THE CORPUS / SIX LENSES
The evidence.
And the arguments.
Sixteen papers, read together. Follow the connections, compare the machinery, and keep the interpretation open to question.
GPT-3 / Few-Shot Learners ↗
Scaling an autoregressive LM to 175B parameters makes task adaptation possible purely by placing examples in the context, with no gradient updates at inference.
WebGPT ↗
Put a language model behind a real text browser, let it search and quote, and train it with human feedback so its answers carry verifiable references.
Chain-of-Thought ↗
Prompting a large model with a few worked examples that show intermediate steps elicits step-by-step reasoning and sharply improves multi-step tasks.
ReAct ↗
Interleaving reasoning traces with environment actions lets each correct the other - thinking plans the next action, and observations correct the thinking.
Toolformer ↗
A model can teach itself when and how to call APIs by self-annotating a corpus with candidate calls and keeping only those that reduce its own loss.
Reflexion ↗
Convert a failed trial into a verbal lesson, store it in episodic memory, and let the next attempt condition on it - reinforcement learning through language instead of gradients.
Self-Refine ↗
One model, no training, three roles: generate a draft, critique it, and revise using the critique - iterated until it stops improving.
Voyager ↗
An agent that writes successful behaviours back into an executable skill library, under an automatic curriculum, keeps getting more capable without any weight update.
InterCode ↗
Treat code as the action space and execution as the observation: a standardised, sandboxed environment where interactive coding agents can be benchmarked with real execution feedback.
DSPy ↗
Stop hand-writing prompts. Express an LM pipeline as declarative modules with typed signatures, then compile it against a metric, letting an optimiser choose the prompts.
MemGPT ↗
Borrow virtual memory from operating systems: give the model a small in-context working set plus a large external store, and let it page between them with its own function calls.
Darwin Godel Machine ↗
Drop the Godel machine's demand for proof. Let agents rewrite their own code, validate each change empirically on benchmarks, and keep an archive of all variants as stepping stones.
Recursive Language Models ↗
Treat a long input as external data: inspect and split it with code, and make recursive model calls over selected parts.
Continual Harness ↗
Formalise the harness as state H = (prompt, sub-agents, skills, memory), then remove the human: a Refiner reads the trajectory mid-episode and applies CRUD edits to all four, without ever resetting.
OpenJarvis ↗
Personal AI stacks fuse prompts, tools, memory, and runtime around one cloud model, so swapping in a local model breaks everything at once. Decompose the stack into five typed primitives and optimise them jointly.
Prime Agent ↗
A harness is a membrane. Make it expressive enough that the model builds its own strategy, and reliable enough that a failure is the model's fault rather than the harness's.
The rung is a teaching index inherited from the corpus. Higher does not mean better. The ordering is neither cumulative nor chronological; Prime Agent’s top-rung assignment is explicitly qualified on its page.