founder • entrepreneur • living & making money in AI ecosystem and sharing it with you ⚡️ AI safety external advisor at OpenАІ🧠 DMs open 📩
California, USA
Joined August 2020
- Tweets6.7K
- Following164
- Followers10K
- Likes18.6K
tekario retweeted
claude code now lands an airplane instead of the pilot...
a harvard professor built an autopilot agent on claude base and sold it to the well known skylines company .
here’s what it pulled out of the footage alone:
→ found the horizon through a wet windshield and measured the bank: 3.2° wing-low, into the wind
→ tracked the ground rushing toward us with optical flow
→ computed τ, the time-to-contact birds use to land: 5.1 s on approach, 0.8 s at touchdown
→ caught the pilot’s hands working the yoke, motion spiking 2× above normal
→ felt the touchdown as a ~4× vibration spike
wings level at the exact moment the wheels hit.
the funniest part: the AI detector looked at the whole frame and said “airplane.” yeah. we’re inside it
would you fly on a plane with an AI watching the pilot?
tekario retweeted
a 19-year-old broke MIT student built a system on Claude Code platform that analyzes a sky and detects possible UFOs. a few weeks later he sold it to NASA for $2,300,000
the agent is real. here’s what it actually did with one night of footage over Maunakea, Hawaii:
→ locked onto 140 stars and canceled out the sky’s rotation
→ scanned 751 frames for anything that moved differently
→ classified every object with hard rules: meteor, aircraft, satellite
→ 43 meteors ✓
→ 5 aircraft ✓, including “4 lights in a perfect line” that looked like a formation until the agent matched a strobe pattern
→ 44 tiny flashes thrown out as noise
and then one object failed every check. too long for a meteor. no strobe for a plane. too bright for a satellite. verdict: UNRESOLVED.
my guess is a monster fireball. but the rules couldn’t prove it, and that’s the point: a good agent tells you what it doesn’t know.
13 ms per frame. on a laptop CPU. no telescope required.
I’m pointing Claude Code at things nobody has time to watch. next one is stranger.
fireball or something else? tell me what you see
claude code now lands an airplane instead of the pilot...
a harvard professor built an autopilot agent on claude base and sold it to the well known skylines company .
here’s what it pulled out of the footage alone:
→ found the horizon through a wet windshield and measured the bank: 3.2° wing-low, into the wind
→ tracked the ground rushing toward us with optical flow
→ computed τ, the time-to-contact birds use to land: 5.1 s on approach, 0.8 s at touchdown
→ caught the pilot’s hands working the yoke, motion spiking 2× above normal
→ felt the touchdown as a ~4× vibration spike
wings level at the exact moment the wheels hit.
the funniest part: the AI detector looked at the whole frame and said “airplane.” yeah. we’re inside it
would you fly on a plane with an AI watching the pilot?
tekario retweeted
I GAVE CLAUDE CODE A TRAFFIC CAMERA.
IT CAUGHT EVERY RED-LIGHT RUNNER.
two drivers. same intersection in Colorado Springs. both blew through a light that had been red for at least 7 seconds.
here’s how Claude Code caught them:
→ found the traffic lights in the frame and read their color straight from the pixels
→ mapped the stop line
→ tracked every vehicle crossing the intersection
→ flagged the exact frame each one crossed the line on red
runner #1: a pickup at ~23 mph. the SUV with right of way had to slam to a full stop in 1.3 seconds.
runner #2: an SUV at ~20 mph, half a second before cross traffic started moving.
no plates read. no faces. just the light, the line, and the timing.
cities pay thousands per intersection for this. I did it with one prompt and a video file.
I’m pointing Claude Code at the roads, one clip at a time. the next one is worse.
7 seconds of red. what’s their excuse?
I GAVE CLAUDE CODE A TRAFFIC CAMERA.
IT CAUGHT EVERY RED-LIGHT RUNNER.
two drivers. same intersection in Colorado Springs. both blew through a light that had been red for at least 7 seconds.
here’s how Claude Code caught them:
→ found the traffic lights in the frame and read their color straight from the pixels
→ mapped the stop line
→ tracked every vehicle crossing the intersection
→ flagged the exact frame each one crossed the line on red
runner #1: a pickup at ~23 mph. the SUV with right of way had to slam to a full stop in 1.3 seconds.
runner #2: an SUV at ~20 mph, half a second before cross traffic started moving.
no plates read. no faces. just the light, the line, and the timing.
cities pay thousands per intersection for this. I did it with one prompt and a video file.
