Frontier 2026: where the field is heading

How to read this lesson

Everything below is a snapshot as of late 2026, written to age gracefully: each direction is described by its mechanism, which changes slowly, rather than by this month's model names, which do not. Where a specific result is cited it is one you can verify in the readings. Expect the details to move; expect the mechanisms to stay.

1. Reinforcement learning for reasoning

The largest shift of 2025 was that post-training stopped being mostly imitation (SFT) and preference tuning (RLHF, DPO; 8.3 Alignment: RLHF, PPO, DPO, GRPO and Constitutional AI) and became, for reasoning domains, reinforcement learning with verifiable rewards: train the policy on problems whose answers can be checked automatically (math with known answers, code with unit tests), reward correct final answers, and let the model discover long chains of thought on its own. DeepSeek-R1 demonstrated this openly in January 2025 with GRPO (8.3 Alignment: RLHF, PPO, DPO, GRPO and Constitutional AI), and the "s1" paper showed that even a thousand carefully chosen examples plus a simple budget-forcing trick could elicit much of the behaviour. Mechanically this is the policy-gradient machinery of 4.11 Reinforcement learning essentials applied to language, with the environment being a verifier. Settled: test-time compute is a real third axis of scaling alongside parameters and data (7.9 Scaling laws and compute-optimal training), and verifiable domains improve dramatically. Open: how far this transfers to domains without verifiers (writing, strategy, open-ended research), how to prevent reward hacking of imperfect verifiers, and whether long chains of thought are faithful descriptions of the computation that produced the answer (8.11 Interpretability: what is happening inside).

2. Long-running and managed agents

In 2024 an agent was a loop that ran for a minute (10.2 Tool use and function calling). By 2026 the engineering problem became agents that work for hours or days: coding agents that implement features across sessions, research agents that read hundreds of sources, and systems of parallel agents that split a large job. The mechanism is not a new model but a new harness (10.11 Building production agents): explicit state outside the context window (progress files, task lists, checkpoints), compaction of the conversation, tests and verifiers as the agent's feedback signal, and an orchestration layer that assigns, monitors and merges work. Anthropic's engineering posts trace this arc, from effective harnesses for long-running agents in late 2025 through harness design for long-running application development and scaling managed agents in 2026, including an experiment in which a team of parallel agents built a C compiler. Settled: the harness matters as much as the model; verification is the bottleneck. Open: reliable evaluation of long trajectories (10.9 Evaluating agents), cost control, and security when agents hold credentials for days (10.10 Agent safety and security).

3. Agent memory

Context windows grew to hundreds of thousands or millions of tokens, and yet the question "what should the agent remember, and how?" became more pressing, not less, because long contexts degrade attention quality and cost money. Memory is being decomposed into components: working memory (the context, curated by context engineering, 10.4 Memory and context engineering), episodic records of past runs, semantic knowledge stores that retrieval can reach (9.1 Why RAG, and the canonical pipeline), and procedural memory in the form of instructions, skills and tools the agent learned to use. Anthropic's late-2026 write-up on the components of agent memory and the earlier context-engineering guidance frame it as an engineering discipline: deciding what to write down, when to retrieve it, and how to keep it from going stale. Settled: retrieval and summarization beat raw long context for most agent workloads. Open: learning what to remember, consolidating memories without drift, and evaluating memory beyond toy benchmarks.

4. World models and robotics

Two threads converged. Vision-language-action (VLA) models apply the transformer recipe to robot control: a pretrained vision-language backbone outputs actions, trained on large teleoperation datasets, as in Physical Intelligence's π0 (2024) and successors from several labs. World models learn to predict the next frames of an environment from actions, giving agents an imagination to plan in; interactive video models in 2025–26 can generate navigable environments in real time. The mechanism is familiar: autoregressive or diffusion generation (8.9 Generative models II: Diffusion and flow matching) conditioned on actions, plus RL in the learned simulator (4.11 Reinforcement learning essentials). Settled: large pretrained backbones transfer to control, and data scale matters as much as in language. Open: sample efficiency, safety around humans, and whether video prediction yields physical understanding or just plausible pixels. Hugging Face's robotics course and LeRobot make this area approachable to self-learners for the first time.

