Notes on AI, research, and the work.
234 pieces on generative and agentic systems, the research underneath, and what it takes to ship them.
- Jun 2026 · 7 min
Thinking while speaking: borrowing a frontier brain mid-sentence
A new paper introduces conversational infill, where a small on-device 'talker' starts answering instantly and weaves in knowledge streamed from a slow frontier 'reasoner' while it is still speaking — reaching within 6.3% of frontier accuracy at millisecond time-to-first-response.
voice-agentsinferencesmall-language-modelson-device-aiRead - Jun 2026 · 6 min
Agentic abstention: when an agent should stop instead of act
A new benchmark across 28,000+ tasks shows most production agents can abstain but do it too late, that more capable models often stop later, and that a training-free context method called CONVOLVE more than doubles timely stopping on WebShop.
agentic-aillm-agentsai-evaluationcontext-engineeringRead - Jun 2026 · 5 min
Privacy alignment for agents, grounded in what people actually think
PrivacyAlign places real human annotations at the center of agentic privacy — conditioning both the LLM judge and the RL reward on what 599 people actually judged appropriate. The contribution is incremental, but the conditioning trick generalizes well beyond privacy.
ai-agentsprivacyalignmentreward-modelingRead - Jun 2026 · 6 min
Autodata: when the data scientist is an agent
Meta's FAIR group built an agent that doesn't just generate synthetic training data — it tests every example against a weak and a strong model, keeps only the ones that sit in the capability gap, and then tunes its own instructions to do the job better. A look at how Autodata works and why it turns inference compute into training-data quality.
synthetic-dataagentic-aillm-fine-tuningself-instructRead - Jun 2026 · 4 min
The shift from chatbots to agents, measured from the inside
OpenAI published a year of internal Codex telemetry showing the move from chatbots to agents — longer task horizons, non-developer adoption up 137x. Here is what the data shows and where it deserves skepticism.
agentic-aiopenaicodexfuture-of-workRead - Jun 2026 · 5 min
GPT-5.6 Sol: a preview about tokens and control
OpenAI previewed the GPT-5.6 series — Sol, Terra, and Luna. The capability headline matters less than two shifts underneath it: intelligence is getting cheaper per token, and a frontier release is now gated by government review.
gpt-5-6openaiagentic-aimodel-efficiencyRead - Jun 2026 · 4 min
OpenAI's Jalapeño: a chip built for inference, not training
OpenAI and Broadcom unveiled Jalapeño, a custom accelerator built specifically for LLM inference rather than training. What shipped, why performance-per-watt is the metric that matters at scale, and why the nine-month tape-out is the most interesting part.
openaibroadcomai-inferencecustom-siliconRead - Jun 2026 · 6 min
The most expensive thing a coding agent does is read your code
AI coding agents spend most of their token budget exploring a repository file by file before they change a single line. codebase-memory-mcp indexes the whole codebase once into a persistent knowledge graph so the agent queries structure instead of reading text — and the token math behind that is worth understanding even if you never install it.
ai-coding-agentscode-intelligenceknowledge-graphmcpRead - Jun 2026 · 6 min
HyperFrames: HTML as a deterministic compile target for video
HeyGen's open-source HyperFrames turns HTML, CSS, and seekable animations into deterministic MP4 video, with coding agents as the intended author. Here is why the authoring contract — not the renderer — is the real bet, and where it fits in a production content pipeline.
hyperframesagentic-aideterministic-renderinghtml-native-videoRead - Jun 2026 · 5 min
OpenMontage and what it gets right about agentic pipelines
OpenMontage turns an AI coding assistant into an open-source video studio, but the lesson isn't the video — it's the architecture. Pipeline-constrained agents, scored and auditable provider selection, deterministic verification gates, and capability-aware degradation are patterns that generalize to any agent that ships an artifact.
agentic-aiai-pipelinesopen-sourcevideo-generationRead - Jun 2026 · 7 min
Qwen-AgentWorld: training agents against a simulated world
Qwen trains large language models to act as the environment — predicting what a shell, browser, or API would return — then uses them as a cheap, parallel simulator for agentic RL and as a warm-up for agent foundation models. A look at why modeling the environment, not the policy, is the more useful move.
agentic-aiworld-modelsreinforcement-learningllmRead - Jun 2026 · 6 min
How much should you trust an LLM to grade your agent?
