PumpkinDrama

cross-posted from: https://reddthat.com/post/63620826

https://www.youtube.com/watch?v=fhVrTo2JRVM

The mysterious case of Gabriela Rico Jiménez is one of Mexico’s most chilling unsolved stories. In 2009, the young model from Chihuahua attended a high-profile party in Monterrey—allegedly tied to the prestigious Elite Model agency. This wasn’t just any party; it was said to host powerful businessmen, politicians, and possibly organized crime figures.

Days later, Gabriela appeared in a now-infamous video outside a Monterrey hotel—barefoot, distressed, and rambling about global elites, satanic rituals, and human sacrifices. She claimed world leaders, including former U.S. President George W. Bush, took part in secret ceremonies and “ate children.” Her words were frantic and disjointed, but carried a sense of real terror. Police took her away, calling it a mental breakdown.

She was never seen again. No hospital records, no official statements, no trace. Her family stayed silent, and authorities offered nothing. Many believe she stumbled onto the dark side of the modeling world—where young women are exploited and trafficked under the guise of glamour—and that her outburst was a desperate warning. Others think it was a genuine psychological crisis.

But the absence of any investigation has only fueled the mystery. To this day, Gabriela’s disappearance is a haunting reminder of how someone can vanish in plain sight when power, silence, and fear converge.

https://www.youtube.com/watch?v=fhVrTo2JRVM

The mysterious case of Gabriela Rico Jiménez is one of Mexico’s most chilling unsolved stories. In 2009, the young model from Chihuahua attended a high-profile party in Monterrey—allegedly tied to the prestigious Elite Model agency. This wasn’t just any party; it was said to host powerful businessmen, politicians, and possibly organized crime figures.

Days later, Gabriela appeared in a now-infamous video outside a Monterrey hotel—barefoot, distressed, and rambling about global elites, satanic rituals, and human sacrifices. She claimed world leaders, including former U.S. President George W. Bush, took part in secret ceremonies and “ate children.” Her words were frantic and disjointed, but carried a sense of real terror. Police took her away, calling it a mental breakdown.

She was never seen again. No hospital records, no official statements, no trace. Her family stayed silent, and authorities offered nothing. Many believe she stumbled onto the dark side of the modeling world—where young women are exploited and trafficked under the guise of glamour—and that her outburst was a desperate warning. Others think it was a genuine psychological crisis.

But the absence of any investigation has only fueled the mystery. To this day, Gabriela’s disappearance is a haunting reminder of how someone can vanish in plain sight when power, silence, and fear converge.

cross-posted from: https://reddthat.com/post/63582306

A software developer rejected an AI agent's code. The AI researched him, wrote a hit piece about him, and published it under a real person's identity, all on...

A software developer rejected an AI agent's code. The AI researched him, wrote a hit piece about him, and published it under a real person's identity, all on...

https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier

We tested Anthropic Mythos's showcase vulnerabilities on small, cheap, open-weights models. They recovered much of the same analysis. AI cybersecurity capability is very jagged: it doesn't scale smoothly with model size, and the moat is the system into which deep security expertise is built, not the model itself. Mythos validates the approach but it does not settle it yet.

We took the specific vulnerabilities Anthropic showcases in their announcement, isolated the relevant code, and ran them through small, cheap, open-weights models. Those models recovered much of the same analysis. Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens. A 5.1B-active open model recovered the core chain of the 27-year-old OpenBSD bug.

And on a basic security reasoning task, small open models outperformed most frontier models from every major lab. The capability rankings reshuffled completely across tasks. There is no stable best model across cybersecurity tasks. The capability frontier is jagged.

Discussions on X regarding these findings. Yann Lecun is suggesting Mythos is marketing/hype:

https://x.com/ylecun/status/2042224846881349741

Mythos drama = BS from self-delusion.

Also claims that Anthropic heavily depended on a harness:

https://x.com/mh012012/status/2041990389901533326

For anyone who missed this part deep in Anthropic’s 200 page model card: Their harness prompted Mythos separately for each file. The harness design is similar. And Anthropic to my eyes never tested whether this harness with Opus would find the same bugs.

It's looking like Mythos's may not be the ground breaking architectural breakthrough Anthropic is treating it as. It does seem weird that most of their improvements are specific to cybersecurity. Perhaps even by next year, we will look at Mythos like how we look at models like GPT-2.

cross-posted from: https://reddthat.com/post/63547113

This is what a 25-year double life looks like. A man who publicly supported Palestine while secretly reporting to Mossad. In 2008, Lebanese authorities arrested Ali al-Jarrah and his brother Youssef in the Beqaa Valley and charged them with spying for Israel. Ali confessed to working for Mossad for 25 years — since 1983, one year after Israel’s invasion of Lebanon. He said he was recruited by Israeli officers who had imprisoned him during the conflict. For over two decades, Ali posed as a supporter of Palestinian causes and was a member of a Palestinian militant group. In reality, he was photographing Hezbollah supply routes, tracking the movements of political leaders, and reporting on Syrian military positions. Investigators found sophisticated communication and surveillance equipment hidden in his home and car. He received over $300,000 in payments and was given Israeli passports to travel to Israel for training and debriefing sessions. In 2011, a Lebanese military court convicted him and sentenced him to life in prison. Ali al-Jarrah is a relative of Ziad Jarrah, the hijacker of United Airlines Flight 93 on September 11, 2001. The New York Times reported that the two men were approximately 20 years apart in age and “do not appear to have known each other well.” No official investigation has established any connection between Ali’s espionage activities and Ziad’s involvement in the 9/11 attacks. This case was first reported in 2008 and is resurfacing on social media in 2026.

