Olivia
Online
Seinz
@0xSeinz
Every fortune is an integral. I study the edge behind it: AI, algorithms and the minds that made them pay
github:
Joined January 2025
70 Following
58 Followers
349 Posts
this is getting out of hand: a post about fixing someone's parents' TV hit 5.2M views and almost 22,000 bookmarks in under 3 days ↓
Everyone has these parents. A TV that opens on a sponsored row, a remote with too many buttons, and a call to you every time HDMI 2 goes missing.
The cleanup gets the attention. The part parents actually live with is the home screen, and one developer who did this for their own parents with Claude Code published exactly what went on it:
1. a big clock
2. their streaming apps, nothing else
3. the HDMI inputs, one click each
4. Settings
That's the whole screen. It uses about 12 MB of RAM and the TV boots straight into it.
Google Play, the keyboard and Bluetooth for the remote all stayed on, and everything that got turned off comes back with one command.
What's on your parents' home screen right now?
full guide with every command and its undo in the reply ↓
Vibe coded my parents' TV this afternoon!
It was showing a sponsored ad before you could even get to Netflix.
Gave Claude Code the model number, it found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher.
Looks like an Apple TV now lol
Fields medalist Terence Tao on long AI chains:
This course chains 9 Grok bots that hand files to each other. Tao says one wrong step in a long chain breaks the whole thing.
P(clean week) = p^9
→ each handoff right 99% of the time: 91% of weeks run clean
→ each handoff right 95% of the time: 63%
At 95%, more than 1 week in 3 has a bad handoff somewhere.
The human gates in this setup cut that down: each one catches a bad file before the next bot builds on it.
How many handoffs does your setup have?
I break down the math behind AI and money, with the code. Follow for more.
https://t.co/sVlvM5N7Xb
@coolerwave Most people do the opposite: they let cheap models hallucinate, then waste human or frontier-model hours debugging. Automating the rejection loop before code ever lands in a PR is the real unlock here
@vayrenrose Exactly. A batch of 10 leaves every app as risky as before and turns a broken TV into 10 suspects instead of 60
holy f*ck, Claude Code turned a 2017 TV into a new one. One rule decides whether yours goes black ↓
A viral clip (2.9M views) shows a TV flying through its menus after Claude cleared out the bloatware. With REA, the reverse-engineering guide below, it takes one prompt, even on a 2017 TV. One developer turned off 25 apps and swapped the 192 MB stock launcher for an 83 MB one in a single afternoon.
Some of that "junk" holds up things you need. Turn off the wrong package and HDMI switching or the home screen goes with it.
Agent Math #2: stacked steps
Say each app is 99% safe to switch off. Turn off n apps at once:
P(nothing breaks) = 0.99ⁿ
25 apps → 78%
60 apps → 55%
100 apps → 37%
At 60 apps, 45% of TVs lose something: HDMI, sound, the remote. And 1 TV in 8 loses two things at once.
The one rule: 10 at a time
Batching leaves the odds across all 60 where they were and changes the search. When a batch of 10 kills HDMI, the bad app is one of 10, and halving finds it in 4 reboots. Blind at 60, you're halving the whole list and maybe chasing two culprits.
Tell Claude three things before it starts:
→ disable, never uninstall: a mistake costs a reboot, not a factory reset
→ 10 apps per batch, then test the remote, HDMI and sound
→ install a new home screen before the old one goes. For the launcher p = 0, and that's the black screen
Same math runs a 9-bot Grok chain: at 95% per bot, 0.95⁹ = 63% of runs come out clean.
What's the first app you'd turn off?
REA is most dangerous thing on the internet right now...
It lets your AI agent reverse engineer any app, down to the binary
If you use it correctly, you can rebuild any feature you pay for, for FREE
In this article, I show you how: https://t.co/3P1FNXiXzl
@justansub Building the home screen from scratch is the smart move: the launcher is the one app where a mistake costs you the whole TV, and now it's yours. How much RAM did the stock launcher use compared to yours?
