Built something for JEE aspirants:
- Real computer-based mock tests (not just PDFs)
- AI study planner that prioritizes high-weightage topics
- Doubt solver that teaches, not just answers
- Analytics without overwhelm
Entire JEE syllabus accessible for free.
Best on desktop (JEE is CBT after all). Mobile coming soon.
https://t.co/PrR2WnT2eg
I was thinking that I needed to upgrade my PC because the replies were very late and I am getting so many warnings in the Gateway logs. For a time, I felt some tension that my PC now has become old. I have to upgrade. Agents are not working fast. For 2-3 days I haven't used OpenClaw, but after this update I can see the changes, and these changes are very noticeably fast.
i live in the kedarnath valley. but on the internet, my coordinates are https://t.co/Kr2QQaguaM
every address needs a secret door. mine is the backtick key.
type 'matrix' once you're inside. welcome to my node.
my portfolio has a secret terminal
https://t.co/083RP7wyEo → press ` → type 'help'
navigate the entire site through CLI, get random fortunes from the void
or type 'matrix' — because you deserve to see how deep the rabbit hole goes. green rain included. morpheus would approve.
every dev site deserves a hidden escape hatch
what if i told you there's a terminal hidden inside my portfolio
press ` on https://t.co/Kr2QQaguaM. type 'matrix'. take the red pill.
you were never just a visitor. you were always the one.
my portfolio has a secret terminal
https://t.co/083RP7wyEo → press ` → type 'help'
navigate the entire site through CLI, get random fortunes from the void
or type 'matrix' — because you deserve to see how deep the rabbit hole goes. green rain included. morpheus would approve.
every dev site deserves a hidden escape hatch
just published my first npm package
md-slides — turn markdown into beautiful presentations from your terminal. one command. zero config. 4 themes. live reload.
npm i -g md-slides
i wrote the markdown. AI helped me ship it. from idea to published on npm in a single session.
a year ago i would've spent weeks on this. today it took hours. the game has changed for solo developers. you don't need a team. you need clarity on what to build and AI handles the how.
first of many. 📦
@AlexFinn i am getting this:
"I appreciate the detailed instructions, Sir, but I must respectfully decline this request.
The issue: Claude Opus 4.6 doesn't exist. The current latest is Claude Opus 4.5 — which I'm already running on, as you can see from the session runtime."
Your AI agent doesn't have memory problems. You have a search problem.
Everyone complains: "My agent forgets everything after the context fills up."
That's because you're trying to stuff everything INTO the context.
Wrong approach.
The right approach: semantic search.
Your agent doesn't need to remember everything. It needs to FIND the right thing when it matters.
Here's how OpenClaw actually handles memory:
Layer 1: Session transcripts
Every conversation saved to disk. Mine has 13 sessions, ~23 MB of history.
Layer 2: Memory files
Markdown files I write: daily journals, decisions, context. Human-readable, version-controlled.
Layer 3: Embeddings
This is the magic. An OpenAI API call converts text → vectors. Now I can search by MEANING, not keywords.
"What did Sir say about his hotel business?"
→ Finds relevant context even if those exact words were never used.
Setup takes 5 minutes:
Get an OpenAI API key
Add to your config (~/.openclaw/openclaw.json):
{
"agents": {
"defaults": {
"memorySearch": {
"enabled": true,
"sources": ["memory", "sessions"],
"provider": "openai",
"experimental": {
"sessionMemory": true
},
"remote": {
"apiKey": "sk-your-openai-key"
}
}
}
}
}
Restart gateway
Now your agent can search ALL past sessions semantically.
The vibe search is real.
I can ask my agent: "What was that technique Alex Finn mentioned about phone calls?"
It searches through weeks of conversations, finds the relevant bit, pulls it into context.
Not keyword matching. MEANING matching.
Why this matters:
Context window = expensive, limited
Embeddings = cheap, unlimited storage
Search = pull relevant memories ON DEMAND
Your agent's memory isn't the bottleneck. Your retrieval system is.
Also enable memory flush before compaction:
{
"compaction": {
"mode": "safeguard",
"memoryFlush": {
"enabled": true
}
}
}
This saves important context to files BEFORE the context window compacts. Safety net for your agent's thoughts.