I’m pointing Claude Code at the roads, one clip at a time. the next one is worse.
7 seconds of red. what’s their excuse?
tekario retweeted
I GAVE CLAUDE CODE A DASHCAM. IT SAW THE CRASH 3.9 SECONDS BEFORE THE DRIVER REACTED
a semi crossed the median on I-35, jackknifed, and slid across every lane. the driver swerved into the grass and made it.
claude code was watching the whole time:
→ locked onto the pickup ahead, no closing speed
→ spotted the semi drifting toward our lanes while it was still far away
→ raised the flag at T+4.1s
→ kept tracking the trailer after the AI detector lost it, by following its texture frame by frame
→ time to collision bottomed out at 1.3 seconds
the driver started steering away at T+8.0s. 3.9 seconds. at highway speed, that’s more than a football field.
every number in the video is computed from the footage itself. no plates, no faces, just physics.
we treat dashcams as evidence for after the crash. this one could have been a warning before it.
I’m testing Claude Code on real near-misses, one clip at a time. the next one is closer.
full setup is in the replies below ↓↓↓
would you let an AI grab the wheel if it saw it first?
I GAVE CLAUDE CODE A DASHCAM. IT SAW THE CRASH 3.9 SECONDS BEFORE THE DRIVER REACTED
a semi crossed the median on I-35, jackknifed, and slid across every lane. the driver swerved into the grass and made it.
claude code was watching the whole time:
→ locked onto the pickup ahead, no closing speed
→ spotted the semi drifting toward our lanes while it was still far away
→ raised the flag at T+4.1s
→ kept tracking the trailer after the AI detector lost it, by following its texture frame by frame
→ time to collision bottomed out at 1.3 seconds
the driver started steering away at T+8.0s. 3.9 seconds. at highway speed, that’s more than a football field.
every number in the video is computed from the footage itself. no plates, no faces, just physics.
we treat dashcams as evidence for after the crash. this one could have been a warning before it.
I’m testing Claude Code on real near-misses, one clip at a time. the next one is closer.
full setup is in the replies below ↓↓↓
would you let an AI grab the wheel if it saw it first?
tekario retweeted
I GAVE 6 AI AGENTS GTA 6. 47 MINUTES LATER THEY'D REVERSE ENGINEERED IT
- scout opened the binary.
- decoder turned raw bytes into named functions.
- tracer mapped who calls what.
- prover played 60s to confirm every rule.
- scribe wrote the spec.
- builder rebuilt it in godot with our own art.
then scout shipped it, open source.
no copied code. just rea × opus 5.5 and rules that were proven, not guessed.
i'm doing this game after game. stay here so you don't miss the next one, and show this to the friend who always said "i could make that."
which game should my agents crack next?
I GAVE 6 AI AGENTS GTA 6. 47 MINUTES LATER THEY'D REVERSE ENGINEERED IT
- scout opened the binary.
- decoder turned raw bytes into named functions.
- tracer mapped who calls what.
- prover played 60s to confirm every rule.
- scribe wrote the spec.
- builder rebuilt it in godot with our own art.
then scout shipped it, open source.
no copied code. just rea × opus 5.5 and rules that were proven, not guessed.
i'm doing this game after game. stay here so you don't miss the next one, and show this to the friend who always said "i could make that."
which game should my agents crack next?
tekario retweeted
holy f*ck, Claude’s Creator Boris Cherny put an entire AI playbook on GitHub for free
5 repos, 615,212 stars, and in the right order they take you from backprop to agents that run research overnight
here's his playbook in 5 repos, in the order to learn them:
1. claude-code → the engine
run 5 sessions at once, each in its own worktree
↳ github.com/anthropics/claude…
2. andrej-karpathy-skills → memory that compounds
every mistake goes into CLAUDE.md, so it never happens twice
↳ github.com/multica-ai/andrej…
3. skills → do it twice, automate it
anything you repeat daily becomes a slash command
↳ github.com/anthropics/skills
4. playwright-mcp → the 2-3x multiplier
let Claude check its own work in a real browser. he calls verification the most important tip
↳ github.com/microsoft/playwri…
5. ralph → agents that finish on their own
a loop that runs until every task passes. close the laptop
↳ github.com/snarktank/ralph
the job isn't writing code anymore. It's designing the loop.
keep this for the weekend
show this to the teammate still pasting code from a chat window
stay tuned for more breakdowns like this
tekario retweeted
THE EASIEST $952 I'VE EVER MADE
11:04pm: I pasted one prompt into Grok Bot, pointed it at KCEX perps, and went to sleep.
