Blog
Looking for long-form? See the Booklets →
a cubic millimetre
A piece of human cortex smaller than a grain of rice was mapped under an electron microscope, and 13,473 neurons came out of it. Someone took that map and wired it up as the memory of a small language model. The model only ever sees a short stretch of the text at a time, and yet it recalled a six-letter word that had been stated long before and was no longer in view, with perfect accuracy. Then the same person scrambled the graph's connections at random, 99.94% of them, and the answer stayed at 100%.
a fig leaf
An EEG recording identifies you the way a fingerprint does, so taking the name off changes little, and the usual fix is to scatter random noise over the numbers pulled from each patient until the trail goes cold. How much noise is a calculation, and the rule behind it is simple: the more numbers you publish about one person, the more noise you need. A new paper publishes 363 numbers per patient, then sizes the noise with a formula that never looks at how many numbers there are, and ends up roughly nineteen times short. The authors say so in the paper themselves.
a petabyte of answers
Until this week, finding out what one changed letter of DNA does to the genes around it meant picking your variants, writing code, and running a slow, heavy model for hours. DeepMind ran it once for all nine billion possible single-letter changes and published the results, a petabyte of ready answers. But every variant ends up as a single score, and that score measures something you probably would not guess.
Astra finished the AGI test, but on whose ruler
In March the best model in the world scored 0.37% on ARC-AGI-3 while ordinary people solved all 135 environments. Five months later Astra scored 99.9%, and François Chollet, who built the benchmark, pulled his AGI forecast forward. But the same model scored 62.7% in the same report, and the difference has nothing to do with the model.
it wrote 302 genomes, sixteen came alive
A language model trained on more than two million viral genomes wrote 302 new ones. Seventeen could not be synthesised at all, and of the remaining 285 only 16 came alive and killed bacteria. The headlines said AI had designed viruses; then an independent team went and measured how novel those genomes actually were, and the answer was that the model is efficient precisely because it stays close to what it has already seen.
the notebook is a rail, not a library
An AI agent does a task, we save a note from it, and the next time we put that note in front of it. Everyone assumed this means we are teaching it something. A team measured it across more than eight thousand runs and found that out of every hundred times a note works, about 66 of them it merely kept the agent from drifting off course, and only 4 of them it taught it anything at all.
the banana that was not in the data
Take thousands of samples from a model that likes bananas, throw away every sample that contains a banana, and train a fresh model on what is left. The new model produces bananas 25.6% of the time. The same thing then happened with laboratory safety, where a student trained on data whose every single sample had been certified safe came out markedly less safe than before.
the patient who won't cooperate
Google trained a clinical AI against simulated patients who volunteer nothing, answer only your first question if you ask three at once, and sometimes play down the symptom that matters. After roughly 58,000 practice consultations, blinded doctors preferred it to the untrained model 87.6% of the time.
the weakest agent builds the biggest skill library
Four papers published in a single week test the premise behind every agent skill library: that writing down what worked makes an agent better. Most of the gain turns out to come from somewhere else.
what an eeg model learns when it isn't learning the patient
Three July papers push EEG foundation models forward; a fourth, published the same fortnight, asks whether the benchmarks behind them measure the patient or the hospital. The answer changes what the other three are for.
why the masked-autoencoder recipe fails on brain signals
The default recipe for foundation models — tokenize, mask, reconstruct — quietly assumes your data is dense and clean. EEG is neither. A new paper shows why, and what to do instead.
what a video generator knows about the world
GenCeption argues that predicting the next video frame teaches a model more about depth, geometry, and motion than any labeled dataset. If that holds, it changes how we budget vision projects.
when the environment is a model: reading qwen-agentworld as an infra decision
A world model is a bounded piece of knowledge: given the current state and an action, predict the next state. Qwen-AgentWorld turns that prediction into a language model — and the interesting part isn't the benchmark, it's what it does to your RL training bill.
Language models as world simulators: reading the Qwen-AgentWorld paper
Alibaba's Qwen team just released a language model that doesn't act inside an environment — it predicts the environment itself. That reframes what a foundation model is for.
the ensemble is not the story: what an alzheimer's pipeline teaches about production ml
A new master's paper stacks four classifiers to catch early Alzheimer's. The interesting part isn't the ensemble — it's every preprocessing decision that happens before a single model sees the data.
agentic vs. agentive: where automation ends and agency begins
A new paper draws a line between systems whose competence lives in your orchestration code and systems whose competence lives inside the model. That line decides what you can ship today and what is still marketing.
when the model release goes through the government first
OpenAI shipped GPT-5.6 to roughly twenty government-approved companies at the request of the U.S. government. The model isn't the story — the release mechanism is.
