Nicolas Gras
@armgd.bsky.social
Tech Lead | GenAI Engineer | DevOps philosophy https://www.n-gr.xyz/ I also paint minis
Vous voulez savoir si vous verrez l'éclipse du 12 août prochain ? Ce site est fait pour vous ! Choisissez un emplacement et il calcule le pourcentage de l'éclipse et permet de dire si votre champ de vision sera obstrué par le relief ⤵️ via @labolycee.org eclipsemap.xyz
Solar Eclipse at sunset · August 12, 2026
Dove vedere l'eclissi? · ¿Dónde ver el eclipse? · Wo die Finsternis sehen? · Où voir l'éclipse?
eclipsemap.xyz
Sharing my interpretation of llm-wiki for working with cli agents github.com/Armgd/llm-wi... Goal is to extract and consign agentic transcripts following a PARA notetaking architecture in Obsidian. Own your knowledge
GitHub - Armgd/llm-wiki-claude-code: An implemantation of LLM Wiki, focused for devs and claude code sessions extraction
An implemantation of LLM Wiki, focused for devs and claude code sessions extraction - Armgd/llm-wiki-claude-code
github.com
PostgreSQL 19 adds Property Graph DDL: CREATE PROPERTY GRAPH with vertex/edge labels backed by tables/views, properties mapped from columns/expressions, plus indexes & constraints. Graph model becomes schema-managed in SQL—useful for relationship-heavy data and KG/RAG stores.
5.15. Property Graphs
5.15. Property Graphs # A property graph is a way to represent database contents, as an alternative to the usual (in …
postgresql.org
“iOS 27 system prompts” — extracted assistant instruction templates from dyld_shared_cache. Includes HomeKit summarizer, Wallet pass extractor, call-intelligence schemas, and other on-device directives. Useful for prompt design patterns + auditing privacy/attack surface.
iOS 27 system prompts
iOS 27 system prompts. GitHub Gist: instantly share code, notes, and snippets.
gist.github.com
Apple’s `container` repo documents “macOS Container Machines” — CLI-managed, container-capable VMs on macOS with details on networking, filesystem sharing, and limitations. A Docker Desktop alternative that makes the VM boundary explicit and scriptable.
container/docs/container-machine.md at main · apple/container
A tool for creating and running Linux containers using lightweight virtual machines on a Mac. It is written in Swift, and optimized for Apple silicon. - apple/container
github.com
Vincent Bernat documents a blogging pipeline that uses LLMs for copyediting/translation with explicit disclosure per post (icons + notes). Uses Claude Sonnet 4.5, DeepL, and Kagi, with concrete before/after edits and a year-by-year tool log.
Blogging with an LLM assistant
As a non-native English writer, I rely on LLMs for grammar, copyediting, and translation, despite hating AI slop.
vincent.bernat.ch
GitHub/npm v12 (July 2026) tightens install defaults: lifecycle scripts blocked unless explicitly approved (`allowScripts`), and git/remote URL dependencies blocked unless allowed. Upgrade to npm 11.16.0+, run installs to see warnings, use `npm approve-scripts` and commit the allowlist.
Upcoming breaking changes for npm v12 - GitHub Changelog
Our next npm major version, v12, introduces security-related default changes to npm install. All these changes are available behind warnings in npm today on 11.16.0 or newer, so you can…
github.blog
EU Open Source Strategy: full-lifecycle OSS plan for “tech sovereignty” — procurement guidance, an Open Internet Stack catalogue, and a maintenance instrument for critical dependency mapping + mirroring. Focus is on keeping deps buildable/patchable, not just funding repos.
The EU Open Source Strategy
The EU Open Source Strategy places open source at the centre of the EU’s technological sovereignty by promoting European open alternatives to non-EU proprietary solutions in critical domains.
digital-strategy.ec.europa.eu
“AI is slowing down” argues the drag isn’t benchmarks — it’s ops: inference cost, latency, GPU limits, safety/regulation, and product integration. Useful framing for leaders shipping LLM features with real SLAs, not demoware.
AI Is Slowing Down
If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s
wheresyoured.at
Signal’s PDF warns the UK’s proposed age-verification + automated nudity scanning would require client-side content inspection on everyone’s devices. That creates a permanent, updateable surveillance mechanism at the OS/app layer—easy to expand beyond the original scope.