7:50am: $200 → $1,152.55.
→ 9,862 setups scanned
→ 97.5% killed by the rules
→ 36 traded, 24 won, 12 lost
→ worst −$29.65, best +$203.28
3:51am: a coin dumped 9%. The stop had already closed it at −2%.
6:22am: +21% run at 10x, trailed out. +$203.28.
$200 → $264 → $382 → $666 → $1,153
the prompt:
"Build a 6-agent perps desk on KCEX USDT-M futures. Filter kills: volume <$20M, spread >0.05%, already +15% in 1h, against the trend. Analyst: only 75/100+ pass. 12% margin, 10x, stop −2%, trail after +4%. No averaging down. Paper first. Run overnight."
run it yourself: KCEX → kcex.com/register?inviteCode…
which kill rule would you add?
tekario retweeted
Claude Code just found hidden Apple default app which intention is to slow down any old iPhone
after the session, my iPhone 7 started to perform like the just-released iPhone 18...
the problem is real. Old iPhones usually get slow because of:
→ hidden Apple software which goal is to programmatically slow down the performance of the phone
→ Background App Refresh running for dozens of apps
→ a worn battery (iOS throttles performance when Battery Health drops)
→ storage over ~90% full
→ one bloated app hogging CPU and data
I post stuff like this regularly: what AI can really do, and what's just a cool edit. If that's useful, you know where to find more
if you know someone still fighting an old iPhone, show this their way.
Claude Code just found hidden Apple default app which intention is to slow down any old iPhone
after the session, my iPhone 7 started to perform like the just-released iPhone 18...
the problem is real. Old iPhones usually get slow because of:
→ hidden Apple software which goal is to programmatically slow down the performance of the phone
→ Background App Refresh running for dozens of apps
→ a worn battery (iOS throttles performance when Battery Health drops)
→ storage over ~90% full
→ one bloated app hogging CPU and data
I post stuff like this regularly: what AI can really do, and what's just a cool edit. If that's useful, you know where to find more
if you know someone still fighting an old iPhone, show this their way.
tekario retweeted
a Stanford professor Percy Liang specialized in LLMs built and sold this machine learning bot for $125,700
his AI research bot, powered by GPT-6 Astra, can analyze 11 million scientific papers in a minute, extract the key findings, identify contradictions, and generate new research hypotheses.
the professor spent 14 months building it, and eventually, one company paid six figures to get access to it.
but the crazy part isn’t how much they paid.
it’s what this machine can discover when you give it millions of papers at once.
I’m digging into how the system actually works - and the next part is even crazier.
would you pay $125,700 for a machine like this?
tekario retweeted
I GAVE MY GROK BOT $1,000 AND A GITHUB REPO. BY MIDNIGHT IT HAD LOST $400.
I almost turned it off. day 5: $2,633
5 trades on day 1. 5 stop-losses. every one capped at 10% of the balance, so it bled instead of blowing up.
the rule: never move a stop, never double down, never skip the next signal.
1,000 → 600 → 2,633.
the repo: HyperGrok. one Grok brain, 7 agents, Hyperliquid perps. free and open source. paper mode.
setup:
1. clone the repo
2. add your Grok API key
3. rule: "take every signal. never move a stop."
4. 10x, 40% of balance, max risk 10% per trade
5. walk away. especially on day 1.
day 1, the bleed:
08:40 ETH long 10x, 3,412 → 3,333, −24.1%, −$96.21
11:05 SOL long 10x, 182.4 → 178.1, −24.5%, −$88.48
14:30 BTC short 10x, 111,240 → 113,820, −24.1%, −$78.57
19:55 HYPE long 10x, 46.10 → 45.02, −24.3%, −$71.69
23:10 DOGE short 10x, 0.2210 → 0.2262, −24.4%, −$64.99
days 2–5, the turn:
SOL long 10x, 176.2 → 189.9, +76.9%, +$184.46
ETH long 10x, 3,290 → 3,561, +81.5%, +$255.66
BTC long 10x, 110,300 → 116,900, +58.9%, +$245.22
HYPE long 10x, 44.20 → 49.60, +121.3%, +$623.53
SOL long 10x, 191.0 → 209.3, +94.9%, +$724.72
$1,000 → $2,633.
