Notes on AI, research, and the work.
234 pieces on generative and agentic systems, the research underneath, and what it takes to ship them.
- Jul 2026 · 7 min
Why VLMs choke in the middle of long documents
SynthDocBench isolates document length, layout, modality, and question type as independent variables and finds that vision-language models fail hardest on the middle third of long documents, with chart comprehension collapsing separately.
vlmbenchmarkslong-contextdocument-understandingRead - Jul 2026 · 7 min
What 18 million hours of accelerometer data taught researchers about wearable motion models
Inertia-1, an open study from UCLA, Duke, and Johns Hopkins, pretrains on 18.2M hours of accelerometry from 115,000+ people to isolate which design choices actually make wearable motion foundation models transfer — and finds that data pipeline decisions and data scale matter more than model size.
wearable-aifoundation-modelsself-supervised-learninghealthcare-aiRead - Jul 2026 · 7 min
Image generators fabricate what they don't know, and search alone doesn't fix it
A new benchmark shows frontier visual generators collapse to 21-28/100 on prompts that require world knowledge outside their training data, and naive retrieval-augmented generation makes the problem worse, not better.
visual-generationretrieval-augmented-generationagentic-aibenchmarksRead - Jul 2026 · 5 min
An operating system for robots, and the benchmark problem underneath it
ABot-AgentOS proposes a general runtime layer — planning, verification, and lifelong graph memory — that sits above existing VLM/VLA controllers, but its headline embodied-task gains rest on an early, self-authored benchmark subset rather than the third-party evaluations it also reports.
roboticsagent-architectureembodied-aibenchmarksRead - Jul 2026 · 6 min
A neural network potential is only as good as the reaction path it's tested on
A new framework called Enerzyme trains electrostatics-aware neural network potentials on enzyme active sites for a fraction of the usual data, and in doing so exposes a real problem with how these models get validated.
neural-network-potentialsmachine-learningcomputational-chemistryenzyme-catalysisRead - Jul 2026 · 7 min
Dense reward, sparse progress: what long-horizon-terminal-bench tells us about agent autonomy
A new 46-task benchmark grades terminal agents on partial credit instead of pass/fail, and the results are humbling: the best frontier model clears only 15.2% of a multi-hour task at a 0.95 reward threshold, with a 15-model mean of just 4.3%.
ai-agentsbenchmarksllm-evaluationagentic-aiRead - Jul 2026 · 5 min
Scaling mixture-of-experts video pretraining for embodied AI
LingBot-Video pairs a from-scratch MoE video architecture with robot-oriented pretraining data and physics-aware rewards, aiming to close the gap between video generators built for content creation and video models useful as robotics world models.
mixture-of-expertsvideo-pretrainingembodied-aiworld-modelsRead - Jul 2026 · 7 min
Why group sampling breaks asynchronous agentic RL, and what GLM-5.2 used instead
A new paper replaces GRPO's group-wise rollout sampling with single-rollout asynchronous training plus strict double-sided token clipping, fixing the straggler and staleness problems that break RL at scale for long-horizon agentic tasks — and reports the method already trained the production GLM-5.2 model.
reinforcement-learningagentic-aillm-traininggrpoRead - Jul 2026 · 6 min
Graphify turns your codebase into a knowledge graph, not a vector index
Graphify parses a codebase locally with tree-sitter into a traversable knowledge graph, tagging every edge extracted or inferred instead of asking you to trust a ranking. I look at how it's built, what its own benchmarks show, and where a confidence-tagged graph beats embeddings for production agent workloads.
knowledge-graphscode-intelligenceai-agentsdeveloper-toolsRead - Jul 2026 · 6 min
A 1.3-billion-parameter chest x-ray generator, and the governance gap it opens
A new paper trains the largest specialist generative model for chest radiographs to date on a rectified flow transformer, producing synthetic x-rays clinical experts can't distinguish from real ones — I break down why the architecture choice matters and what fidelity this good actually demands of the teams who use it.
