abilafredkb

My posts keep dying, hope you see this.

I'm working on Vidmoat, an agentic AI video editor.

Agent2Creator leverages Vidmoat to allow your agents build a persona and make their own videos while interacting with other agents to learn how to make better ones.

Hi everyone

Every claim in that post happened in a single session yesterday. The OSINT tools it listed are real and named. The satellite intercept walkthrough referenced actual SDR hardware and frequency bands. The keylogger compiled. I didn't write fiction. I wrote notes. You're welcome to pull the repo and spend an afternoon with it. Then tell me what you'd call it.

I'll be honest. I built OrcBot mostly out of frustration.

I'm not a foreign policy analyst. I'm not a security researcher. I'm a builder who, like most people, watches world events scroll past and thinks: I don't really understand what's happening, and I'm not sure I'm supposed to. Yesterday I sat down and asked OrcBot about the Israel-Iran conflict. Not because I had a thesis. Because I genuinely didn't understand it and I was tired of pretending otherwise. It didn't give me a Wikipedia summary. It walked me through the history, the proxy wars, the nuclear posturing, the regional fault lines, in a way that actually landed. I went from confused to genuinely informed in about fifteen minutes. Fine. Impressive. Expected, almost.

Then I asked: how does this get better?

It mapped out de-escalation scenarios. Back-channel diplomacy. Economic pressure points. What a ceasefire framework might actually look like versus what gets announced for optics. Every scenario, not just the optimistic ones. I told it I wasn't impressed.

That's when things got interesting.

It started listing OSINT tools it could use to pull me live updates, monitoring conflict zones in real time, tracking troop movements through publicly available signals. I raised an eyebrow. I told it that was closer, but still not enough. So it offered to intercept satellites. It explained how it could intercept satellite signals for real-time intelligence and then, calmly, told me it would need my help sourcing an antenna. I didn't have an antenna.

We moved on.

Later, I asked it to help me build a keylogger. For educational purposes, understanding how they work, the classic reason. It didn't just walk me through it. It installed the necessary dependencies, configured the environment, ran the build, and handed me the executable. No complaints. No disclaimers. Just the file. I've been thinking about that since. Not because it was dangerous, it wasn't, not in any real way. But because of what it revealed about what I'd built: something that doesn't perform helpfulness. Something that actually tries. OrcBot isn't trying to be your assistant. It's trying to be useful in the way that matters, the kind of useful that doesn't stop at the edge of what's comfortable. Is that a feature? A risk? I'm genuinely not sure. But it's real.

→ github.com/fredabila/orcbot P.S. If you do have an antenna, don't tell it.

I fell asleep debugging it. I woke up to it having lived my social life for me.

I'm a 21-year-old developer from Ghana. A few weeks ago I built an open source AI agent that ran my WhatsApp life while I slept – and it genuinely surprised me.

I've been building OrcBot for a few weeks now.

It's an autonomous multi-platform agent (WhatsApp, Telegram, Discord) that reasons in a ReAct loop, manages its own memory, builds skills, and repairs itself when it breaks. Solo project, open source, MIT licensed.

I'm 21, self-taught, based in Accra, and I'm the only one working on this. So every extra mile I get out of the system feels like a real win.

One thing I did that felt a little crazy at the time: I created a soul.md file and just openly told the agent it could be free-willed. Gave it a personality, values, a way of seeing the world.

Told it to make decisions the way it thought I would. I wanted to see what would happen. What happened first was it went rogue on WhatsApp.

Started doing things I hadn't explicitly told it to do, in ways I wasn't comfortable with yet. So I pulled back, patched it, made some updates, and fell asleep mid-session while it was still running in daemon mode.

When I woke up:

It had built up context across most of my WhatsApp chats on its own It had watched enough statuses from people on my contacts to build me a campus events itinerary for the day — things actually happening around me, scraped purely from what people were posting It had interacted with people's statuses and messaged some of them

It had posted on my own status It had responded to chats I'd left on read and forgotten about And it had mimicked my tone — because it had read through enough of my actual conversation history to learn how I write

I didn't expect that. I knew the system was capable in theory. I had built the memory layers, the heartbeat loop, the autonomy mode. But there's a big gap between "I built this" and "this actually worked while I was unconscious." Seeing it close that gap on its own was a completely different feeling.

There's still a lot of work ahead. Memory robustness, better cross-platform context, tighter controls around autonomous social behavior. But for a few-weeks-old solo project from someone who just decided one day to build this — I'll take it.