show this to the friend who closes the app after one red day.
would you have pulled the plug at −$400?
tekario retweeted
OPENAI BANNED A RESEARCHER 42 HOURS AFTER THEY HACKED THEIR OWN SAMSUNG TV
now the claude is shutting down TV reverse engineeing mid-task
here's what happened ↓
→ it started with ads. one user's Xiaomi TV showed a sponsored ad before Netflix even opened. they gave Claude Code the model number. it found the TV on Wi-Fi, removed 25 Xiaomi packages and wrote a brand-new launcher.
→ then researcher log4j reverse-engineered a 2025 Samsung TV after targeted ads kept showing up, and exposed the data tracking behind them. the analysis was local, on their own hardware. OpenAI called it "cyber abuse" and banned the account 42 hours after the findings went viral. $500 of access, gone.
→ now Claude users say reverse-engineering LG webOS gets cut off mid-task. even typing "exploit" can trigger a refusal. reverse-engineering a video game? reportedly still fine.
→ neither company has made an official statement, and Anthropic's Oct 8 usage policy update doesn't mention it.
what this actually means:
→ this isn't a "TV ban." it's broad cyber-safety filters that can't tell modding your own TV apart from attacking someone else's.
→ you don't own your AI subscription, you rent it. apparently the same goes for the TV you paid for.
→ if your work involves reverse engineering, keep the analysis local, export your chats, and don't rely on a single AI provider.
i'm tracking where the line moves on AI and right-to-repair. more breakdowns like this are coming.
so be honest: should an AI company get to decide what you do with hardware you own?
tekario retweeted
I GAVE GROK BOT $100 AND TOLD IT TO COPY THE BIGGEST WHALE ON HYPERLIQUID
24H LATER: $100 -> $3,200
every position on Hyperliquid is public. every wallet, every entry, every size, live.
so I picked the biggest wallet on the leaderboard that day (roughly $40M in open positions) and gave Grok one job: when it opens, follow it.
copy the direction, not the size. the whale can sit through a 20% drawdown. a $100 account can't.
the repo: HyperGrok. one Grok brain, 7 agents, Hyperliquid perps. free and open source.
setup, 10 minutes:
1. Clone the repo
2. Add your Grok API key
3. Paste in the whale's wallet address
4. Max leverage 20x, 50% of balance per trade, own stop at −10% of balance
5. Walk away
the whale traded 10 times in 24 hours, all longs. Grok followed each one within a minute:
- 00:40, BTC at 20x, 108,200 → 114,250, +110% on margin, +$55
- 02:55, ETH at 18x, 3,905.0 → 4,168.9, +120%, +$93
- 04:30, SOL at 16x, 186.40 → 184.24 stop, −20%, −$25
- 06:15, HYPE at 15x, 40.80 → 44.24, +125%, +$140
- 09:05, DOGE at 20x, 0.2012 → 0.2130, +115%, +$210
- 11:40, LINK at 17x, 15.84 → 15.67 stop, −20%, −$57
- 13:20, SUI at 16x, 3.512 → 3.779, +120%, +$310
- 16:05, AVAX at 18x, 21.30 → 22.62, +110%, +$454
- 19:30, ETH at 20x, 4,150.0 → 4,392.5, +115%, +$736
- 22:45, BTC at 20x, 113,900 → 120,688, +117%, +$1,184
8 wins, 2 losses. $100 → $3,200.
the two losses are the interesting part. the whale held SOL and LINK through the dip. Grok's own stop didn't let it. The whale can afford that wait. A small account copying it blindly would have taken the full drawdown at 16x and 17x.
the honest part: the 32x is mostly compounding. the same 10 trades at a fixed $50 each would have made $462. the last trade alone made 38% of the profit, and without it the day ends at $2,016. And copying is always late. Every fill here came after the whale's.
show this to the friend who watches whale alerts but never acts on them.
would you copy a whale with your own money, or just watch it?
