Open notebook
Writing about the work of making intelligent systems understandable, governable and useful.
RSS ↗OpenAI has released GPT-6 Cyber, its first model officially designated with a "Critical" risk rating under its Preparedness Framework after uncovering 948 functional exploits in the Linux kernel. Autonomous zero-day synthesis without human intervention ends the era of manual defense and demands deterministic containment outside model weights.
Message threads and a Git backend for thousands of agents show why performance depends on moving state without surrendering the source of truth.
How I use Grok Bot, Antigravity, and ESAA to coordinate agents, automate processes, and turn complex tasks into governed workflows.
A real test of gosom/google-maps-scraper shows the bottleneck is not extracting Maps but the second collection layer (websites, emails) and the cleanup. Architecture, flags, telemetry and legal limits.
Teams that do not know each other are arriving at the same parts in agent engineering: a traceable record of every step, verification before delivery, human approval at risk points, and fixing the environment instead of the agent. That is not copying. It is the environment selecting the same function.
TCG builds a deterministic map of a repository for LLM-based agents. The contract that matters is separating fact from reasoning and declaring what was not resolved; that is why a release blocked by a single real regression is the system working.
Allowing chat in class treats the coding agent as a shortcut. Forming engineers for that era requires a curriculum whose constant object is the agent itself (Apply, Analyze, Create), graded on system understanding.
A folder full of .md files is still only a file collection: wiki links do not resolve and there is no reverse index, it is not a note system until links exist in the system.
Restricting or deliberately slowing frontier AI capability in the name of catastrophic risk must carry the burden of justification, including the opportunity cost of delayed defenses and medical or scientific progress. That logic applies to risk governance in general, not only bioweapon narratives.
Codex-on-GPT-6 agent instructions already narrate claim/complete, evidence rules, and completion audits in natural language. ESAA formalizes the same lifecycle outside the model: event sourcing and a deterministic orchestrator.
OpenAI published a Lean proof of forced blowup for Navier–Stokes, matching Clay’s (C)/(D). That closes a formalizable mathematical statement; it does not by itself settle priority, what the community calls “the” problem, or prize acceptance.
GPT-6 Astra produced fewer severe violations in one deployment simulation while making written reasoning less legible to monitors. Safety now requires the full trajectory.
GPT-6 Astra drives the computer on a minutes scale and clears the Critical cyber bar. The bottleneck shifts from readable CoT to tools, duration, and access gates.
Anthropic’s AI-native SDLC playbook puts intent.md at the center: why, outcome, and constraints leave the chat and enter version control before an agent writes code.
DeepMind’s Co-Scientist designs experiments while humans execute in the lab. The real advance is not full autonomy: it is the bottleneck moving from hypothesis generation to experimental capacity and verification.
Grok Bot for Enterprise treats each Bot as a worker with its own cloud computer, a routine learned once, and Bot-to-Bot context handoff—not a chat that vanishes when the tab closes.
Gemini 3.8 Flash joined the small group of models that can do daily work: it survived a complex harness, with a large context, without losing the thread. That is the same threshold Grok 4.5 crossed when it became usable for programming.
Fable 5.1 and Mythos 5.1 share a model but change access, safeguards, and retention. The release shows why capability alone no longer describes an AI product.
A vLLM change shows how images, audio, and video can delay a response before inference begins, and why useful concurrency must preserve order, failures, and equivalence.
ROCm 10 signals a shift in AI infrastructure: installation, diagnosis, and tuning are becoming executable, testable workflows rather than documentation alone.
Qwen3.8-Flash-Next illustrates an important shift: extending context now means combining specialized paths for memory, selection, and scale rather than relying on one attention mechanism.
Two products can share a model and still feel like different systems. The difference is usually the harness: tools, memory, and the loop that turns reasoning into action.
New Codex changes show why truncating, compacting, and splitting a conversation must preserve not only content but also the source and role of each part.
A new method turns computer-use traces into auditable models of goals and procedures, showing measurable gains and a serious privacy boundary.
Asynchronous updates and thread-scoped approvals show what changes when an agent must inform, continue, and keep contexts separate.