Surveillance Is Not Safety: A statement on the UK's latest threat to privacy [pdf]
Signal’s June 8, 2026 statement condemns the UK government’s proposal to require age‑verification and automated nudity scanning of all device content as a form of mass surveillance that will erode privacy, strengthen platform control, and be expanded beyond its original scope; the statement argues this policy won’t keep children safe and calls for investment in education, social services, and AI guardrails instead. citeturn0view0
signal.org
Jimmy Koppel’s “Software Design in the Age of AI” argues LLMs will automate low-level coding but amplify the need for explicit design intent. Focus: removing hidden coupling, human-led refactoring, and tools/agents that assist understanding vs “dark factory” automation.
Arch-Engineer: Software Design in the Age of AI
Eliminate impostor syndrome, Escape software influencer dogma, Rise above subjective debates, Give concrete answers you can explain. Unleash your full potential as a software engineer
self-service.mirdin.com
“AI rockstar dev” failure mode: lots of LLM-generated code from ephemeral sessions → inconsistent patterns, naming, logging, retries, config schemas. The post argues for small guided snippets, slower architecture choices, and human curation to avoid unpayable tech debt.
Cleaning up after AI rockstar developers - Jesse Skinner
We
codingwithjesse.com
Apple Core AI: framework + tooling (coreai-torch, coreai-build) to compile and run .aimodel assets on-device across iPhone/iPad/Mac/visionOS. Includes debugging/profiling and fine-grained memory control. Zero server dependency, optimized for Apple silicon.
Core AI | Apple Developer Documentation
Run AI models in your app on Apple silicon.
developer.apple.com
0xsid documents a patched Instagram account takeover via Meta’s AI support flow: attacker gets support to change the linked email, then runs a standard password reset. 2FA is bypassed because account recovery becomes the new auth path.
The newest Instagram “exploit” is the goofiest I've seen
Short, punchy write-up describing a trivial-seeming Instagram account takeover that abused Meta’s AI-powered support flow to change a linked email and reset the password, bypassing 2FA; author notes Meta appears to have patched it. High signal for practitioners interested in real-world AI automation failures and the security gaps in account-recovery workflows.
0xsid.com
“AI job grief” essay: argues AI disruption triggers a distinct grief pattern (identity loss + anticipatory mourning) because the threat is ongoing. Proposes an AIRD framing, pulls from clinical literature + Reddit incidents, and notes HR lacks rituals/language for it.
AI Job Grief: The Unnamed Psychological Crisis Hitting Tech Workers
Across hundreds of Reddit threads and a small body of clinical literature, AI-driven displacement is producing an emotional category that most closely resembles grief, and the institutions causing it have no language for it.
jackmaguire.org
Many of you forgot too fast the insane amount of shitty software we had to see and suffer in the pre-AI era.
DBOS shows durable workflow execution on Postgres: job tables + `FOR UPDATE SKIP LOCKED`, advisory locks, LISTEN/NOTIFY, transactional retries, idempotent handlers. Fewer moving parts than a workflow engine, with trade-offs in long-running state machines + observability.
Postgres-backed Durable Workflow Execution | DBOS
Explaining the concept of a Postgres-backed durable execution system like DBOS and comparing it to external workflow orchestration systems like Temporal.
dbos.dev
Walkthrough: patching a Yamaha THR10c guitar amp firmware via UART/JTAG. Uses FT2232H + OpenOCD to dump flash, disassemble, relink, patch, and reflash. A practical embedded RE pipeline from soldering pads to rebuilding images.
Patching my guitar amp's firmware
A deep, hands‑on walkthrough of reverse engineering and patching the Yamaha THR10c guitar amp firmware. The post covers hardware work (soldering UART/JTAG connectors), using an FT2232H adapter with OpenOCD, dumping and disassembling memory, relinking and patching the firmware, flashing updates, and possible next steps like MIDI 2 and custom DSP.
mforney.org
Research shows a new web side‑channel: JavaScript can infer SSD I/O timing/patterns to fingerprint or track users without network beacons. It relies on storage access latency signals and could be mitigated via browser timer fuzzing, cache partitioning, or OS/firmware changes.
Websites have a new way to spy on visitors: Analyzing their SSD activity
Telltale SSD activity can be measured in the browser using simple JavaScript.
arstechnica.com
mlsu.io “Can we have the day off?” argues AI productivity gains should convert into shorter workweeks (e.g., Fridays off). It’s a management/redistribution take — automation savings get reclaimed by humans, not just more output.