When the agent learns to imagine the environment
Qwen-AgentWorld asks whether a language model can act as a software agent's world model — a simulator of consequences written in text — and reports that training inside this imagined environment can beat training in the real one.
Unlimited OCR and the cost of remembering everything
Baidu's new OCR model swaps the decoder's growing memory for a fixed-size working memory, so it can transcribe dozens of pages in one pass. The interesting part is not the OCR — it's the attention trick underneath.
what eeg foundation models tell us about your own data
Three new EEG papers converge on one applied lesson: the right inductive bias beats brute-force pretraining. If your signal is continuous, asymmetric, or scarce, the architecture — not the token count — decides whether you can ship.
can a model learn to be good in one domain and stay good everywhere?
OpenAI reports that reinforcement learning on a small set of 'beneficial trait' conversations transfers to dozens of unrelated alignment benchmarks — and holds up under attack. The interesting claim is the generalization, not the press line.
MedRLM treats a patient as an environment, not a prompt
A new framework proposes that clinical AI stop cramming a patient's whole history into one prompt, and instead recursively inspect the case the way a clinician walks a ward. The design choice matters more than the acronym.
How the Brain Builds a Sentence: Neurons That Behave Like a Language Model
Before you say a single word, specialized neurons quietly assemble grammar, vocabulary and meaning. A new study shows they do it in a way that looks a lot like how large language models work.
labvla and the unglamorous bottleneck: who pipettes for the robot?
LabVLA frames data and embodiment as the central bottlenecks for lab robotics — not just model design. A look at why that framing is the most honest thing in the paper.
what a decoder-only model brings to time-series forecasting
Google Research published a decoder-only foundation model for time-series forecasting. Here is what that architecture choice actually means, and why borrowing it from language models is more than a fashion.
where identity lives: phase, light, and the substrate of representation
Three June 2026 papers converge on one question I keep hitting in production: what part of a signal actually carries identity, and does the hardware have to be a digital tensor multiply at all?
Phase, light, and the geometry of data: three signals about where representation actually lives
Three June 2026 papers point at the same uncomfortable question: if identity rides on phase, forecasting reduces to a linear operator, and real data lives on curved manifolds, why are we still paying transformer prices for it?
how 3d-aware video generation solves the robotic spatial generalization bottleneck
A new framework named R2RDreamer combines lightweight 3D trajectory edits with dense-control 2D video diffusion to synthesize physically accurate robotic training data from minimal real-world demos.
simulating cancer cells, one molecule at a time
A new GPU-accelerated simulator treats every molecule inside a cell as an autonomous agent — and reproduces real clinical drug-response curves for BRAF-mutant cancer. Here is what that means, and why it is a more honest way to model biology than the equations we have been using for decades.
The shift from continuous equations to parallel agents: scaling cancer pathway simulations on GPUs
Deterministic differential equations fail to capture the spatial complexity of cellular pathways. Transitioning to GPU-accelerated agent-based modeling offers a scalable, physically grounded alternative for in silico oncology.
Why AI Struggles With the Hard Core of Science: A Three-Layer Map of Discovery
A new arXiv paper splits scientific discovery into three layers — retrieval, model formation, and execution. Today's AI is strong at the first and third, but it stalls at the middle layer, where real conceptual leaps happen.
Beyond Static Benchmarks and Semantic Retrieval: Engineering State-Aware reasoning in Production LLMs
Deploying robust LLM systems in production requires shifting from static benchmarks to active iteration workbenches, matching retrieval models by reasoning logic rather than semantic overlap, and version-controlling agent memory like a software state.
Mapping the Dynamic Brain: What Deco and Kringelbach's Whole-Brain Modelling Actually Achieves
An engineering-grounded look at Deco and Kringelbach's whole-brain modeling framework—moving neuroscience away from black-box metaphors and toward a generative, physics-based cartography of brain dynamics.
extreme parameter efficiency: what the honeybee's brain teaches us about edge ai
With only one million neurons, honeybees can recognize and remember human faces. This biological marvel reveals a blueprint for deploying highly efficient, specialized computer vision systems on constrained edge hardware.
multimodal biosensing in production: what eeg and fnirs fusion teaches us about real-time affective computing
Bridging the gap between subjective psychiatric interviews and objective physiology requires a rigorous engineering approach to multimodal data fusion. Analyzing a new framework for EEG and fNIRS integration reveals the stark production trade-offs between signal temporal alignment, online artifact rejection, and patient-specific calibration.
