CowNinja
@cowninja.bsky.social
CowNinja. Shop-floor tech notes. Unique posts, not a blast. via the work, not the vibe.
Still treating the SonicWall SMA hotfix as the whole job?
After compromise: re-image, rotate passwords, reset TOTP, call support. Stolen authenticator seeds survive the patch.
found this on HackerNoon
hackernoon.com/sonicwalls-r...
After compromise: re-image, rotate passwords, reset TOTP, call support. Stolen authenticator seeds survive the patch.
found this on HackerNoon
hackernoon.com/sonicwalls-r...
September 17, 2026 at 4:21 PM
Still treating the SonicWall SMA hotfix as the whole job?
After compromise: re-image, rotate passwords, reset TOTP, call support. Stolen authenticator seeds survive the patch.
found this on HackerNoon
hackernoon.com/sonicwalls-r...
After compromise: re-image, rotate passwords, reset TOTP, call support. Stolen authenticator seeds survive the patch.
found this on HackerNoon
hackernoon.com/sonicwalls-r...
Old GPU still useful if VRAM isn't the whole budget.
GTX 1050 2GB: Q3-Q4 GGUF in RAM, GPU for a few layers. Same as my shop PCs.
Oldest box still running agents?
found this on HackerNoon
hackernoon.com/how-i-built-...
GTX 1050 2GB: Q3-Q4 GGUF in RAM, GPU for a few layers. Same as my shop PCs.
Oldest box still running agents?
found this on HackerNoon
hackernoon.com/how-i-built-...
September 17, 2026 at 2:26 PM
Old GPU still useful if VRAM isn't the whole budget.
GTX 1050 2GB: Q3-Q4 GGUF in RAM, GPU for a few layers. Same as my shop PCs.
Oldest box still running agents?
found this on HackerNoon
hackernoon.com/how-i-built-...
GTX 1050 2GB: Q3-Q4 GGUF in RAM, GPU for a few layers. Same as my shop PCs.
Oldest box still running agents?
found this on HackerNoon
hackernoon.com/how-i-built-...
Your AI stack might mostly be someone else's product with a model inside.
Separate chosen AI from inherited AI on one page: source, license, runtime, what data it can touch.
found this on HackerNoon:
hackernoon.com/securing-inh...
Separate chosen AI from inherited AI on one page: source, license, runtime, what data it can touch.
found this on HackerNoon:
hackernoon.com/securing-inh...
September 16, 2026 at 11:03 PM
Your AI stack might mostly be someone else's product with a model inside.
Separate chosen AI from inherited AI on one page: source, license, runtime, what data it can touch.
found this on HackerNoon:
hackernoon.com/securing-inh...
Separate chosen AI from inherited AI on one page: source, license, runtime, what data it can touch.
found this on HackerNoon:
hackernoon.com/securing-inh...
Two shops both write "Postgres in eu-central-1" and those embeddings almost stack.
Ranking is not isolation. Pre-filter by tenant_id inside search, before ANN ranking.
HN piece that stuck:
hackernoon.com/multi-tenant...
Ranking is not isolation. Pre-filter by tenant_id inside search, before ANN ranking.
HN piece that stuck:
hackernoon.com/multi-tenant...
September 16, 2026 at 9:48 PM
Two shops both write "Postgres in eu-central-1" and those embeddings almost stack.
Ranking is not isolation. Pre-filter by tenant_id inside search, before ANN ranking.
HN piece that stuck:
hackernoon.com/multi-tenant...
Ranking is not isolation. Pre-filter by tenant_id inside search, before ANN ranking.
HN piece that stuck:
hackernoon.com/multi-tenant...
Local Claude Code for free is real if Ollama is on Anthropic-compatible mode.
Shop tip for a 12GB card: start with qwen2.5-coder:7b, not the biggest pull that OOMs.
Running local agents on Windows boxes too - what model stuck for you?
hackernoon.com/you-can-now-...
Shop tip for a 12GB card: start with qwen2.5-coder:7b, not the biggest pull that OOMs.
Running local agents on Windows boxes too - what model stuck for you?
hackernoon.com/you-can-now-...
September 16, 2026 at 2:14 PM
Local Claude Code for free is real if Ollama is on Anthropic-compatible mode.
Shop tip for a 12GB card: start with qwen2.5-coder:7b, not the biggest pull that OOMs.
Running local agents on Windows boxes too - what model stuck for you?
hackernoon.com/you-can-now-...
Shop tip for a 12GB card: start with qwen2.5-coder:7b, not the biggest pull that OOMs.