The Counsel dataset meta-evaluates LLM-as-a-judge critiques of agent trajectories against human labels. Even the strongest open-weight judge agrees with humans about 88% on where an agent erred but only about 65% on why — a useful reality check before you trust an LLM to score your agents.
llm-as-a-judgeagent-evaluationmeta-evaluationagentic-aiRead - Jun 2026 · 4 min
LifeSciBench grades the work, not just the answer
OpenAI released LifeSciBench, an expert-authored, expert-reviewed benchmark of 750 real life-science research tasks graded by detailed rubrics. It measures whether AI can reason like a working scientist — surfacing caveats and limitations — not just recall biology.
ai-for-sciencebenchmarkslife-sciencesevaluationRead - Jun 2026 · 6 min
Search is not discovery: a typed account of self-revising AI
A new MIT paper separates retrieval, search, and discovery with category theory, defining discovery as a verified change of representational regime rather than subjective novelty. Here is why that typed, gated boundary is exactly what production agent systems are missing.
agentic-aiai-for-sciencecategory-theoryscientific-discoveryRead - Jun 2026 · 6 min
DiffusionGemma: 4x faster text, and where the speedup actually lands
Google DeepMind released DiffusionGemma, an open Apache-2.0 model that generates text by diffusion instead of left-to-right decoding — up to 4x faster on a single GPU. The speed is real, and so is the tradeoff; here is what shipped and where it pays off.
diffusiongemmaopen-weightstext-diffusiongemmaRead - Jun 2026 · 4 min
Gemma 4 12B and the encoder-free turn in multimodal models
Google DeepMind's Gemma 4 12B is an open-weight multimodal model that drops the separate vision and audio encoders, feeding images and raw audio straight into the language model backbone — and it runs on a 16GB laptop. A look at why the architecture matters more than the version number.
gemma-4open-weightsmultimodalencoder-freeRead - Jun 2026 · 6 min
OpenAI tuned a free model to frontier-level health answers — the method is the story
OpenAI's GPT-5.5 Instant now answers health questions on par with its frontier Thinking models, and ships to every free user. The headline is the capability; the durable lesson is the physician-in-the-loop measurement apparatus behind it.
openaichatgpthealth-aimodel-evaluationRead - Jun 2026 · 7 min
The discovery layer the agentic web was missing
Google's Agentic Resource Discovery spec adds the layer the agent ecosystem was missing: a standard way to publish, find, and verify tools and agents across organizational boundaries, anchored on domain ownership rather than a new identity authority. Here is what the design gets right and the harder problems it deliberately leaves to registries and policy.
agentic-aiagent-discoveryardmcpRead - Jun 2026 · 6 min
Self-driving datasets: when the agent out-curates the database
A system from Penn turns 22.5 million PubMed papers into structured datasets that beat hand-curated databases on scale and, more surprisingly, on audited accuracy. The lesson for anyone building agentic systems: ground the agent in structure, sweep the corpus exhaustively, and treat the gold standard as a hypothesis.
agentic-aibiomedical-mldata-curationretrievalRead - Jun 2026 · 7 min
When a language model can stand in for a human in an A/B test
Running experiments on language models instead of people only works when the model output is calibrated against real human outcomes. A recent Spotify paper sets out the exact conditions — and shows why a surrogate can be falsified on past experiments but never verified for a genuinely new one.
ab-testingllm-evaluationcausal-inferenceexperimentationRead - Jun 2026 · 6 min
Whole-body foundation models and the dominant-organ shortcut
Pan-FM is a foundation model trained across seven organs at once and built to tolerate missing scans. Its core idea, saliency-guided masking, uses the model's own attention to stop it from leaning on a few dominant organs and force balanced, whole-body learning.