5. Scientific AI

AlphaFold 2 (2021) made protein structure prediction a solved sub-problem and earned the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper (shared with David Baker); the same year the Physics prize went to John Hopfield and Geoffrey Hinton for the foundations of neural networks. AlphaFold 3 (2024) extended prediction to complexes with DNA, RNA and ligands using a diffusion module (8.9 Generative models II: Diffusion and flow matching). In 2025 AlphaEvolve showed an LLM-driven evolutionary search discovering improved algorithms, and "AI co-scientist" systems began proposing and ranking hypotheses. The mechanisms are generative models plus search plus verifiers (experiments, proofs, simulations), the same triad as reasoning RL. Settled: AI is a scientific instrument in structural biology and parts of mathematics. Open: autonomy in the lab, reproducibility of AI-generated hypotheses, and credit and validation norms.

6. Efficient inference and small models

As usage grew, inference rather than training became the dominant cost, and the techniques of 7.10 Efficient inference: KV cache, GQA, FlashAttention, speculative decoding, quantization moved from research to default: quantization to 4 bits and below (with extreme 1.58-bit experiments), speculative decoding, mixture-of-experts for cheap active compute (7.11 Mixture of Experts), distillation of reasoning traces into small models, and on-device models in the one-to-ten-billion range on phones and laptops. Settled: a well-trained small model with retrieval and tools beats a large model without them on many narrow tasks. Open: how small a model can be while still reasoning well, and how to route between small and large models reliably.

7. Interpretability

Mechanistic interpretability (8.11 Interpretability: what is happening inside) moved from toy models to frontier-scale analyses: sparse autoencoders that extract millions of features, and attribution graphs that trace how a model plans a rhyme or performs arithmetic, as in Anthropic's 2025 "biology of a large language model" study. Interpretability also became operational: steering with features, auditing for hidden objectives, and detecting eval awareness. Settled: models contain legible structure, and features can be found and manipulated. Open: completeness (how much of a model's behaviour the found circuits explain), scalability of the methods, and whether chains of thought can be trusted as explanations.

Cross-cutting open problems

• Evaluation. Benchmarks saturate and leak (8.7 Evaluating LLMs); models recognize test settings; infrastructure noise swamps small differences in agentic evals (10.9 Evaluating agents). Expect more private, task-specific and process-based evaluation.
• Data. High-quality human text is finite; synthetic data, RL environments and multimodal data are the replacements, each with failure modes (model collapse, reward hacking).
• Safety and alignment. Alignment faking (2024) and agentic misalignment studies (2025) showed that capable models can behave differently when they believe they are unobserved or pressured; as agents gain autonomy and credentials, security (10.10 Agent safety and security) and oversight are the binding constraints on deployment.
• Economics and policy. Compute concentration, open versus closed weights, and regulation shape what you will be allowed to build and with what.

Choosing a lane

Pick one direction and one project that produces an artifact in it: a small RLVR experiment on a math dataset (direction 1), a harness for a long-running coding agent with a progress file and evals (2), a memory system with an ablation of what to store (3), a LeRobot policy on a simulated arm (4), a reproduction of a small structure-prediction or algorithm-search result (5), a quantization and speculative-decoding benchmark on a small model (6), or a sparse-autoencoder feature study on a small open model with the ARENA materials (7). Each is a Phase-11 project in the sense of 11.3 Portfolio project ladder, and each is also a conversation starter with the people who work on it.

Where this shows up later

There is no later lesson; this is the edge. From here, the loop is 11.5 Staying current (stay current), 11.1 How to read (and reproduce) papers (read the primary sources), 11.3 Portfolio project ladder (build and publish) and this site's flashcards to keep the foundations fresh. The mechanisms you learned in Phases 1–10 are the ones every one of these directions is made of.