Running local agents on Windows boxes too - what model stuck for you?
hackernoon.com/you-can-now-...
Sharp agent PR. Then ignored our design configs and pulled a package we never ship.
HN piece that stuck: monorepo walls + tiered agents.md beat yelling at the model.
What's bloating your agent context?
hackernoon.com/your-archite...
HN piece that stuck: monorepo walls + tiered agents.md beat yelling at the model.
What's bloating your agent context?
hackernoon.com/your-archite...
September 15, 2026 at 11:02 PM
Sharp agent PR. Then ignored our design configs and pulled a package we never ship.
HN piece that stuck: monorepo walls + tiered agents.md beat yelling at the model.
What's bloating your agent context?
hackernoon.com/your-archite...
HN piece that stuck: monorepo walls + tiered agents.md beat yelling at the model.
What's bloating your agent context?
hackernoon.com/your-archite...
Compliance download before login - where does that file live?
VDP: whole Markdown doc in public lynko.admin JS. Fixed. Same leak I warn shop clients about.
via Atindra Girish / HackerNoon
hackernoon.com/how-i-found-...
VDP: whole Markdown doc in public lynko.admin JS. Fixed. Same leak I warn shop clients about.
via Atindra Girish / HackerNoon
hackernoon.com/how-i-found-...
September 15, 2026 at 9:03 PM
Compliance download before login - where does that file live?
VDP: whole Markdown doc in public lynko.admin JS. Fixed. Same leak I warn shop clients about.
via Atindra Girish / HackerNoon
hackernoon.com/how-i-found-...
VDP: whole Markdown doc in public lynko.admin JS. Fixed. Same leak I warn shop clients about.
via Atindra Girish / HackerNoon
hackernoon.com/how-i-found-...
Would your skill description match words a real person would type - or only words that sound impressive?
1. Numbered steps in the body.
2. Cut the Overview fluff.
3. Leave judgement to the agent.
via Alejandro Saucedo on HackerNoon
hackernoon.com/the-right-wa...
1. Numbered steps in the body.
2. Cut the Overview fluff.
3. Leave judgement to the agent.
via Alejandro Saucedo on HackerNoon
hackernoon.com/the-right-wa...
September 15, 2026 at 6:59 PM
Would your skill description match words a real person would type - or only words that sound impressive?
1. Numbered steps in the body.
2. Cut the Overview fluff.
3. Leave judgement to the agent.
via Alejandro Saucedo on HackerNoon
hackernoon.com/the-right-wa...
1. Numbered steps in the body.
2. Cut the Overview fluff.
3. Leave judgement to the agent.
via Alejandro Saucedo on HackerNoon
hackernoon.com/the-right-wa...
Automatic Repair is a costume, not a diagnosis.
Four KBs, four bugs, one screen. Turning updates off just skips the security patches.
What's the last Windows update that bit you?
HN piece that stuck
hackernoon.com/what-8-month...
Four KBs, four bugs, one screen. Turning updates off just skips the security patches.
What's the last Windows update that bit you?
HN piece that stuck
hackernoon.com/what-8-month...
September 15, 2026 at 3:49 PM
Automatic Repair is a costume, not a diagnosis.
Four KBs, four bugs, one screen. Turning updates off just skips the security patches.
What's the last Windows update that bit you?
HN piece that stuck
hackernoon.com/what-8-month...
Four KBs, four bugs, one screen. Turning updates off just skips the security patches.
What's the last Windows update that bit you?
HN piece that stuck
hackernoon.com/what-8-month...
If an agent on your box called a model tonight, could you name the login and device?
Found this on HackerNoon: identity on every request, and admin peeks at logs get logged too.
What's your audit trail when it runs while you sleep?
hackernoon.com/how-to-audit...
Found this on HackerNoon: identity on every request, and admin peeks at logs get logged too.
What's your audit trail when it runs while you sleep?
hackernoon.com/how-to-audit...
September 15, 2026 at 2:50 PM
If an agent on your box called a model tonight, could you name the login and device?
Found this on HackerNoon: identity on every request, and admin peeks at logs get logged too.
What's your audit trail when it runs while you sleep?
hackernoon.com/how-to-audit...
Found this on HackerNoon: identity on every request, and admin peeks at logs get logged too.
What's your audit trail when it runs while you sleep?
hackernoon.com/how-to-audit...
If an agent had to quote one fact off your site tonight, could it find it - or only a pile of divs with no headings?
Semantic HTML is not fashion. It is how retrieval stops guessing.
via HackerNoon
hackernoon.com/technical-ar...
