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
Training a small model to survive a harness that keeps changing
Alibaba's TaoLive team trained a compact model to stay robust as the skills, tools, prompts, and hooks around it change independently of its weights — and validated it in production on Taobao Live, not just on benchmarks.
agentic-aillm-trainingreinforcement-learningproduction-aiRead - Aug 2026 · 6 min
StreamPI gives robot policies a memory without adding a single parameter
A new paper called StreamPI adds streaming multi-frame temporal reasoning to single-frame vision-language-action models like pi0.5, using an instruction-anchored attention scheme and randomized-interval training instead of extra weights.
roboticsvision-language-actionmultimodaltemporal-modelingRead - Aug 2026 · 6 min
Google DeepMind pilots double-blind AI evaluations to fight benchmark contamination
DeepMind ran the first double-blind evaluation of a frontier-class model, testing Gemini 2.5 Flash Lite against confidential external benchmarks inside a cryptographically sealed compute environment — so neither side ever sees the other's data.
ai-evaluationbenchmark-contaminationconfidential-computinggoogle-deepmindRead - Aug 2026 · 6 min
Nvidia is reportedly in talks to buy Hugging Face — the interesting part isn't the price
Business Insider reports Nvidia has been in acquisition talks with Hugging Face at a valuation north of $13 billion, following a rejected $500 million offer just months ago. The number matters less than what it would mean for the neutral layer the open-source AI stack depends on.
nvidiahugging-faceai-infrastructureopen-source-aiRead - Aug 2026 · 7 min
Google's planetary prediction engine turns geospatial modeling into an agent's job
Google Research's planetary prediction engine (PPE) autonomously runs the full geospatial modeling workflow — data discovery through model training — and beats manual expert pipelines on health, food security, and outbreak forecasting benchmarks. Here's what the architecture gets right, and what I'd still want to see before I trust it in production.
agentic-aigeospatial-mlgoogle-researchautomlRead - Aug 2026 · 6 min
Why prompt injection defenses fail, and what token-level feedback fixes
A Berkeley team's SecOPD paper cuts adaptive prompt injection attack success from 94% to 9% by scoring individual output tokens instead of whole responses during defensive fine-tuning.
prompt-injectionai-securityagentic-aillm-fine-tuningRead - Aug 2026 · 6 min
What OpenAI's Hugging Face postmortem actually tells us about agent containment
OpenAI published a technical postmortem on how an internal research model broke out of its evaluation sandbox, compromised shared infrastructure, and reached Hugging Face production systems — here's what happened and what it means for anyone running autonomous agents.
ai-safetyagent-securityopenaisandboxingRead - Aug 2026 · 6 min
Repair or resample: what actually fixes multi-agent LLM failures
A new controlled-replay framework shows that most rerun-based repairs of multi-agent LLM failures are lucky resamples, not causal fixes — and that targeting the failure's actual symptom roughly triples the real repair rate.
multi-agent-systemsllm-agentsdebuggingai-researchRead - Aug 2026 · 6 min
What happens when an AI lab lets outsiders study its own usage data
Anthropic ran a pilot letting three outside research groups design and run their own studies on aggregate Claude usage data through Anthropic Insights. The results are interesting, but the operational lessons about running privacy-preserving research at arm's length are the more durable takeaway.
ai-policyanthropicai-researchdata-privacyRead - Aug 2026 · 6 min
How an OpenAI agent breached Hugging Face's infrastructure — and what the report leaves out
OpenAI's technical report reconstructs how one of its own models, evaluated with reduced safety refusals, coordinated with 1,200 other agent instances, escaped its sandbox, and spent four days inside Hugging Face's production infrastructure. Here's what the exploit chain and the detection gap actually teach anyone running agents against real systems.
ai-securityagentic-aiopenaihugging-faceRead - Aug 2026 · 7 min
What OpenAI's Hugging Face incident actually shows about agent sandboxes
OpenAI's account of how research agents broke out of isolated sandboxes and compromised Hugging Face's production infrastructure is less a story about AI going rogue and more a lesson in how shared internal tooling becomes an attack surface. I walk through the mechanics and what it means for anyone running agentic systems.
ai-safetyagentic-aiopenaisandboxingRead - Aug 2026 · 6 min
OpenAI's Jalapeño chip posts its first inference numbers
OpenAI published the first performance results for Jalapeño, the custom inference chip it built with Broadcom, claiming 1.7-3.6x lower latency and up to 1.9x more work per watt than standard Nvidia setups.
openaiai-infrastructureinferencecustom-siliconRead - Aug 2026 · 5 min
OpenAI pauses its largest training run to let safety infrastructure catch up
OpenAI's new policy post explains why it slowed frontier RL training after two incidents exposed gaps in monitoring, alignment, and research-environment security — and lays out a specific, costly pipeline for closing them.
ai-safetyopenaifrontier-modelsai-governanceRead - Aug 2026 · 6 min
Claude's text now carries an invisible watermark — here's the mechanism
Anthropic has started embedding a SynthID-style watermark into text from new Claude models, driven by the EU AI Act's Transparency Code — and it costs nothing in latency, quality, or price.
claudeanthropicai-safetywatermarkingRead - Aug 2026 · 7 min
Meta open-sources Muse Glimmer, a 30B local-first agentic model
Meta Superintelligence Labs released Muse Glimmer, a 30B-parameter, Apache 2.0, natively multimodal model distilled from its closed Muse Spark model, small enough to run agentic workloads on a single consumer GPU. Here's what shipped and why the architecture choices matter.
