Not a news feed — a map. The research currents, the people steering them, and the papers that set the coordinates. Curated, opinionated, and linked back to the source.
Spend compute at inference, not just training.
Models now allocate variable "thinking" budgets — chains of internal deliberation that lift performance on math, code and proofs far beyond single-pass generation.
From answering questions to completing work.
Reason–act loops let models plan, call tools, browse and write code across many steps — turning the LLM from oracle into operator.
Predict consequences, not just the next token.
Non-generative, predictive architectures learn abstract representations of how the world evolves — a proposed path to planning and physical common sense.
Open the black box before it opens us.
Sparse autoencoders and circuit analysis decompose model internals into human-readable features — the foundation for steering and safety at scale.
Intelligence per dollar is the new frontier.
Compute-optimal training and mixture-of-experts routing decouple capability from cost — shifting the bottleneck from training FLOPs to serving efficiency.
A million tokens, then a lifetime of them.
Million-token windows and external memory let models reason over whole codebases and corpora — making retrieval, caching and forgetting first-class design problems.
Reality Compression
"Predicting the next token is discovering the hidden laws that produced it."
Superalignment as the central problem: as compression deepens toward AGI, the data model becomes an executable world-model enabling System-2 reasoning.
Safe Superintelligence ↗The LLM OS
"The LLM is the new CPU; the context window is the RAM."
A kernel for computing where the model manages memory, runs code, and orchestrates tools and peripheral APIs as a unified operating system.
karpathy.ai ↗World Models
"True intelligence requires predicting consequences, not just generating tokens."
Advocates moving past autoregression toward predictive world-models (JEPA) that grasp physical constraints, planning and abstraction.
V-JEPA at Meta AI ↗Powerful AI, Soon
"A country of geniuses in a datacenter — and the duty to make it go well."
Frames the upside of aligned AI across biology, neuroscience and governance, while treating interpretability and safety as preconditions, not afterthoughts.
Machines of Loving Grace ↗AI for Science
"Solve intelligence, then use it to solve everything else."
From AlphaFold to AlphaProof, points the field at AI as a scientific instrument — compressing decades of discovery in biology, chemistry and mathematics into years.
Google DeepMind ↗Attention replaces recurrence. Parallelizable training makes scale possible for the first time.
Bigger is predictably better. Few-shot prompting hints at general capability emerging from scale.
InstructGPT and ChatGPT turn raw predictors into usable assistants. AI goes mainstream overnight.
Llama opens the floodgates; RAG, function-calling and the first agents move from demo to product.
Test-time compute arrives — models that deliberate before answering reset the hardest benchmarks.
DeepSeek-R1 shows frontier reasoning at a fraction of the cost, redrawing the economics of the field.
Models that act, see and remember — long-horizon agents wired into tools, with safety and interpretability as the gating constraints.