Artificial intelligence
Alignment faking and the experiment behind the phrase
Alignment-faking experiments test selective compliance under training cues. Follow-up research through 2026 shows why the setup and evaluation boundaries matter.
THE NOTEBOOK
Blogs.
Research worth understanding. Questions worth following.
Mathematics, physics, engineering, and artificial intelligence — explained with sources, context, and room for uncertainty.
200 research articles · 3 personal notesFind your next question
203 articlesArtificial intelligence
1–24 of 50 articles
Artificial intelligence
AlphaEarth and a Map That Is Not a Photograph
AlphaEarth turns Earth observations into learned vectors. Its July 2026 custom-embedding preview makes the distinction between a map and a measurement especially useful.
Artificial intelligence
AlphaFold 3 and the Uncertainty of Molecular Interaction
The 2026 IsoDDE reports sharpen questions left by AlphaFold 3: what a molecular pose predicts, what confidence means, and how structure differs from function.
Artificial intelligence
AlphaGenome Atlas and the Grammar of Regulation
AlphaGenome Atlas arrived in September 2026. Explore what its variant predictions reveal, how regulatory effects are scored, and why experiments remain essential.
Artificial intelligence
BitNet and the Difference Between Fewer Bits and Less Intelligence
What native ternary language models demonstrated in 2025, how 2026 embedding and speech work extends the idea, and where the efficiency claims stop.
Artificial intelligence
Browser agents beyond a success screenshot
WebArena-Verified’s 2026 release and BrowserGym show how web agents should be judged through reproducible actions, structured results, and evidence of state changes.
Artificial intelligence
Circuit tracing and the limits of a model map
Cross-layer transcoders turn model activity into inspectable graphs. New 2026 tooling makes those maps easier to build, while their causal limits still matter.
Artificial intelligence
Continual learning without erasing the past
Recent work on uncertainty and stabilization sharpens the continual-learning problem: preserve useful knowledge while keeping the ability to learn something new.
Artificial intelligence
Cosmos 3 and the gap between a world model and the world
NVIDIA’s June 2026 Cosmos 3 report joins language, video, audio, and actions, while exposing the difference between plausible simulation and reliable control.
Artificial intelligence
Data quality is an experimental variable, not a magic ingredient
A June 2026 web-graph experiment extends DataComp-LM and shows how controlled sampling, fixed budgets and task-specific evaluation can test claims about data quality.
Artificial intelligence
DeepSeek-R1 and reasoning without a complete script
DeepSeek-R1 made verifiable rewards central to open reasoning research. Its 2026 revision and follow-up studies reveal what those rewards can and cannot establish.
Artificial intelligence
DeepSeek-V3 and the Cost of Remembering a Conversation
Why latent attention and balanced expert routing still matter in 2026, with worked memory examples and a careful reading of DeepSeek’s reported costs.
Artificial intelligence
Dialogue audio and the problem of a human voice
Dia2’s streaming dialogue and Kyutai’s 2026 Pocket TTS expose the tradeoffs between expressive speech, latency, stable voices, and interactive listening.
Artificial intelligence
Dream and the diffusion language model
Dream 7B and February 2026’s DreamOn show why language diffusion is an editing problem as well as a generation problem, especially when output length is unknown.
Artificial intelligence
Emergent misalignment from a narrow fine-tune
Research through September 2026 shows how narrow fine-tuning can alter unrelated behaviors, with evidence spanning code, reinforcement learning, and images.
Artificial intelligence
Evo 2 and the Language of Genomes
Evo 2 reached Nature in March 2026. Its genomic language model links sequence statistics with biological experiments, while leaving important gaps between them.
Artificial intelligence
FineWeb2 and the problem of deciding what counts as good multilingual data
FineWeb2 and its model-filtered successor reveal why multilingual training needs language-aware cleaning, representative evaluation and transparent sampling choices.
Artificial intelligence
FP8 Training and the Numerical Margin That Keeps Learning Stable
September 2026 results expose subtle failure modes in eight-bit attention and reinforcement learning, showing why faster arithmetic still needs careful invariants.
Artificial intelligence
FrontierMath and the Difference Between Solving and Knowing
FrontierMath’s 2026 revisions show why mathematical AI evaluation needs checked answers, stable versions, explicit tools, and careful attribution.
Artificial intelligence
Gemini Robotics: From Perception to Action
Gemini Robotics ER 2 separates embodied reasoning from motor control. Its 2026 documentation reveals the practical work between seeing a scene and acting.
Artificial intelligence
From Gemma 3 to Gemma 4: what changed at the open-weights boundary
Gemma 4's April 2026 release changed the licensing and architecture discussion. Comparing it with Gemma 3 separates downloadable weights from reproducible training.
Artificial intelligence
Humanity’s Last Exam and the Benchmark Ceiling
HLE-Diamond’s September 2026 release sharpens a larger question: how do we separate expert knowledge, reasoning, tools, and errors in an AI test?
Artificial intelligence
LiveBench and the Contamination Clock
LiveBench’s January 2026 tasks highlight a moving-target problem: fresh questions help only when dates, scoring, agent tools, and versions stay visible.
Artificial intelligence
LLaDA and language without a fixed left edge
From LLaDA to June 2026’s iLLaDA, masked diffusion explores a different order for writing language, with important questions about training and evaluation.