generative-aimedical-imagingrectified-flowfoundation-modelsRead - Jul 2026 · 5 min
ChatGPT work turns OpenAI's chatbot into a long-horizon agent
OpenAI has repositioned ChatGPT around a new mode, ChatGPT Work, that plans multi-step projects, acts across your connected apps and files, and stays on task for hours before handing back finished documents, decks, and sites. Here's what actually shipped and why the orchestration layer matters more than the model behind it.
chatgptopenaiai-agentsgpt-5-6Read - Jul 2026 · 4 min
Gemini API's managed agents get background execution, remote MCP, and credential refresh
Google shipped four production-focused updates to Managed Agents in the Gemini API — async background execution, direct remote MCP server connections, custom function calling alongside sandbox tools, and mid-session credential refresh. None of it is a new model; all of it targets what actually breaks when you run agents unattended.
gemini-apimanaged-agentsmcpagentic-aiRead - Jul 2026 · 6 min
SurfDesign: treating molecular surfaces as manifolds, not point clouds
A new protein design framework conditions generation on molecular surface geometry instead of backbone structure alone, using an SE(3)-equivariant surface encoder grafted onto a pretrained protein language model via lightweight fine-tuning — and it breaks the 70% recovery barrier that backbone-only and prior surface methods couldn't clear.
protein-designgeometric-deep-learningequivariant-networksgenerative-aiRead - Jul 2026 · 6 min
A brain foundation model that skips tokenization entirely
B[FM]² pretrains an EEG foundation model with continuous-time flow matching instead of masked patches, using a fraction of the data prior models needed and setting a new state of the art on most benchmark tasks.
eegflow-matchingfoundation-modelsgenerative-aiRead - Jul 2026 · 4 min
Anthropic details Fable 5's cyber safeguards and proposes a jailbreak severity framework
Anthropic published the four-tier classifier logic behind Fable 5's cybersecurity safeguards and, with Amazon, Microsoft, Google, and other Glasswing partners, drafted a shared industry framework for scoring jailbreak severity, alongside a new HackerOne bounty program.
anthropicclaudeai-safetycybersecurityRead - Jul 2026 · 7 min
SeKV: what a resolution-adaptive KV cache buys you, and what it costs
SeKV keeps the full KV cache alive by splitting it into a coarse GPU-resident summary and a full-detail CPU-resident SVD basis, reconstructing token detail on demand, cutting GPU memory 53% at 128K context, though the paper leaves the latency cost of that CPU round trip unmeasured.
kv-cachelong-contextllm-inferencememory-optimizationRead - Jul 2026 · 6 min
Treatment reasoning as iterative tool use: what ATHENA-R1 gets right
A new agent reframes clinical treatment reasoning as iterative evidence-gathering over 212 biomedical tools, trained with a two-level pipeline of multi-agent trajectory synthesis, supervised fine-tuning, and RL on scientific feedback — then validated against 5.4 million patient records, not just benchmarks.
agentic-aireinforcement-learningai-in-medicinellm-agentsRead - Jul 2026 · 7 min
NVIDIA's Nemotron team ships unified audio intelligence without breaking the text model underneath
Nemotron's new Audex model unifies audio understanding, speech recognition, TTS, and audio generation into one decoder without the usual collapse in reasoning ability — and the interesting part isn't the architecture, it's the training order that protects it.
audio-llmnvidianemotronmultimodalRead - Jul 2026 · 6 min
Data mixing, not filtering, is what actually improves VLM pretraining data
DataComp-VLM, a 160-dataset, 6T-token benchmark for vision-language data curation, finds that how you mix data types matters far more than how aggressively you filter them -- and that instruction-heavy mixtures pull ahead of caption-heavy ones as scale increases.
vision-language-modelsdata-curationpretrainingdatacompRead - Jul 2026 · 6 min
Inside GeneBench-Pro: a benchmark that measures research taste, not recall
OpenAI's GeneBench-Pro scores frontier models on 129 research-level genomics judgment calls, and the best model gets less than a third right — a rare, cross-lab-comparable read on scientific reasoning rather than a product launch.