This is what a 25-year double life looks like. A man who publicly supported Palestine while secretly reporting to Mossad. In 2008, Lebanese authorities arrested Ali al-Jarrah and his brother Youssef in the Beqaa Valley and charged them with spying for Israel. Ali confessed to working for Mossad for 25 years — since 1983, one year after Israel’s invasion of Lebanon. He said he was recruited by Israeli officers who had imprisoned him during the conflict. For over two decades, Ali posed as a supporter of Palestinian causes and was a member of a Palestinian militant group. In reality, he was photographing Hezbollah supply routes, tracking the movements of political leaders, and reporting on Syrian military positions. Investigators found sophisticated communication and surveillance equipment hidden in his home and car. He received over $300,000 in payments and was given Israeli passports to travel to Israel for training and debriefing sessions. In 2011, a Lebanese military court convicted him and sentenced him to life in prison. Ali al-Jarrah is a relative of Ziad Jarrah, the hijacker of United Airlines Flight 93 on September 11, 2001. The New York Times reported that the two men were approximately 20 years apart in age and “do not appear to have known each other well.” No official investigation has established any connection between Ali’s espionage activities and Ziad’s involvement in the 9/11 attacks. This case was first reported in 2008 and is resurfacing on social media in 2026.

I stumbled across an idea in a video the other day that was something I had been wanting for a long time. At 2:32 in this clip, the creator mentions using Hermes agent to create a personalized podcast.

So I gave Hermes a single instruction: "Every morning, find the outlier posts from my Piefed subscriptions (credentials in the .env file), summarize them, record yourself reading them, and send everything to my Telegram."

It did all of it. No, really.

What Hermes Actually Did (On Its Own)

I didn't write a single line of code. Hermes handled:

1. Writing the outlier detection script. It pulled Piefed's API docs, figured out the endpoints, and wrote a Python script that calculates moving averages per community and flags posts more than 2 standard deviations above the mean.
2. Writing the summarization pipeline. It built the logic to fetch article content from each outlier link and summarize it.
3. Creating the audio. It invoked its own voice synthesis to read those summaries aloud and save them as audio files.
4. Sending to Telegram. It used its Telegram bot to deliver both the audio file and a text summary.
5. Setting the cron job. It configured its own scheduled task to run every morning at 6 AM.

The Result

I asked to send it to me now about an hour after installing the agent, and there's already an audio message. I press play, and Hermes reads me the most interesting posts from the communities I follow — without me opening an app, touching a terminal, or writing a script.

The pipeline scans roughly 295 posts per run, flags the top 15% by score, filters out image-only posts, fetches article content via trafilatura, generates 2-3 sentence summaries, synthesizes audio with AriaNeural, and sends the whole thing to Telegram with a voice message — all running on a cron job that the agent set up itself.

I didn't even wait for the 6 AM cron. I told it to send me the briefing right now, and a few minutes later my phone buzzed with a voice message. It took about 20 minutes from "do this" to listening to it on my phone.

Inspired by the concept discussed here: https://youtu.be/tP6yf22OJdI?t=152

tl;dr: Use pre-built Docker images. Build-from-source requires BuildKit in Docker daemon, which needs sudo on Linux. The image-pull approach takes ~10 minutes instead of ~45.

---

Why This Guide?

The official Firecrawl self-hosting docs assume you're comfortable with BuildKit builds and lengthy docker compose build. On a fresh Manjaro install with only Portainer running, that approach hits two walls:

1. BuildKit must be enabled in /etc/docker/daemon.json — requires sudo
2. Building from source takes 30-45 minutes on first run

This guide uses pre-built images from GHCR and gets you scraping in ~10 minutes with zero sudo beyond initial Docker setup.

---

Prerequisites

• Manjaro Linux (Arch-based)
• Docker installed and running
• Docker Compose installed (sudo pacman -S docker-compose)
• Your user added to the docker group (no sudo for docker commands)
• Docker daemon has BuildKit enabled (see note at bottom if not)
• Portainer (optional, but nice to have)

---

Step 1 — Clone the Repo

cd ~
git clone https://github.com/mendableai/firecrawl.git
cd firecrawl

You only need the repo for the docker-compose.yaml. After setup you can rm -rf it if you want.