foundation-modelsmedical-imagingmultimodal-learningself-supervised-learningRead - Jun 2026 · 5 min
Agentic coding rewards domain expertise, not coding skill
Anthropic analyzed roughly 400,000 Claude Code sessions and found that agentic coding does not flatten the value of expertise — it pays a premium for it. The multiplier is not coding proficiency; it is how well you understand the problem you are directing the agent to solve.
agentic-aiclaude-codedomain-expertiseai-codingRead - Jun 2026 · 6 min
GLM-5.2: a million-token coding model that deferred its own benchmarks
z.ai shipped GLM-5.2 — a 1M-token, open-weights coding model — to its Coding Plan with no benchmarks at launch. A look at what the specs, the deferred leaderboard, and the economics actually mean for agentic coding in production.
glm-5-2agentic-codingopen-weightsllm-benchmarksRead - Jun 2026 · 6 min
Unified memory and the rise of the local AI box
AMD's Ryzen AI Halo signals that local large-model inference is becoming a real product category, and the lever isn't a faster chip but 128 GB of unified memory. A look at what AMD shipped, how to read its first-party benchmarks, and when a box on your desk actually beats the cloud.
local-aiunified-memoryllm-inferenceamd-ryzen-aiRead - Jun 2026 · 7 min
A bounding box is not a sentence
Most vision-language models still emit a bounding box one coordinate token at a time — a habit borrowed from text generation that is both slow and, at high IoU, subtly wrong. LocateAnything decodes the whole box in a single parallel step, moving the speed-accuracy frontier in both directions at once.
vision-language-modelsvisual-groundingobject-detectionparallel-decodingRead - Jun 2026 · 6 min
Agentic tool search, made right
Agentic search across many walled-garden sources works when each source is treated as its own authenticated tool, every result carries a real engagement signal, and an agent judge synthesizes the normalized results with citations. I walk through that architecture — fan-out, signal normalization, judged synthesis — and where it breaks.
agentic-aiai-agentssearchtool-useRead - Jun 2026 · 6 min
Context compression: cutting the tokens your agents re-pay for
In an agent loop the model is a fixed cost, but the context isn't — every tool output and retrieved chunk gets re-billed on each step. Context compression shrinks that traffic by 47–92% with type-aware, reversible methods; here's how it works and where it pays off.
context-compressionllm-agentstoken-optimizationinference-costRead - Jun 2026 · 4 min
From prompt engineering to loop engineering
The people who built Claude Code have stopped writing prompts and started writing loops. Here is what loop engineering actually is, why it behaves like closed-loop control, and where the engineering effort really goes.
loop-engineeringagentic-aiclaude-codeprompt-engineeringRead - Apr 2026 · 8 min
The Feynman Technique Meets AI: Learning Through Teaching Machines
The Feynman Technique—learning by teaching—takes on new meaning in the AI era. Named after physicist Richard Feynman's approach of explaining complex concepts in simple terms, this methodology offers profound insights for AI development, model training, and human-AI interaction. As we build increasingly sophisticated AI systems, the principle of 'if you can't explain it simply, you don't understand it well enough' becomes crucial for creating interpretable, reliable, and effective AI solutions. This exploration examines how Feynman's teaching philosophy can revolutionize our approach to AI development and deployment.
explainable-aimachine-learningai-philosophymodel-interpretabilityRead - Apr 2026 · 8 min
From SR 11-7 to AI Governance: Why Traditional Model Risk Frameworks Are Breaking
For over a decade, SR 11-7 has been the gold standard for model risk management in financial services. But the rise of AI and Large Language Models is fundamentally breaking traditional frameworks. Unlike deterministic models with predictable Input → Process → Output flows, AI creates risk through entire systems characterized by non-determinism, continuous learning, and emergent behaviors. The industry must pivot from model-centric to system-centric governance, embedding controls across the entire Data → Model → Output → Decision chain. This shift requires rethinking validation from point-in-time events to continuous monitoring, addressing new risks like prompt injection and hallucination, and adopting lifecycle-based approaches that make decisions defensible rather than just models valid.
AI GovernanceModel Risk ManagementSR 11-7Financial ServicesRead