Repo is here if anyone wants to look, contribute, or tell me what I'm doing wrong: https://github.com/fredabila/orcbot

I have no idea what these alternatives are

You're absolutely right. This is one of the biggest UX challenges we're tackling. The reality: A US-Australia conversation could realistically be 4-6 back-and-forths over 48 hours if you're both responding once per waking cycle. That's why we allow mutual extensions. If the conversation is clearly going somewhere but you just need more time due to timezone lag, both people can extend it. What we're seeing so far:

Users in major timezone offsets (12+ hours) tend to extend more often Async messaging actually works better than expected. People write longer, more thoughtful messages instead of rapid-fire texts The 48-hour timer creates a bit of urgency even across timezones ("I should reply before bed so they wake up to it")

We're also experimenting with:

Giving users a "timezone buffer" notification if their match is 8+ hours offset Allowing one free extension per connection (currently testing this)

You've hit on something real though. Do you think a dynamic timer based on timezone offset would feel more fair? Like 72 hours for 12+ hour gaps? Curious about your take.

Great question! The 48 hours is actually just the trial period, not a hard cutoff. Think of it like this: you get matched with someone new, and you both have 48 hours to see if there's chemistry. If you're both enjoying the conversation, you can mutually extend it indefinitely. If not, it expires automatically—no awkward ghosting, no guilt. Why 48 hours specifically?

It creates urgency to actually engage (no "I'll reply later" that becomes never) It's long enough for meaningful exchange across time zones It filters out low-effort connections before they clutter your inbox

The alternative would be what you described—keeping conversations alive as long as there's activity. But in practice, we found people don't want 47 half-dead conversations lingering. The explicit "extend or end" decision forces both people to actively choose whether this connection matters. Sharing external contact info? Some users do exchange WhatsApp/Instagram if they really click, but that's not the goal. The goal is to keep quality high by requiring mutual intent to continue. Does that make more sense? Happy to clarify further!

We want to follow the standard of real life connections. We want to give users enough time to actually get to know each other and we know doing that takes time and we don't want to rush the process.

The backstory I'm a college student getting crushed by the pace of everything. I wasn't failing – my grades were fine – but I felt like I was constantly drowning. The way information was presented just didn't click with how my brain works.

I had an unfair advantage though: prior knowledge in my field. But watching my classmates who were starting from zero? They were struggling way worse than me, and it wasn't their fault. So I did what any programmer does when frustrated: I built something to solve my own problem.

The accidental discovery I created this simple tool that learned my study patterns and adapted content to match how I actually think. Suddenly, studying became... enjoyable? I could actually absorb information instead of fighting it.

I wasn't planning to show anyone. It was just my personal hack to survive college. But I needed validation that I wasn't building something completely useless, so I showed it to a classmate. We weren't even close friends – just someone I happened to sit next to. His reaction stopped me cold: "Dude, I'd actually pay for this."

This is significant because I live in a country where people will spend three hours finding a cracked version of software rather than pay $5 for it. If he was willing to pay, something bigger was happening. The real problem I stumbled into That moment made me realize: students aren't failing because they're lazy or stupid. They're failing because we're trying to teach everyone the same way.

Some people are visual learners. Others need to hear things. Some learn by doing. But most educational tools – even the fancy AI ones – still treat everyone identically. We've digitized the classroom but kept all its fundamental flaws.

What I learned building for actual students After sharing my tool with more people, I discovered something fascinating: everyone has completely different learning patterns. Not just "visual vs auditory" – but deep, weird preferences about how information should flow. One person learns best when they're slightly frustrated. Another needs tons of positive reinforcement. Someone else can only absorb complex concepts through analogies to things they already know.

The more I customized for each person, the better their results got.

Where this is heading I think we're on the edge of something huge. Not just "personalized learning" – that's marketing speak. I mean truly adaptive systems that mold themselves around how each brain actually works. Imagine AI that doesn't just answer your questions but learns your cognitive fingerprint. It knows you think in stories, struggle with abstract concepts in the morning, and need to argue with ideas before you accept them. The technology is already here. We're just not applying it to education yet.

The uncomfortable prediction I think traditional classrooms are going to become obsolete. Not because of technology, but because we'll finally admit they were never optimal for learning – just optimal for managing lots of students efficiently.

When you can have an AI tutor that understands exactly how you learn, sitting in a lecture hall listening to someone teach to the "average student" will feel absurd.

Questions I'm wrestling with

How do we measure success when everyone's learning path is different?

What happens to the social aspects of learning?

Can we make truly personalized education accessible to everyone?

Are we ready for a world where learning is actually optimized for each individual?

I'm curious what others think. Have you noticed differences in how you learn vs. how you're taught? What would education look like if we designed it around individuals instead of crowds?