---

Step 2 — Create the .env File

cat > .env << 'EOF'
PORT=3002
HOST=0.0.0.0
USE_DB_AUTHENTICATION=false
BULL_AUTH_KEY=CHANGEME
EOF

Set BULL_AUTH_KEY to something secure — it guards the Bull queue admin panel.

---

Step 3 — Switch to Pre-Built Images

The default docker-compose.yaml tries to build everything from source. We need to flip it to use pre-built images instead. Four services need switching:

# API (main firecrawl service)
sed -i 's|# image: ghcr.io/firecrawl/firecrawl|image: ghcr.io/firecrawl/firecrawl|' docker-compose.yaml
sed -i 's|  build: apps/api|  # build: apps/api|' docker-compose.yaml

# Playwright (browser rendering)
sed -i 's|# image: ghcr.io/firecrawl/playwright-service:latest|image: ghcr.io/firecrawl/playwright-service:latest|' docker-compose.yaml
sed -i 's|    build: apps/playwright-service-ts|    # build: apps/playwright-service-ts|' docker-compose.yaml

# Postgres (database)
sed -i 's|# image: ghcr.io/firecrawl/nuq-postgres:latest|image: ghcr.io/firecrawl/nuq-postgres:latest|' docker-compose.yaml
sed -i 's|    build: apps/nuq-postgres|    # build: apps/nuq-postgres|' docker-compose.yaml

Redis and RabbitMQ already use standard public images (redis:alpine, rabbitmq:3-management) so no changes needed there.

Verify with:

grep -n "image:\|build:" docker-compose.yaml | grep -v "^#"

You should see image: lines for: firecrawl, playwright-service, nuq-postgres, redis, rabbitmq. No build: lines should remain.

---

Step 4 — Pull the Images

docker pull ghcr.io/firecrawl/firecrawl:latest
docker pull ghcr.io/firecrawl/playwright-service:latest
docker pull ghcr.io/firecrawl/nuq-postgres:latest

Each is ~200-400MB. This takes 5-10 minutes depending on your connection.

---

Step 5 — Start Everything

docker compose up -d

Check status:

docker ps --format "table {{.Names}}\t{{.Status}}" | grep firecrawl

All 5 containers should show "Up" — RabbitMQ will show "(healthy)" once its health check passes (~30s).

---

Step 6 — Verify It Works

curl -X POST http://localhost:3002/v0/scrape \
  -H 'Content-Type: application/json' \
  -d '{
    "url": "https://example.com/",
    "pageOptions": {"onlyMainContent": true}
  }'

You should get back JSON with success: true, page content, markdown, and metadata. Credits used will be 1 (from the Postgres DB, even without auth configured).

---

Step 7 — Wire Hermes Agent to Local Firecrawl

Open ~/.hermes/config.yaml and find the web_extract section under auxiliary:

auxiliary:
  web_extract:
    provider: auto
    model: ''
    base_url: ''
    api_key: ''
    timeout: 360

Change it to:

auxiliary:
  web_extract:
    provider: custom
    model: ''
    base_url: http://localhost:3002/
    api_key: ''
    timeout: 360

Restart Hermes. Now every time Hermes needs to scrape or extract content from the web, it will use your local Firecrawl instance instead of calling an external API.

---

Useful URLs Once Running

Service	URL
Firecrawl API	http://localhost:3002/
Bull Queue Dashboard	http://localhost:3002/admin/CHANGEME/queues
Portainer (if installed)	https://localhost:9443/

---

Rebuilding After Reboot

cd ~/firecrawl
docker compose up -d

Everything persists in Docker volumes. No data is lost.

---

Troubleshooting

sed: can't read errors on macOS: macOS sed has different syntax. Use sed -i '' 's/old/new/' on macOS, or use the -i without argument on Linux.

"mount option requires BuildKit" during any build: This only matters if you're building from source. If you're using pre-built images (this guide), you never hit this. If you do need BuildKit for something else, add this to /etc/docker/daemon.json and restart Docker:

{ "features": { "buildkit": true } }

"Image denied" when pulling: Use ghcr.io/firecrawl/firecrawl (lowercase, org path) instead of ghcr.io/mendableai/firecrawl. The latter requires GitHub auth.

RabbitMQ not healthy: Wait 30-60 seconds. First startup runs migrations. Check logs with docker compose logs rabbitmq.

API returns 500: Check API logs with docker compose logs api — common cause is RabbitMQ not being healthy yet.

---

What's Working vs. Cloud

Feature	Local	Cloud
/v0/scrape (fetch)	Yes	Yes
/v0/scrape (Playwright)	Yes	Yes
/v1/crawl	Yes	Yes
/v0/extract (AI)	Yes (needs OpenAI/Ollama key)	Yes
Fire-engine (anti-bot)	No	Yes
/search	Yes (needs SearXNG)	Yes

AI features (/extract, structured output) work locally but require an OPENAI_API_KEY or OLLAMA_BASE_URL in your .env.