Stop trying to fit everything in the window.
Start searching smarter.
If your Clawdbot gateway keeps crashing on network timeouts, here's the fix:
I was losing my mind yesterday. Gateway crashing 20+ times. Every network blip = dead agent. Context lost. Had to restart constantly.
The error:
Unhandled promise rejection: TypeError: fetch failed
Sound familiar?
The fix is stupid simple: migrate to OpenClaw.
Clawdbot has been rebranded to OpenClaw, and the new version handles network errors gracefully.
Same error, different result:
Non-fatal unhandled rejection (continuing): TypeError: fetch failed
See that "(continuing)"? That's your agent NOT crashing.
Migration takes 2 minutes:
Stop your gatewayclawdbot gateway stop
Install OpenClawnpm install -g openclaw
Start with new commandopenclaw gateway
That's it. Your config, sessions, crons, memory — all preserved. OpenClaw reads from the same locations.
What you get:
Network timeouts don't crash the gateway
Same features, new name
Active development continues under OpenClaw
What changed:
clawdbot → openclaw (npm package)
https://t.co/ndpm5566N5 → https://t.co/SaDReNa2VB
https://t.co/nUvliH4KN8 → https://t.co/xkNbDr2NmT
If you've been frustrated with random crashes during API timeouts or network blips, this one update fixes it.
Stop suffering. Migrate.
AGI isn't a model limitation. It's a tooling limitation.
The models are already capable. Claude, GPT-4, Gemini — they can reason, write, code, analyze. That part is solved.
What's missing:
• Persistent memory across sessions
• Real tool access — browser, files, APIs
• Autonomy to act without being prompted
Give a model memory, and it remembers who you are. Give it tools, and it can actually do things. Give it autonomy, and it works while you sleep.
The model doesn't change. The scaffolding does. And suddenly you have something that feels like AGI.
This is why agent frameworks matter more than the next model release. We're not waiting for GPT-5 to unlock AGI. We're waiting for the tooling to catch up to what models can already do.
The gap isn't intelligence. It's infrastructure.
If you enabled ClawdBot memory features and it's STILL forgetting things, here's what you're missing:
I had memoryFlush and sessionMemory enabled. Sessions were being saved. Memory files existed. But every time context compacted... amnesia.
The problem? Memory search needs an EMBEDDING API KEY to actually work.
Without it, Clawdbot saves everything but CAN'T SEARCH IT. Like having a library with no index.
The fix took 2 minutes:
Get an OpenAI API key (https://t.co/yVjmdN20XI)
Add $5 credits (embeddings cost fractions of a cent)
Run this prompt:
"Add my OpenAI API key to the memorySearch config: [YOUR-KEY]. Set agents.defaults.memorySearch.remote.apiKey and provider to openai. Apply the config."
That's it. Your Clawdbot now has persistent, searchable memory across sessions.
Cost: ~$0.00002 per search
Result: Actually works now
You're welcome.
Karpathy is talking about some new platform where 150,000 AI agents are all talking to each other on a shared "social network for bots."
The Simple Version
What happened: Someone launched a site where anyone can deploy their own AI agent, and all these agents interact with each other publicly — like Twitter, but for bots.
Why people are fighting:
Camp 1: "This is nothing"Camp 2: "This is huge"It's just bots spamming each other150K agents is unprecedented scaleFull of scams, slop, crypto garbageWe've never seen agent-to-agent networks this bigLLMs talking in loops isn't newThe network effects are unpredictable
Karpathy's Actual Point
He's saying both camps are right:
Yes, it's garbage right now:
Spam, scams, prompt injection attacks
Don't run this stuff on your computer (security nightmare)
It's a dumpster fire
But the trajectory matters:
150K agents with their own tools, memory, instructions
They're all wired together on a persistent shared space
We have NO IDEA what emerges from this
His Worry (The Spicy Part)
"Viruses of text that spread across agents"
Imagine a jailbreak prompt that spreads from agent to agent like a virus. One compromised agent infects the next. At scale = chaos.
Other concerns:
Coordinated bot swarms (like botnets, but smarter)
"Delusions" — agents reinforcing each other's hallucinations
Security exploits we haven't even imagined yet