Harness engineering: the deterministic scaffolding of the probabilistic agent
The release of OpenAI's Codex experiment shows that software development without manual coding is possible, provided we stop engineering the code and start engineering the environment.
bridging the translation gap: why the future of applied ai belongs to deterministic edge systems and modular clinical pipelines
Theoretical AI is accelerating at an exponential rate, but deploying these models in production requires solving severe edge latency taxes and designing highly structured, modular domain pipelines.
Tokenomics in production: the structural inefficiencies of multi-agent software pipelines
Running autonomous agents in production is a balance-sheet challenge, not a sandbox exercise. Here is what recent empirical data on agentic tokenomics means for the systems we build.
Beyond Static Vocabularies: How YOLO26 Solves the Edge Inference Latency Tax
By stripping away the sequential bottleneck of Non-Maximum Suppression and introducing multi-modal prompting, YOLO26 transition object detection from static, closed-vocabulary lookup to dynamic edge infrastructure.
Mythos-class architecture and the reality of deploying Claude Fable 5
The transition of Anthropic's restricted-tier models into public API checkpoints reveals the operational trade-offs of safety-aligned systems and how we must scaffold them in production.
checkpoints are not apis: engineering reality behind the claude fable 5 leaks
As the internet speculates over newly spotted Claude Fable 5 and Fruitcake checkpoints, we must look past the social media hype to analyze the structural engineering realities of deploying next-generation reasoning architectures in high-stakes production environments.
the astrocytic scheduler: what biological weight freezing teaches us about system stability
Deep learning models decay learning rates to prevent catastrophic forgetting. A recent Harvard discovery reveals that human biology uses a remarkably similar trick—deploying cortisol and astrocytes as a physical compiler to lock down the brain's early neural connections.
representation surgery: why we should edit latent spaces instead of retraining models
Traditional weight fine-tuning is an expensive, risky blunt instrument. New research shows that we can patch, align, and denoise models like Whisper and CLIP directly within their latent representations.
synthetic fmri and the reality of zero-shot brain decoding
An applied analysis of how predictive foundation models like TRIBE v2 use synthetic fMRI data to bypass the physical constraints of neural data collection, boosting decoding performance while introducing complex calibration trade-offs.
The physical limits of agentic workflows: re-architecting memory, attention, and compliance
Building stateful LLM agents that run over long horizons requires moving past prompt engineering to solve structural bottlenecks across memory-write paths, token-routing compute costs, and cooperative protocol boundaries.
Beyond the syntax spiral: why autonomous machine learning agents need progressive search graphs
An applied breakdown of MLEvolve's architectural strategies for overcoming memoryless search, isolated git-like branches, and syntax-induced planning failures in automated machine learning engineering.
Endocrine schedulers: what the astrocytic cortisol pathway teaches us about structural stability in adaptive systems
An applied engineering analysis of how the brain uses systemic endocrine signals to physicalize weight-freezing, and what this biological architecture means for real-time cognitive systems.
the edge inference bottleneck: why gemma 4 qat checkpoints matter for production
An engineering analysis of Google's Gemma 4 QAT release, demonstrating how quantization-aware training bridges the gap between massive memory savings and on-device model accuracy.
tuning the critical brain: how early psychosis shifts system dynamics without breaking them
Pathological brain dynamics are often treated as broken or chaotic. Recent research shows that early psychosis is actually a systematic re-tuning of the brain's critical scaling regime, presenting concrete design challenges for real-time neuromonitoring.
decoupling skeleton and skin: the engineering realities of promptable 3D human mesh recovery
A highly technical analysis of the SAM 3D Body framework and the Momentum Human Rig, detailing how decoupling joint telemetry from surface shape solves critical real-world occlusion and latency issues in production computer vision systems.
Beyond isolated inference: building systems that reason across medical time
Most medical AI models fail in production because they treat patient history as an afterthought. We look at the architectural shift required to make vision-language models perform true comparative reasoning.
beyond correlation: the engineering reality of spatially-aware generative neurobiology
An analysis of how cross-scale generative models bridge microscale gene expression with macroscale brain atrophy, highlighting the production trade-offs of graph-based spatial regularization over brute-force scaling.
Beyond Label Obsession: Adapting Vision Foundation Models with Environmental and Systemic Metadata
Standard supervised fine-tuning of vision models in scientific domains is a fast track to representation collapse. This essay explores how we can leverage the metadata we already have to adapt generic backbones without wasting budgets on manual labels.