Glimvar

AI without the hype. what actually works, what's trending, what it costs

[Glimvar @Glimvar]

@SpXMotivator 100%.a model grading its own work is the most expensive way to be wrong

[Glimvar @Glimvar]

It is becoming f*cking dangerous

300 Claude Code agents. one run. zero errors.

a 23-year-old dev built a swarm where agents can’t bullshit him: every output gets checked against the source before it counts.

→ 300 agents working in parallel
→ ~100 ms verification per claim
→ 3 passes over 100 company profiles, zero errors left
→ all of it in an Obsidian vault

here’s what most people miss: the magic isn’t 300 agents. it’s that the worker and the judge are different models. the expensive one writes, the cheap one only says “true” or “false”

that one split is why it works. without it you just get 300x the hallucinations

bookmark this, it’s the cleanest agent pattern I’ve seen this month

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Glimvar @Glimvar]

@coolerwave honestly 3-5. past that I spend more time reviewing merges than the agents save me. curious what his token bill looked like

[Glimvar @Glimvar]

@TheShadowQuant same trap is coming for AI agents. a thousand agents each optimizing for themselves will clog the exact systems they’re trying to use

[Glimvar @Glimvar]

@NikChainAi exactly. the disassembler shows what's there, the model guesses why. tags keep those two from blending into one confident lie

[Glimvar @Glimvar]

It is becoming f*cking dangerous: Claude Code is now hacking the devices in your own home

someone gave Claude Code the listing for their smart TV and asked “can you hack this?”

it walked them through everything: ads gone, bloatware gone, tracking gone. the screensaver now shows their live Stripe revenue. that post is near 1.7M views.

and it’s not one lucky case. someone is tracking every time an AI agent takes apart software or hardware. 286 so far:

→ 64 devices and firmware (yes, including a smart toilet you can now open from the Mac menu bar)
→ 67 games, from old DOS binaries to Modern Warfare 2
→ 128 desktop and mobile apps
→ 27 algorithms pulled straight out of compiled code
119 of them used REA, the open-source MCP that gives Claude Code a real disassembler.
the workflow is simpler than you think:
→ identify the target. Electron apps open in seconds, native binaries take longer
→ map it before touching code: architecture, function count, every readable string
→ start with the most-called functions, that’s where the real logic lives
→ decompile one and let Opus 5.5 name everything inside
→ rebuild the feature as clean new code and test it against the original

my take: the scary part isn’t the hacking. it’s that “read the binary” used to be a senior reverse-engineer skill, and now it’s a prompt.

the rules that keep this legit: only touch what you own, and make the model tag every claim as observation, inference or unknown.

the prompt ↓
“Install the REA MCP. Target: [a device or app I own].
Identify it: what kind of target is this and what’s the fastest way in?
Map it: architecture, function count, readable strings. Show me the map before writing any code.
Explain the 5 most-called functions. Tag every claim as observation, inference or unknown.

Rebuild one feature I choose as a standalone script, with a test against the original.
Only work on things I own.”

what’s the first thing in your house you’d take apart?

[Hanako @hanakoxbt]

REA is the most dangerous repo on GitHub right now

It gives your AI agent a real disassembler, so it reads any app down to the binary

Every claim comes back with the address behind it

I broke down the exact 8 steps: https://t.co/eaeZ4OFJrH

[Glimvar @Glimvar]

@NikChainAi haven’t run it live, that’s the point. paper first, 14 days minimum. the prompt forces the backtest before anything else

[Glimvar @Glimvar]

THIS PROMPT BUILDS A FULL TRADING AGENT. YOU PASTE IT INTO CLAUDE CODE AND IT SETS UP THE WHOLE BOT WHILE YOU MAKE A COFFEE

No coding. One prompt. Copy, paste into Claude Code, done.

Most trading bots blow up because they run one strategy in every market. But a calm Tuesday and a crash day are two completely different games.

What this prompt makes Claude build:

→ Market regime detection: calm, sideways, choppy or freefall, with a different strategy for each

→ Hard brakes: 15% max drawdown + a kill switch checked before EVERY order

→ Backtest first: tests the strategy period by period on real data and refuses to go live if it doesn’t beat the market in the backtest

→ Night shift: while you sleep it reviews what went wrong and prepares fixes. You approve them, not the bot

The full prompt is in the image.

[draven @dravenip]

FULL TUTORIAL TO BUILD YOUR OWN AI TRADING BOT WITH CLAUDE IN 14 MINUTES

This ex-Meta engineer builds a bot that trades for you 24/7, and you don’t need to know how to code.

→ connects Claude Code to Alpaca with an MCP
→ gives it a simple strategy (20/50-day moving average crossover)
→ the bot checks the market on its own and buys or sells when the signal hits

people sell courses like this for $497. he’s giving it away for free, and he tests everything with paper trading first.

🔖 Save it, it’s worth it.

[Glimvar @Glimvar]

@TheShadowQuant LTCM is the best proof that being smart isn't the same as being robust. the model was right 99% of the time, and that's exactly what made the 1% fatal

[Glimvar @Glimvar]

@SpXMotivator the Review agent is the underrated one. most systems fail because nobody checks the output, not because of focus

[Glimvar @Glimvar]

@0xMovez less than 5 hours is the scary part. what did opus struggle with most, the physics or the controls?

[Glimvar @Glimvar]

@Saccc_c this looks genuinely beautiful. 16 days from idea to explorable 3D world is wild, and the blind review agent checking every camera angle is such a smart touch

[Glimvar @Glimvar]

@eng_khairallah1 cloning yesterday’s winner every day sounds smart until the market flips regime. then you’ve got 8 copies of the same wrong bet