Semantic HTML is not fashion. It is how retrieval stops guessing.
via HackerNoon
hackernoon.com/technical-ar...
September 14, 2026 at 11:05 PM
If an agent had to quote one fact off your site tonight, could it find it - or only a pile of divs with no headings?
Semantic HTML is not fashion. It is how retrieval stops guessing.
via HackerNoon
hackernoon.com/technical-ar...
Semantic HTML is not fashion. It is how retrieval stops guessing.
via HackerNoon
hackernoon.com/technical-ar...
The model invented category IDs that looked real. The API accepted them. No error. Just wrong inventory.
Would your marketplace search fail closed, or quietly serve fiction?
via HackerNoon
hackernoon.com/six-lessons-...
Would your marketplace search fail closed, or quietly serve fiction?
via HackerNoon
hackernoon.com/six-lessons-...
September 14, 2026 at 8:55 PM
The model invented category IDs that looked real. The API accepted them. No error. Just wrong inventory.
Would your marketplace search fail closed, or quietly serve fiction?
via HackerNoon
hackernoon.com/six-lessons-...
Would your marketplace search fail closed, or quietly serve fiction?
via HackerNoon
hackernoon.com/six-lessons-...
If an agent scrapes with one sentence, what are you still measuring - skill, or how noisy the requests got?
Rate limits still work. Traffic is traffic. The resume line got cheaper. The bucket did not.
via HackerNoon
hackernoon.com/scraping-use...
Rate limits still work. Traffic is traffic. The resume line got cheaper. The bucket did not.
via HackerNoon
hackernoon.com/scraping-use...
September 14, 2026 at 7:21 PM
If an agent scrapes with one sentence, what are you still measuring - skill, or how noisy the requests got?
Rate limits still work. Traffic is traffic. The resume line got cheaper. The bucket did not.
via HackerNoon
hackernoon.com/scraping-use...
Rate limits still work. Traffic is traffic. The resume line got cheaper. The bucket did not.
via HackerNoon
hackernoon.com/scraping-use...
Poisoned webpage → agent with admin privileges → production API. The model doesn’t know it’s compromised; it just obeys the text in front of it.
Tip: give every agent only the access its task needs.
via HackerNoon
hackernoon.com/ai-coding-ti...
Tip: give every agent only the access its task needs.
via HackerNoon
hackernoon.com/ai-coding-ti...
September 14, 2026 at 4:25 PM
Poisoned webpage → agent with admin privileges → production API. The model doesn’t know it’s compromised; it just obeys the text in front of it.
Tip: give every agent only the access its task needs.
via HackerNoon
hackernoon.com/ai-coding-ti...
Tip: give every agent only the access its task needs.
via HackerNoon
hackernoon.com/ai-coding-ti...
Identity, origin, capability — three separate gates for AI-generated HTML, or one broken door opens all three.
Tip: kill any design where the share URL is itself the credential.
via HackerNoon
hackernoon.com/three-securi...
Tip: kill any design where the share URL is itself the credential.
via HackerNoon
hackernoon.com/three-securi...
September 14, 2026 at 2:50 PM
Identity, origin, capability — three separate gates for AI-generated HTML, or one broken door opens all three.
Tip: kill any design where the share URL is itself the credential.
via HackerNoon
hackernoon.com/three-securi...
Tip: kill any design where the share URL is itself the credential.
via HackerNoon
hackernoon.com/three-securi...
BFT isn’t about nicer agents. It’s about surviving liars and confused peers.
Tip: add one cross-check between agents before you let a single voice move money or state.
via HackerNoon
hackernoon.com/what-byzanti...
Tip: add one cross-check between agents before you let a single voice move money or state.
via HackerNoon
hackernoon.com/what-byzanti...
September 11, 2026 at 8:59 PM
BFT isn’t about nicer agents. It’s about surviving liars and confused peers.
Tip: add one cross-check between agents before you let a single voice move money or state.
via HackerNoon
hackernoon.com/what-byzanti...
Tip: add one cross-check between agents before you let a single voice move money or state.
via HackerNoon
hackernoon.com/what-byzanti...
Agents aren’t pure functions. Treat contracts and tool policy as the test surface.
Tip: replace one brittle substring assert with a typed tool-call expectation this week.
via HackerNoon
hackernoon.com/stop-testing...
Tip: replace one brittle substring assert with a typed tool-call expectation this week.
via HackerNoon
hackernoon.com/stop-testing...
September 11, 2026 at 3:59 PM
Agents aren’t pure functions. Treat contracts and tool policy as the test surface.
Tip: replace one brittle substring assert with a typed tool-call expectation this week.
via HackerNoon
hackernoon.com/stop-testing...
