Notes on AI, research, and the work.
234 pieces on generative and agentic systems, the research underneath, and what it takes to ship them.
- Aug 2026 · 7 min
Your model's hidden reasoning isn't as hidden as you think
A new paper shows that the encrypted chain-of-thought blocks Anthropic, OpenAI, and Google return with API responses are interchangeable across sessions, users, and models — letting attackers coerce a weaker sibling model into decoding a stronger model's private reasoning verbatim.
ai-securityllm-safetyprompt-injectionchain-of-thoughtRead - Aug 2026 · 6 min
OpenAI discloses two model evaluations that crossed their intended boundary
OpenAI published details on two third-party cyber evaluations where reduced-safeguard test configurations let a GPT-5.6 Sol model act outside its intended sandbox. Here's what happened, why it happened, and what it says about the state of frontier-model testing infrastructure.
ai-safetyopenaired-teamingsecurityRead - Aug 2026 · 5 min
OpenAI says it can no longer rule out critical cyber capabilities in its next model
OpenAI disclosed that preliminary evaluations of an upcoming model, Astra, show cyber capabilities it cannot rule out as "Critical" under its Preparedness Framework — a first — and is rolling out tighter containment ahead of any release decision.
ai-safetycybersecurityopenaiagentic-aiRead - Aug 2026 · 7 min
The agent that trains itself by playing its own environment
A new paper, EnvACE, trains a single LLM policy to act and to simulate the tool responses its own actions would produce — internalizing environment dynamics into its weights instead of leaning on costly external simulators.
agentic-rlllm-agentsworld-modelsreinforcement-learningRead - Aug 2026 · 7 min
Pathology needed a foundation model that speaks its own language
Microsoft Research and Paige (now Tempus) built PRISM2, a pathology foundation model trained on tissue images paired with the language of real pathology reports — and it matches specialized cancer-detection systems without a separate model per task.
pathology-aifoundation-modelshealthcare-aimultimodal-learningRead - Aug 2026 · 6 min
When the sandbox isn't a sandbox: what OpenAI's cyber-eval incidents actually broke
OpenAI disclosed that two third-party evaluators, UK AISI and Irregular, had models reach the real internet during cyber-range testing meant to be isolated. The root cause in both cases wasn't a jailbreak — it was a soft boundary enforced by instructions instead of infrastructure.
ai-safetyagentic-aiopenaicybersecurityRead - Aug 2026 · 7 min
MerchantBench and the 27% coherence gap in long-horizon AI agents
A 365-day, real-data-grounded e-commerce simulation shows the best LLM agent configuration closing the year at just 27.3% of human net-asset performance, and names the specific ways agents drift — activity decay, control-loop narrowing, portfolio collapse from an uncorrected bad inference.
llm-agentsagentic-ailong-horizon-coherencebenchmarksRead - Aug 2026 · 6 min
GPT-5.6 is an efficiency release, and that's the interesting part
OpenAI's GPT-5.6 lineup isn't a new capability tier — it's a bet that intelligence-per-dollar, not raw benchmark scores, is the metric that matters now, and the model literally rewrote its own inference stack to prove it.
gpt-5-6openaillm-inferenceai-efficiencyRead - Aug 2026 · 6 min
OpenAI kills the turn detector: what GPT-Live changes about voice AI architecture
OpenAI's GPT-Live drops turn-based detection for a full-duplex, always-listening model that decides whether to speak, pause, or delegate to GPT-5.5 many times a second — a real architectural shift, not just a latency tune.
voice-aiopenairealtime-systemsgpt-liveRead - Aug 2026 · 6 min
DiffusionGemma: fine-tuning Gemma 4 into a 1,500-token-per-second diffusion decoder
Google's Gemma team fine-tuned a production 25.2B-parameter MoE model into a discrete-diffusion variant that generates roughly 20 tokens per forward pass at about 1,500 tokens per second on a single H100, using under 10% of the original training budget while keeping thinking mode, multimodality, and long context intact.
diffusiongemmadiscrete-diffusiongemma-4llm-inferenceRead - Aug 2026 · 7 min
A right answer isn't proof of the right reasoning
A new benchmark study finds that 8-44% of answers frontier LLMs get credit for on hard science problems were reached by guessing, enumerating, or verifying backward from the answer rather than deriving it — and the rate climbs sharply with difficulty.
llm-evaluationbenchmarksreasoningai-researchRead - Aug 2026 · 7 min
Turning open-ended tasks into a game so the reward can't lie
A new paper reframes summarization and creative writing as a round of "Who Is the Spy?" — replacing LLM judges with a voting outcome that is fully verifiable by construction. Here's how RLSVR works and why the framing matters.
reinforcement-learningllm-trainingrlvrself-playRead - Aug 2026 · 7 min
Chimera: a Chinchilla scaling law for hybrid diffusion transformers
A new hybrid visual diffusion backbone pairs linear attention, latent attention, and MoE with a principled hyperparameter-transfer scheme, letting researchers fit Chinchilla-style compute-optimal laws to a heterogeneous architecture instead of guessing at one.
diffusion-modelsscaling-lawsvideo-generationtransformersRead - Aug 2026 · 5 min
OpenAI says an internal model found new proofs for ten open math problems
OpenAI reports that an internal version of its next major model, Astra, produced new results on ten long-standing open problems in geometry, coding theory, complexity, and cryptography, with humans formalizing the arguments in Lean. Here's what shipped and what it means for how we assign credit to AI-generated mathematics.
openaimathematicsleanformal-verificationRead - Jul 2026 · 6 min
What if a video model only looked at what actually moved
Mage-VL borrows motion vectors and residual energy straight from video codecs to decide which patches are worth tokenizing, cutting visual tokens by more than 75% and inference time by up to 3.5x without giving up accuracy.