genebench-proai-benchmarksgenomicsscientific-reasoningRead - Jul 2026 · 6 min
Genes as queries: the trick behind DeepSpot-M's virtual spatial transcriptomics
DeepSpot-M reframes spatial gene-expression prediction from histology as a query over gene embeddings instead of a fixed output list, letting it generalize zero-shot to genes and cancer types never seen in training.
spatial-transcriptomicsfoundation-modelscomputational-pathologyzero-shot-learningRead - Jul 2026 · 5 min
DeepMind's AI Control Roadmap treats AI agents as insider threats
Google DeepMind published its first AI Control Roadmap — a defense-in-depth plan that treats internal AI agents as potential insider threats, watched by trusted AI supervisors and measured on coverage, recall, and time-to-response. Here is what shipped and what production teams can borrow.
ai-safetyai-agentsai-controlsecurity-engineeringRead - Jul 2026 · 6 min
When the answer is a set, not a chain: set-distance rewards for radiology reports
A close read of SDR (arXiv:2606.00440), which replaces exact-match RL rewards with permutation-invariant set-to-set embedding distances for chest X-ray report generation — then reuses the same signal for best-of-N selection and mid-generation pruning that cuts tokens by half.
reinforcement-learningreward-designvision-language-modelsradiologyRead - Jul 2026 · 5 min
Fable 5 is back: inside the first regulatory pause of a frontier model
Anthropic restored Claude Fable 5 globally on July 1 after the US lifted the export controls that had suspended it for nineteen days, adding a new cyber safety classifier that reroutes blocked requests to Opus 4.8. Here is what happened, what shipped, and what the episode means for teams building on frontier models.
fable-5anthropicexport-controlsai-safetyRead - Jul 2026 · 4 min
Claude Sonnet 5: near-Opus agentic capability at Sonnet pricing
Anthropic shipped Claude Sonnet 5, the first Sonnet-tier model in the Claude 5 family: close to Opus 4.8 on agentic work, a 1M-token context window, an effort dial up to xhigh, and intro pricing of $2/$10 per MTok through August 2026. Here is what shipped, the tokenizer fine print, and what it changes for production model routing.
claude-sonnet-5anthropicagentic-aillm-pricingRead - Jul 2026 · 6 min
Judging a code patch without running it
Training a coding agent needs a referee to grade its patches, and the standard one spins up a Docker environment per repository to run the tests. A new paper judges correctness by exploring the repo instead — a fully environment-free SFT and RL pipeline that matches the execution-based version.
coding-agentsreinforcement-learningllm-verifiersswe-benchRead - Jul 2026 · 6 min
GeneBench-Pro is measurement infrastructure, and that's the point
OpenAI's new GeneBench-Pro benchmark scores whether AI agents can make the judgment calls that real computational biology demands. It isn't a new model or a capability leap — it's a carefully built ruler, and the way it's built is the interesting part.
ai-benchmarkscomputational-biologyagentic-aiscientific-aiRead - Jul 2026 · 6 min
A small model, a whole health system: what DT-Transformer gets right
DT-Transformer treats each patient's medical history as a token sequence and predicts the next diagnosis and its timing, trained on 57 million records from 1.7 million patients across an 11-hospital system. The result worth studying is not the architecture — it is the health-system-scale data and the honest prospective test.
foundation-modelshealthcare-aidisease-predictionehrRead - Jun 2026 · 5 min
Claude Science: reproducibility as a first-class artifact
Anthropic shipped Claude Science, an agentic research workbench that runs on a lab's own compute and bundles every figure with the code, environment, and history that produced it. The interesting bet isn't the chat interface — it's provenance by construction and a reviewer agent built into the loop.
claude-scienceagentic-aireproducibilityai-for-scienceRead - Jun 2026 · 5 min
Claude Sonnet 5: agentic capability moves down the cost curve
Claude Sonnet 5 brings near-Opus agentic capability to a lower price tier, collapsing the Sonnet-versus-Opus choice into a single cost-performance frontier you tune with an effort dial. A practitioner's read on what shipped, the tokenizer-driven pricing nuance, and how to route work between the two.
claude-sonnet-5agentic-aianthropicmodel-routingRead