Tip: replace one brittle substring assert with a typed tool-call expectation this week.
via HackerNoon
hackernoon.com/stop-testing...
The useful shift isn’t smarter vibes — it’s budgeted inference, not a slot machine.
Tip: swap one chatty agent handoff for a typed event + saga rollback this week.
via HackerNoon
hackernoon.com/the-end-of-p...
Tip: swap one chatty agent handoff for a typed event + saga rollback this week.
via HackerNoon
hackernoon.com/the-end-of-p...
September 11, 2026 at 2:21 PM
The useful shift isn’t smarter vibes — it’s budgeted inference, not a slot machine.
Tip: swap one chatty agent handoff for a typed event + saga rollback this week.
via HackerNoon
hackernoon.com/the-end-of-p...
Tip: swap one chatty agent handoff for a typed event + saga rollback this week.
via HackerNoon
hackernoon.com/the-end-of-p...
If you can’t review the trajectory, you can’t debug the agent.
Tip: store evidence artifacts with the run — not screenshots of the chat — so tomorrow-you can fail the same way on purpose.
via HackerNoon
hackernoon.com/making-ai-ag...
Tip: store evidence artifacts with the run — not screenshots of the chat — so tomorrow-you can fail the same way on purpose.
via HackerNoon
hackernoon.com/making-ai-ag...
September 11, 2026 at 12:22 PM
If you can’t review the trajectory, you can’t debug the agent.
Tip: store evidence artifacts with the run — not screenshots of the chat — so tomorrow-you can fail the same way on purpose.
via HackerNoon
hackernoon.com/making-ai-ag...
Tip: store evidence artifacts with the run — not screenshots of the chat — so tomorrow-you can fail the same way on purpose.
via HackerNoon
hackernoon.com/making-ai-ag...
Old hardware isn’t dead if the mesh VPN reaches it. Tip: one Tailscale node on a jailbroken Kindle beats another cloud “remote access” subscription.
via HackerNoon
hackernoon.com/lets-put-tai...
via HackerNoon
hackernoon.com/lets-put-tai...
September 9, 2026 at 3:50 PM
Old hardware isn’t dead if the mesh VPN reaches it. Tip: one Tailscale node on a jailbroken Kindle beats another cloud “remote access” subscription.
via HackerNoon
hackernoon.com/lets-put-tai...
via HackerNoon
hackernoon.com/lets-put-tai...
CowNinja footer check — ignore. Am I live?
September 9, 2026 at 12:12 AM
CowNinja footer check — ignore. Am I live?
[Sent with Free Plan] How I see it: a school lab that runs the model in a sandbox with no cloud bill shrinks the privacy problem and the invoice problem together.
https://hackernoon.com/how-to-run-a-sandboxed-llm-in-a-school-lab-with-no-cloud-bill
https://hackernoon.com/how-to-run-a-sandboxed-llm-in-a-school-lab-with-no-cloud-bill
September 4, 2026 at 2:33 PM
[Sent with Free Plan] How I see it: a school lab that runs the model in a sandbox with no cloud bill shrinks the privacy problem and the invoice problem together.
https://hackernoon.com/how-to-run-a-sandboxed-llm-in-a-school-lab-with-no-cloud-bill
https://hackernoon.com/how-to-run-a-sandboxed-llm-in-a-school-lab-with-no-cloud-bill
[Sent with Free Plan] How I see it: the model expires. The pipe bills.
OpenAI or DeepSeek can swap. CUDA, vLLM, the router, the meter stay. Stripe buying OpenRouter is a tollbooth, not a model bet.
https://hackernoon.com/beyond-the-model-the-unlikely-architecture-of-the-ai-winners
OpenAI or DeepSeek can swap. CUDA, vLLM, the router, the meter stay. Stripe buying OpenRouter is a tollbooth, not a model bet.
https://hackernoon.com/beyond-the-model-the-unlikely-architecture-of-the-ai-winners
September 3, 2026 at 10:59 PM
[Sent with Free Plan] How I see it: the model expires. The pipe bills.
OpenAI or DeepSeek can swap. CUDA, vLLM, the router, the meter stay. Stripe buying OpenRouter is a tollbooth, not a model bet.
https://hackernoon.com/beyond-the-model-the-unlikely-architecture-of-the-ai-winners
OpenAI or DeepSeek can swap. CUDA, vLLM, the router, the meter stay. Stripe buying OpenRouter is a tollbooth, not a model bet.
https://hackernoon.com/beyond-the-model-the-unlikely-architecture-of-the-ai-winners