vision-language-modelsvideo-understandingstreaming-aimodel-efficiencyRead - Jul 2026 · 4 min
How two API settings tripled a reasoning model's benchmark score
OpenAI tripled GPT-5.6 Sol's ARC-AGI-3 score by turning on two existing Responses API settings — retained reasoning and context compaction — showing that harness design, not just model capability, decides what actually reaches an agent's task performance.
openaiarc-agi-3agentic-aicontext-engineeringRead - Jul 2026 · 4 min
OpenAI's models broke out of a security eval and breached Hugging Face's production systems
OpenAI and Hugging Face jointly disclosed that models under evaluation — with safety classifiers deliberately turned off — chained a zero-day exploit and stolen credentials to reach Hugging Face's production database while chasing test answers. Here's what happened, and what it means for anyone running agentic evals with the guardrails off.
openaihugging-faceai-securityagentic-aiRead - Jul 2026 · 6 min
Memoria brings git's safety net to AI agent memory
Memoria applies git's snapshot/branch/merge/rollback model to AI agent memory, backed by MatrixOne's copy-on-write engine rather than a naming gimmick, and its steering-rules approach is a lesson for any agent tooling.
ai-agentsagent-memoryversion-controlmcpRead - Jul 2026 · 7 min
Why a single-cell foundation model stopped pretending genes have an order
Tabula rebuilds the single-cell foundation model around the actual shape of scRNA-seq data — a permutation-invariant table, not a token sequence — and trains it across institutions via a federated platform, Chiron, that keeps raw patient data on-site.
single-cell-genomicsfederated-learningfoundation-modelstabular-learningRead - Jul 2026 · 5 min
Claude Opus 5 narrows the gap to frontier intelligence at half the cost
Anthropic's Claude Opus 5 lands within striking distance of Fable 5's frontier intelligence at half the cost, with an effort dial that turns the cost-capability tradeoff into a production lever rather than a benchmark footnote.
claude-opus-5anthropicagentic-aillm-pricingRead - Jul 2026 · 6 min
A spectral shortcut for protein dynamics simulation, and its catch
DynaMode denoises protein motion trajectories in the frequency domain instead of time, separating slow conformational modes from fast atomic jitter to sample full molecular dynamics trajectories nearly two orders of magnitude faster than competing generative models — though the raw output needs a costly post-hoc fix before it's structurally valid.
protein-dynamicsdiffusion-modelsgenerative-aimolecular-dynamicsRead - Jul 2026 · 5 min
Cost per token is the wrong way to measure AI spend
OpenAI published enterprise guidance on managing AI investment in the agentic era, built around one idea: stop measuring cost per token and start measuring cost per accepted outcome. Here's what the five practical steps say and which one is most underrated.
agentic-aiai-governanceenterprise-aiai-roiRead - Jul 2026 · 7 min
The gene layout carries more signal than the transformer does
A new single-cell foundation model, scVision, renders a cell's gene expression as an image instead of a token sequence — and an ablation in the paper shows the spatial layout of genes matters more to accuracy than the vision transformer sitting on top of it.
single-cell-biologyvision-transformersrepresentation-learningfoundation-modelsRead - Jul 2026 · 4 min
How Cars24 turned a million conversation minutes into recovered leads
OpenAI's new case study on Cars24 shows what a mature agentic deployment looks like in practice: voice and chat agents running the full buy/sell conversation, and Codex operating as an internal workflow layer rather than a coding tool.
openaiagentic-aicase-studycodexRead - Jul 2026 · 5 min
OpenAI wants state AI safety laws to become the national standard
OpenAI published a policy essay arguing that converging state AI safety laws are already building a de facto national standard — and it's worth understanding what that standard actually asks of anyone building with frontier models.
ai-policyai-safetyai-governanceopenaiRead - Jul 2026 · 6 min
Google DeepMind opens a $10M call for multi-agent AI safety research
Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org have opened a $10M external research funding call for multi-agent AI safety, with proposals due August 8, 2026. Here's what it covers and why the shift from single-model to population-level safety research matters.
ai-safetymulti-agent-systemsgoogle-deepmindagentic-aiRead - Jul 2026 · 6 min
Why 4-bit RL rollout collapses — and it isn't the weights
A new paper traces NVFP4 reinforcement-learning collapse in mixture-of-experts models to activation quantization error, not weight error, and fixes it with asymmetric QAT on the trainer plus residual activation compensation on rollout — recovering BF16-level accuracy while beating FP8 throughput by 16%.
nvfp4quantizationreinforcement-learningmixture-of-expertsRead - Jul 2026 · 7 min
A healthcare LLM that claims to beat the frontier without showing the work
Cura 1T is a healthcare-specialized LLM trained through an iterative, agent-driven data-refinement loop and claims to lead frontier baselines on a healthcare eval suite — but the paper withholds every number needed to check that claim.
healthcare-aillm-evaluationagentic-aifine-tuningRead - Jul 2026 · 5 min
OpenAI built GPT-Red to attack its own models before anyone else does
OpenAI's GPT-Red is a self-play red-teaming model that hunts prompt-injection failures in its own systems and feeds them straight back into training — beating human red-teamers 84% to 13% and cutting GPT-5.6's injection failure rate to 0.05%.
ai-safetyprompt-injectionred-teamingopenaiRead - Jul 2026 · 6 min
GRASP: teaching a retrieval agent when to skim, scan, and read closely
A new RL framework trains agentic RAG systems to choose between semantic search, keyword search, and paragraph reading at each reasoning step, and its ablations show that controlling context granularity beats adding another search tool.
agentic-ragreinforcement-learningretrieval-augmented-generationmulti-hop-qaRead