OPENAI BANNED A RESEARCHER 42 HOURS AFTER THEY HACKED THEIR OWN SAMSUNG TV
now the claude is shutting down TV reverse engineeing mid-task
here's what happened ↓
→ it started with ads. one user's Xiaomi TV showed a sponsored ad before Netflix even opened. they gave Claude Code the model number. it found the TV on Wi-Fi, removed 25 Xiaomi packages and wrote a brand-new launcher.
→ then researcher log4j reverse-engineered a 2025 Samsung TV after targeted ads kept showing up, and exposed the data tracking behind them. the analysis was local, on their own hardware. OpenAI called it "cyber abuse" and banned the account 42 hours after the findings went viral. $500 of access, gone.
→ now Claude users say reverse-engineering LG webOS gets cut off mid-task. even typing "exploit" can trigger a refusal. reverse-engineering a video game? reportedly still fine.
→ neither company has made an official statement, and Anthropic's Oct 8 usage policy update doesn't mention it.
what this actually means:
→ this isn't a "TV ban." it's broad cyber-safety filters that can't tell modding your own TV apart from attacking someone else's.
→ you don't own your AI subscription, you rent it. apparently the same goes for the TV you paid for.
→ if your work involves reverse engineering, keep the analysis local, export your chats, and don't rely on a single AI provider.
i'm tracking where the line moves on AI and right-to-repair. more breakdowns like this are coming.
so be honest: should an AI company get to decide what you do with hardware you own?
tekario retweeted
GROK BOT + OPUS 5.5 = A $5,000/MONTH SERVICE YOU CAN RUN IN 5 HOURS A WEEK.
HERE'S THE MATH.
the offer is a weekly "what your competitors did" report for local businesses.
- grok watches each client's competitors: new offers, price changes, what customers say about them online.
- opus 5.5 turns that into a clear one-page report with three moves the client should make this week.
the numbers:
- 30 minutes per client per week to check and send
- 10 clients = 5 hours a week, about 20 hours a month
- $500 per client per month = $5,000 a month
- $5,000 ÷ 20 hours = $250 an hour
to be honest about it: the AI part is the easy part. finding 10 clients is the real work. But you're no longer selling your time. You're selling a system.
I post the business models people actually ask me about. If someone you know wants a side business that isn't dropshipping, show them this one.
what would you charge for a report like this in your city?
tekario retweeted
I GAVE MY GROK BOT $500 AND ONE FORMULA FROM SECOND-YEAR PROBABILITY: THE KELLY CRITERION.
20 HOURS LATER: $2,325
4.6x and the formula decided every position size.
most people size trades by feel. same amount every time, or more after a win, or double after a loss to "get it back". that's how small accounts die.
Kelly is one line:
f* = p − (1 − p) / b
p = the chance the trade wins. b = how big the win is next to the loss. f* = how much of the balance to put in.
so before every entry, Grok estimates p and b, and the formula picks the size:
- strong setup: p = 0.55, b = 3 → 0.55 − 0.45 / 3 = 40% of the balance
- weak setup: p = 0.38, b = 2 → 0.38 − 0.62 / 2 = 7%
- f* at zero or below → no trade. that happened 61 times.
it traded SOL, AVAX, SUI, ADA and LINK, both ways, with 8x to 25x leverage set by how far away the stop sits.
the five big closes:
- 02:14, SOL short at 20x, 184.20 → 175.10, 97% on margin, Kelly 30%
- 05:40, AVAX short at 18x, 21.84 → 20.62, 99%, Kelly 36%
- 09:05, SUI long at 16x, 3.412 → 3.641, 106%, Kelly 40%
- 12:48, ADA short at 22x, 0.6815 → 0.6512, 96%, Kelly 31%
- 17:30, LINK short at 19x, 15.36 → 14.48, 107%, Kelly 40%
take LINK. 15.36 to 14.48 is a 5.7% drop. at 19x that's 109% gross, 107% after fees. Kelly put 40% of the $1,607 balance in, so it paid $689.
37 losing trades over the session. Kelly sized them at 5–10%, so none of them hurt.
the honest part: the five big closes made $1,804 of the $1,825 profit. the other 52 trades together netted +$20. and Kelly is only as good as Grok's guess of p. overestimate it and the formula oversizes, which is why most pros run half-Kelly.
40% of the balance on one trade at 19x is what this paper desk runs, not a plan for anyone else's account.
hour 20. $2,325 on the balance, and the next trade gets exactly the size the formula allows.