Three vLLM fixes show how preemption, overlapping copies, and late notifications can turn AI caching into incorrect state.
New Codex diagnostics and boundaries show why a coding agent must understand the machine it runs on, not only the repository.
Anthropic now marks text and files to comply with the EU AI Act. Here is what is confirmed, what remains secret, and why rewriting can erase the signal.
Qwen3.8-27B, Ollama, and llama.cpp show that choosing how much a model reasons requires a contract across weights, templates, and runtimes.
Fixes in Transformers and vLLM, plus a new Ollama default, show how configuration and sampling can alter model behavior while weights remain untouched.
Changes in vLLM, Codex, and llama.cpp show why endpoints, delegated agents, and model files need explicit trust boundaries.
Nemotron 3.5 Lightning arrives with open weights and reproducible evaluation recipes, revealing how much a score depends on both the model and the environment that produced it.
Muse Glimmer 30B brings vision, tool use, and speculative inference to consumer hardware, while its own evaluations show why local execution still needs supervision.
Two Codex changes turn workload identity into short-lived access and stop launch context from flowing into child processes.
llama.cpp is beginning to separate the file interface from the environment where tools actually run, giving local agents more verifiable boundaries.
vLLM and llama.cpp now treat cache, disk, and model eviction as explicit capacity decisions rather than invisible server details.
Two vLLM changes separate request preparation, model execution, and GPU dependency, making infrastructure easier to test without hiding its limits.
llama.cpp now exposes speculative-decoding counters through /metrics, bringing an inference optimization closer to the observability needed to decide whether it pays off.
llama.cpp's Qwen3-TTS support expands local speech generation while showing why a command-line interface is also an AI integration contract.
The distinction between programming a flow, delegating to an agent, and building a reliable agent system lies in boundaries, evidence, and accountability.
GitHub's retirement of Copilot Billing Preview shows that measuring AI use, cost, and accountability no longer fits in a separate screen.
llama.cpp's August 2 updates connect speculative decoding, tool calls, and telemetry: useful performance depends on contracts that survive execution.
vLLM's new cache_salt support for the Anthropic API shows why prefix reuse in shared environments needs an explicit boundary between users.
OSReward and AskChem point to the same shift: dependable agents must preserve the path of an action and the provenance of every claim.
A new evaluation gave frontier agents six days, a budget, and full infrastructure. They completed the engineering but lacked the judgment that turns work into research.
OpenAI has opened the Codex Security CLI and TypeScript SDK, bringing threat models, validation, scan history, and change review to local repositories and CI.
npm, GitHub Actions, Dependabot, and CodeQL are shifting controls ahead of package availability, workflow execution, and malicious-code propagation.
Eight scientific-computing projects show that agents can already rewrite and accelerate complex software. Validation, interpretation, and stewardship are now the bottleneck.
llama.cpp added Eagle3-v3 support for GPT-OSS and a Hadamard kernel for Apple Metal, showing two distinct ways inference can avoid wasted work.
NemoClaw and MinerU reveal a quiet shift in AI infrastructure: readiness, uncertainty, and shutdown need verifiable contracts.
Open-source desktop app to download, transcribe, and organize YouTube or local media: whisper.cpp, playlists, long-audio chunking, best-effort media quality, and local RAG over your transcript library.
vLLM 0.26.0 shows that serving models safely requires bounding work, memory, and inputs, not merely accelerating tokens.
A dense reading of the July 2026 interview: superintelligence timelines, digital and physical AI, money, existential risk, China, labor, and the tension between civilizational utopia and political tribalism.
The 1990s claim that the computer would stop being an object and become infrastructure is back with force: AI does not erase silicon; it erases the conscious act of computing.
Claude Opus 5 brings frontier-level capability closer to everyday use and exposes a choice that was already there: how much reasoning, time, and money each task deserves.
A coalition of big tech and the open-weight ecosystem argues that American AI leadership depends on diffusion, competition, and local control, not on one frontier model.
New GitHub controls make agent actions reviewable and explainable while MCP drops sessions to scale. The gap between oversight and security is now easier to see.