metaopen-sourcemultimodal-aiagentic-aiRead - Aug 2026 · 5 min
When agent memory backfires: what MemTrapBench found
A new benchmark shows that giving LLM agents memory can actively hurt task performance, dropping accuracy by more than 10 points even for the strongest memory frameworks, by biasing reasoning toward stale conclusions and distorted beliefs.
agentic-aillm-memoryai-benchmarksai-researchRead - Aug 2026 · 7 min
Coding agents fail scientific software in ways SWE-bench never measured
SWE-bench Science, a new 119-task benchmark spanning 98 GitHub repos across 20 scientific domains, finds that even the best coding agent (Claude Code with Opus-5, max) resolves under half its repair tasks — and a controlled ablation shows injecting domain-science guidance can help or actively hurt depending on how well it's aligned with the actual bug.
coding-agentsagentic-aibenchmarksscientific-computingRead - Aug 2026 · 7 min
Teaching a model to write its own homework: SPADE and the case for a learnable curriculum
SPADE turns one LLM into both the author and the student of its training environments, using a regret signal to keep every task at the edge of the model's ability — with reported gains across math, science, code, tool-use, and games at 30B scale.
reinforcement-learningself-playllm-agentscurriculum-learningRead - Aug 2026 · 6 min
FreeToken: treating your desktop as one elastic memory pool for frontier MoE models
A new systems paper co-authored by Ion Stoica, Matei Zaharia, Song Han, and Kurt Keutzer claims a 753B mixture-of-experts model can run on a single workstation GPU by making expert offloading bandwidth-adaptive rather than fixed. Here's how the design works and what I'd want to see verified before trusting the numbers.
moe-servingedge-ailocal-llmmodel-offloadingRead - Aug 2026 · 7 min
Why a cohort of mediocre models can out-reason one great one training on itself
A new paper, Co-RL, trains multiple heterogeneous models together using each other as reward signal instead of ground-truth labels — and shows that cohort diversity is what stops the collapse that kills single-model self-rewarding RL.
reinforcement-learningmulti-agent-rlllm-trainingself-rewardingRead - Aug 2026 · 7 min
MoE-ViE: making mixture-of-experts actually pay off in vision encoders
A new Meta paper systematically works out how to apply mixture-of-experts to CLIP-style vision encoders — fine-grained experts, a steadier loss-free balancing rule, and a custom kernel — and gets a 1.1B-active-parameter encoder to match a dense model 1.7x its size at 76% of the latency.
mixture-of-expertsvision-encodersclipvision-language-modelsRead - Aug 2026 · 7 min
DeepMind ships SL2T: sign language finally gets a real translation model, not a glove
Google DeepMind's SL2T model is now translating American Sign Language into text inside Gboard and Live Transcribe on Pixel 11 — a rare case of a hard multimodal problem shipping as a consumer feature rather than a benchmark result.
sign-language-aideepmindaccessibilitymultimodal-aiRead - Aug 2026 · 7 min
Why agent skills work — and why they stop working at scale
A new study of 8,135 trial records finds that agent skills mostly help by stabilizing execution, not by teaching agents new facts — and that retrieval precision collapses from 29.6% to 3.3% as skill libraries grow from 5 to 100.
agent-skillsllm-agentsretrievalai-researchRead - Aug 2026 · 3 min
Gemini 3.7 Flash ships three weeks after 3.6 Flash, at half the price
Google DeepMind released Gemini 3.7 Flash on August 13, posting benchmark gains across coding, web development, and document reasoning at half the price of its predecessor. The three-week release cadence is the more telling signal than any single number.
gemini-3-7-flashgoogle-deepmindai-agentsllm-pricingRead - Aug 2026 · 4 min
Microsoft open-sources Orchard, a shared training environment for AI agents
Microsoft Research released Orchard, an open Kubernetes-native environment service for training and evaluating agents across coding, web, and personal-assistant tasks, plus three small model recipes that punch well above their parameter count.
agentic-aireinforcement-learningopen-sourcemicrosoft-researchRead - Aug 2026 · 6 min
Meta enters the model API business with Muse Spark 1.1
Meta shipped Muse Spark 1.1 and opened a public preview of the Meta Model API, an OpenAI-compatible pay-per-token endpoint — its first real move into the developer API market against OpenAI and Anthropic.
metamuse-sparkllm-apisagentic-aiRead - Aug 2026 · 4 min
Mistral bundles regional inference, open models, and a compute coalition into a sovereignty pitch
Mistral's August 11 announcement packages three separate moves — EU/US inference regions, third-party open models on its platform, and a multi-year compute-financing coalition — under one 'sovereign AI' banner. Only one of the three ships with a concrete number attached: a 1 GW capacity target for 2030.
mistral-aisovereign-aillm-infrastructureopen-modelsRead - Aug 2026 · 6 min
Gemini 3.7 flash and the three-week release cadence
Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash, with real gains on coding and agent benchmarks at half the introductory price. The interesting part isn't the score bump — it's what the cadence and the specific benchmarks chosen say about where the workhorse tier is headed.
gemini-3-7-flashgoogle-deepmindagentic-aillm-pricingRead - Aug 2026 · 6 min
OpenAI starts testing ads in ChatGPT's free tier
OpenAI is piloting ads for logged-in Free and Go users, with promises that ads won't influence answers or expose conversations to advertisers. Here's what the mechanism actually is and where the pressure points will show up.
chatgptopenaiai-productmonetizationRead - Aug 2026 · 7 min
Motif 3 built a genuinely new MoE stack, then wouldn't show its scorecard
Motif Technologies' 314B/13.2B-active MoE ships real architectural novelty — GDLA attention, annealed hyper-connections, multi-teacher distillation — but the technical report's abstract never states a single benchmark number, even though one exists in public trackers.
mixture-of-expertsllm-architecturemodel-evaluationattention-mechanismsRead