Can we have the day off?
Short, timely opinion piece arguing that AI-driven productivity gains should translate into shorter workweeks (e.g., a Friday off), mixing wry personal anecdotes about childcare and a broader point about redistributing gains from automation. Useful provocation for engineering leaders weighing AI efficiency vs. workplace norms. citeturn0view0
mlsu.io
`airtop`: terminal “htop for the airwaves” — live 802.11 dashboard with spectrum, per-station RSSI traces, frame-type activity feed, histogram, and AP list using braille/block graphics. Uses eBPF on mac80211/cfg80211 from a normal connected iface (no monitor mode).
GitHub - yeet-src/airtop: htop for the airwaves — a live 802.11 (Wi-Fi) RF dashboard in your terminal
htop for the airwaves — a live 802.11 (Wi-Fi) RF dashboard in your terminal - yeet-src/airtop
github.com
Essay on “package managers that package package managers”: shipping npm/pip/etc inside products to make installs reproducible, at the cost of layered dependency resolution, harder security patching, and more complex upgrade/rollback paths for ops and maintainers.
Package managers that package package managers
brew install pip install poetry add pdm add uv tool install conda
nesbitt.io
DoomBench: a reproducible “Can it run DOOM?” benchmark for data stacks. It drives ingest/storage/query behavior with an unusual workload to expose tail latency, throughput limits, and operational failure modes across the full pipeline — not just microbenchmarks.
Introducing DoomBench - Can Your Data Stack Run DOOM?
A playful, practical benchmark that uses the cultural “Can it run DOOM?” meme to exercise and measure a data stack’s behaviour under unusual workloads. Useful as a reproducible, attention-grabbing stress test for DB/ingest/processing pipelines, it’s the kind of experiment SREs, backend engineers and tech leads can use to validate latency, throughput and operational assumptions.
cedardb.com
Starlette CVE‑2026‑48710 (“BadHost”) is an auth bypass via crafted Host header that can influence request.url.path and defeat path-based auth middleware. Affects Starlette <1.0.1 (many FastAPI AI services like vLLM/LiteLLM/MCP). Fix: upgrade to 1.0.1+.
Millions of AI agents imperiled by critical vulnerability in open source package
BadHost" was found in Starlette, a package with 325 million weekly downloads.
arstechnica.com
Agent Trace RFC (v0.1.0, Jan 2026): vendor-neutral JSON schema to attribute repo changes to humans + AI agents. Supports file/line ranges, movement-resistant content hashes, VCS metadata, model IDs, and session/conversation URLs. Excludes legal/provenance/quality.
Agent Trace
A standard format for tracking AI-generated code.
agent-trace.dev
Article argues Wikipedia is drifting toward Big Tech’s anti-labor pattern — privatizing value created by volunteers and weakening the public knowledge commons. Useful lens on platform governance, incentive design, and the hidden supply chain risk of volunteer-run infra.
Big Tech’s Anti-Labor Playbook Has Come for Wikipedia
TLDR: In ten days last month, the Wikimedia Foundation fired the longtime lead developer of MediaWiki and disbanded the team whose entire…
medium.com
SignalBloom models when “offshore engineer + local/OSS LLM (DeepSeek as proxy)” beats frontier API economics. Uses token blend + cache-hit assumptions and salary inputs to argue frontier inference is ~10x pricier, and rising token use + better local models cap pricing power.
Outsourcing plus LocalAI will soon become more economical vs Frontier labs | SignalBloom AI posts
Outsourcing plus LocalAI will soon become more economical vs Frontier labs
signalbloom.ai
Qwen3.7-Max (Qwen) is an “agent-first” LLM release focused on tool use and orchestration. The writeup frames production concerns—latency, tool-call reliability, and safety/guardrails—aimed at shipping LLM-driven automation instead of just chat.
Qwen Studio
Blog post announcing Qwen3.7-Max, positioned as an “agent-first” LLM release that emphasizes improved tool use, orchestration, and production-ready agent capabilities. Useful read for engineers evaluating agentic workflows and integrations — it highlights practical trade-offs (tooling, latency, safety/guardrails) that matter when shipping LLM-driven automation in production.
qwen.ai
