A strictly-curated feed of serious technical progress: foundational model releases, architectural work, safety and alignment research, and primary papers — drawn only from arXiv, Nature/Science, MIT Technology Review, IEEE Spectrum, and frontier research labs.
High-signal only · no product news, no hype
Multiple pre-training reports now show routed MoE architectures matching dense frontier models at a fraction of the active-parameter cost, shifting the efficiency frontier for the whole field.
A learned verifier steering inference-time search over candidate solution traces yields large gains on olympiad-level maths and competitive programming, at a controllable compute budget.
New alignment work studies whether a weaker supervisor can reliably elicit the full capability of a stronger model without also amplifying its failure modes — a core safety-research question for superhuman systems.
State-space and linear-attention hybrids report near-parity with softmax transformers on long-context language modelling while keeping memory linear in sequence length.
Dictionary-learning features extracted from a deployed model are shown to be causally editable and human-auditable at a scale previously only demonstrated on toy networks.
Production deployments report meaningful drops in cost and power draw from combining speculative decoding, paged KV-caches, and low-bit weight quantisation without measurable quality loss.
FAIR describes an early-fusion architecture that encodes text, image, and audio into one shared vocabulary, avoiding the adapter stacks that bolt modalities onto a frozen language model.
New rotation- and calibration-based methods hold accuracy within a point of full precision at 3-bit and below, changing what hardware a frontier model can run on.
A single policy trained on pooled data from many robot morphologies transfers zero-shot to unseen manipulators, echoing the pre-train-then-adapt recipe that reshaped NLP.
Diffusion-designed binders are now confirmed experimentally at hit rates high enough to compress the earliest, slowest stage of drug and enzyme discovery.
AI2's benchmark refreshes its questions on a fixed cadence and publishes canary strings, directly targeting the train-on-the-test-set problem that inflates leaderboard scores.
Independent measurements of non-von-Neumann accelerators show large joules-per-token gains for transformer inference, though tooling and model-mapping remain immature.
With high-quality text becoming scarce, new analyses quantify the returns from repeated epochs, synthetic data, and code-heavy mixtures — and where each stops helping.
Supervising intermediate steps rather than only final outputs produces more reliable long-chain reasoning and reduces confidently-wrong answers on multi-step tasks.
Mechanistic work isolates induction-style heads whose ablation selectively removes a model's ability to copy and generalise from examples in its prompt.
A controlled study finds that agent success rates collapse as task length grows because small per-step error rates compound — arguing for verification and checkpointing over bigger models.
A family of reference-free and length-controlled objectives reports steadier training and less reward hacking than both PPO-style RLHF and first-generation direct preference optimisation.
Graph-network screening followed by autonomous synthesis confirms a batch of previously unknown inorganic crystals, a template for closed-loop scientific discovery.
Recent history