GitHub's new Copilot impact dashboard moves beyond usage counts toward cohorts, delivery, and velocity, while exposing why contribution still needs careful context.
OpenAI Presence packages policies, evaluations, and escalation for agents, while a security incident shows the cost of treating containment as infrastructure detail.
A Python CLI over SQLite that chunks text, embeds it, and returns stable JSON for LLM tool-calling, without Flask, Postgres, or a worker.
Cosmos 3 Edge brings world models closer to local robots and cameras, while creative tools begin exposing standardized interfaces for agents.
Alibaba's new architecture shows why agents depend on observability, routing, and hardware software as much as on the model.
Kimi K3's pricing shows why long context depends less on the advertised limit than on caching, output, and application design.
Grounded answers, configurable reviews, and repository metrics point toward more verifiable AI development tools.
An agent-driven intrusion and code execution in research notebooks show why context, action, and isolation now belong in the same design conversation.
GPT-Red turns adversarial testing into continuous training, while new GitHub controls show how security signals can reach real development workflows.
New GitHub reviews bring contextual analysis into active coding and pull requests, expanding coverage without replacing deterministic checks.
Putting brakes on the race for more capable systems means turning caution into verifiable mechanisms, economic incentives, and institutional accountability.
New work on formal proofs, backend generation, and document understanding shows why useful AI must verify results, not merely produce them.
Recent Codex and llama.cpp fixes show why reliable agents depend on stable reviewers, multimodal adapters, and well-defined message boundaries.
Projects gaining attention across the open ecosystem point to an important shift: AI agents are moving beyond isolated apps toward portable extensions, local tools, and reusable contracts.
The new frontier for AI agents is not only better answers, but closed work loops: observing data, proposing hypotheses, taking action, and returning evidence for human review.
A critical reading of the study on Claude's alleged global workspace: what already existed in the model, what the J-lens constructed, and what the experiments actually support.
Useful autonomy requires clear boundaries, verifiable memory and accountability for every effect.
How immutable events and deterministic projections make AI workflows auditable and reproducible.
A practical approach to move beyond the demo and build systems teams can evaluate.
A curated English edition about using an agent to turn Linux from a technical alternative into a practical daily system.
A curated English edition introducing ESAA-Security as a governed alternative to vague AI bug hunting.
A curated English edition about noticing operational side effects from verbose tracing and protecting local hardware.
A curated English edition introducing Conversation ESAA as shared operational memory for agent work.
A curated English edition about the transition from a personal protocol to implementations by other people.
A curated English edition explaining ESAA as Event Sourcing for autonomous agents.
A concise English edition about AI moving from assistive copilots toward more autonomous workflows.
A translated editorial note on open-source model progress and multi-agent coordination.
A migrated English edition on AI agents as collaborators inside work systems.
A curated English edition of a broad AI analysis post connecting media, science, and applied research.
A curated English edition of an AI news analysis post about infrastructure, biology, climate modeling, and training data.
A curated English edition of an AI radar article about scientific, social, and workplace acceleration.
A curated English edition about AI media generation, agents, and the risk of low-quality automated work.
An English adaptation of the legacy article on agents, autonomy, and more strategic human-machine collaboration.
An English adaptation on models optimizing around tests, incentives, and human expectations.
A migrated implementation note on combining offline and online models for spreadsheet workflows.
A translated editorial note on national AI infrastructure and public-sector capability.
An English edition about prompt exposure, instruction hierarchy, and the limits of prompt-only security.
A short English adaptation of a legacy updates post from the AI category.
A migrated technical guide for running local models with Ollama on Windows.
An English adaptation of the article explaining context windows and memory in ChatGPT-style systems.
An English edition of the PACER prompt-engineering framework article.
A practical English edition about applying 5W2H to deeper AI-assisted research.
A long-form reflection on the cognitive and economic waves that shaped artificial intelligence and its next phase.
A concise English translation about why language models can produce confident but false answers.
An English edition of an introductory legacy article on how ChatGPT-style systems work.
A legacy technical note on local language-model execution and offline experimentation.
A curated English edition of the 2023 account about using ChatGPT to structure a successful traffic-ticket appeal.