM / TGI inference at scale
·Transformer architecture mods

demand: Extreme

RLHF / post-training

+15 to +30%

$220k - $300k

Alignment, fine-tuning, reward modelling. Highly compressed talent pool.

Tooling stack

·Reward-model training
·Online preference learning
·DPO / KTO methods
·Synthetic data pipelines
·Eval framework design

demand: Very high

Agentic systems

+10 to +20%

$200k - $275k

Tool-use, multi-step planning, autonomous agent frameworks. Emerging field.

Tooling stack

·Function calling / tool use
·Planning and search
·Agent eval methodology
·Browser and shell environments
·Memory systems

demand: High (emerging)

MLOps / platform

+0 to +10%

$190k - $260k

Training and inference infrastructure, experiment tracking, ML platforms at scale.

Tooling stack

·Kubeflow / Airflow / Argo
·MLflow / W&B
·Feature stores
·Triton / TGI / vLLM serving
·CI/CD for ML

demand: Very high

Multi-modal / vision-language

+10 to +18%

$200k - $270k

Vision-language models, image generation, video understanding, robotics-adjacent.

Tooling stack

·CLIP / SigLIP architectures
·Diffusion models
·Vision transformers
·3D representations (NeRF, GS)
·Cross-modal alignment

demand: High

Computer vision (classical)

+0 to +10%

$185k - $245k

Object detection, segmentation, real-time CV for autonomous and industrial use.

Tooling stack

·YOLO / Detectron / SAM
·ONNX / TensorRT
·Edge deployment
·Camera calibration
·Real-time CV pipelines

demand: Mature

Recommendation systems

+5 to +15%

$190k - $250k

Two-tower, graph, and sequence models powering ranking and personalisation.

Tooling stack

·Two-tower retrieval
·Embedding at scale
·Online learning
·Counterfactual eval
·Cold-start strategies

demand: High at platform companies

NLP (pre-LLM)

baseline

$170k - $225k

Classical IR, sentiment, NER, translation. Increasingly absorbed into LLM work.

Tooling stack

·BERT family fine-tuning
·Semantic search
·NER / classification
·Sequence-to-sequence
·Information retrieval

demand: Declining as standalone

5.2Frequently asked

3 questions

Q.What is the highest-paying ML specialisation in 2026?

▸

A.LLM and foundation-model engineering commands the highest premium, +15 to +35 percent above generalist ML at the same level. The premium is concentrated at T1 frontier labs and is highest for engineers with hands-on pre-training experience at the 100B+ parameter scale.

Q.Should I specialise in MLOps or stay a generalist ML engineer?

▸

A.MLOps offers a smaller but more durable premium (typically 0 to +10 percent on base) and very high demand. The role is critical at any company shipping production ML. If you enjoy infrastructure, tooling, and reliability over model research, MLOps is one of the highest-ROI specialisations because the premium compounds with hyperscaler tier multipliers.

Q.Is RLHF a real specialisation or just LLM work?

▸

A.It is increasingly a distinct specialisation. RLHF, post-training, and preference learning have their own labour pool, often with reinforcement learning or recommendation systems backgrounds. The supply is compressed enough that premiums of +15 to +30 percent are routine at T1 and T3 labs.

→ next: total comp → vs data scientist