Frontier radar
Where each research direction and practice stands, from research (papers and prototypes) through emerging (first real releases, claims still contested) to shipping (in products people use). Positions are reviewed every month; history is never rewritten.
History starts in November 2026. From next month on, a slider here replays how entries moved.
- 1Hybrid / linear attention
Most layers keep a fixed-size state; a few full-attention layers keep exact recall. Read more
- 2Mixture of Experts
Huge total capacity, small per-token compute. Standard in frontier models. Read more
- 3Diffusion language models
Parallel, iterative decoding. Several-fold faster in recent work; reasoning and long sequences still open. Read more
- 4Test-time training
Writes what the model reads into fast weights. Targets the frozen-knowledge gap directly. Read more
- 5Latent reasoning / recursive depth
Thinking in hidden vectors or looped layers instead of written tokens. Harder to monitor. Read more
- 6World models / JEPA
Predicting in an abstract representation space. Heavily funded, no public product yet. Read more
- 7RL with verifiable rewards
The recipe behind reasoning models and coding agents. Read more
- 8Test-time compute scaling
Buying accuracy with longer reasoning at inference time. Read more
- 9Synthetic data and distillation
Models generating and filtering training data for the next generation. Read more
- 10Automated AI research agents
Agents that edit training code and run experiments themselves. Read more
- 11Coding agents
RL-trained models in a tool loop with tests as the verifier. Read more
- 12MCP tool servers
A standard way to connect models to tools and data. Read more
- 13Small / local models
Distilled, quantised models for cheap, private, on-device steps. Read more
- 14Decision models
Typed answers with probabilities, no text generation. Jev, Strands Decider. Benchmarks still disagree. Read more
- 15Prompt caching
Reusing the KV cache of a repeated prefix. What makes agent loops affordable. Read more
- 16Prompt-injection defences
Classifiers, privilege separation, confirmation. All probabilistic; no architectural fix. Read more
- 17Decision-model tool gates
A fast model scores risk; plain code enforces the policy. Calibration has to be measured. Read more
- 18Agent sandboxing and least privilege
Scoped permissions, isolated execution, no auto-approval where network meets secrets. Read more
- 19Interpretability tooling
Sparse autoencoders and attribution graphs that show what a model computes internally. Read more