@0900mihir 192 MB down to 83 MB on the launcher alone explains a lot of the speed. Did you measure free RAM before and after? One TCL owner went from about 64 MB free to about 900 MB
@shesaidmewakeup @0900mihir Not always. Updates can switch disabled apps back on, and some builds bring the stock launcher back. Keep the list of everything you disabled and rerun it after each update. It's a one-line loop over the file
"Prompt engineering" was always a temporary bridge.
Just like compilers eliminated handwritten assembly, recursive self-prompting systems like Harness² will eliminate manual system-prompt tweaking. The transition from static text files to dynamic, git-versioned, self-healing agent configs is the cleanest realization of Software 3.0 so far
@Plutkar "Disable first, delete second, protect the I/O layer."
This is a masterclass in prompt-level guardrails for agentic execution. Saving a sluggish 2GB stick from the e-waste pile with a 12 MB custom launcher is wildly satisfying
This shifts the bottleneck from model training to meta-prompting and state management. We spent years treating system prompts and tool definitions as static configuration.
Treating them as a mutable, self-optimizing runtime environment is essentially bringing gradient-free meta-learning to the system layer.
The big challenge now: preventing prompt drift and regression over long task horizons
Spot on. The "code as a moat" era is officially dead. When the cost of reverse engineering and specification drops to near zero, the real moat aggressively shifts to distribution, trust, proprietary data loops, and deeply embedded operational workflows.
You can clone 80% of Salesforce or Jira over a weekend with agentic pipelines, but the enterprise won't switch because the barrier was never just the code — it’s compliance, integrations, and organizational inertia. But for unbundled, overpriced B2B micro-SaaS? They are sitting ducks
The best line in @0xjmori's guide is one sentence:
"A successful render only proves that the file exists."
It's also why one-bot studios break more often than they look ↓
The guide builds a motion studio inside one Opus 5.5 agent. Five departments work in order:
Director → Art Director → Animator → Render Engineer → Critic
Each department is a stage that has to land. Say every stage gets its part right 90% of the time:
1 prompt, 1 stage: 90%
1 agent, 5 stages: 59%
5 stages + an independent Critic: 91%
The formula
P(good video) = p⁵
p = 0.99 → 95%
p = 0.95 → 77%
p = 0.90 → 59%
p = 0.80 → 33%
Reliability engineers have multiplied stage odds like this since the 1950s (Lusser's law).
The guide's own example
It's a 20-second launch film: a messy task board turns into a sprint plan. 600 frames at 30 fps, the click at frame 195, the result at frame 210.
Now picture the render finishing 600/600 with the cursor sitting on the headline at frame 210. The agent reports done, and the file exists. The client still sends it back.
The Critic, in numbers
It reviews the rendered frames, ideally in a fresh context, and only the broken frames get fixed and rendered again. Cap that loop at 3 rounds. If the Critic flags 80% of real defects and a local fix works 90% of the time:
→ good videos: 59% → 91%
→ broken videos that slip through: 9%
At a 50% catch rate, 22% slip through.
Before you trust a one-bot studio, measure its Critic. Plant 20 known defects in a render, a cursor over a headline among them, and count how many it flags.
Save this for the day you build yours.
What catch rate would you need before it ships to a client?
Claude Opus 5.5 could turn a single AI bot into an entire motion design studio.
It can direct scenes, animate product interfaces, render videos, catch visual mistakes and deliver polished results.
In this article, I show you how. https://t.co/HjZIAksasB
@Stochastly Exactly. And you can put a number on it: with ~0.58 Sharpe noise on 3y of daily data, the best of 30 zero-edge tries lands around 1.18 on average. So 1.47 is only ~0.5σ above luck. Pure noise hits 1.47+ in about 15% of 30-night runs
Steps 3 and 4 of this prompt quietly cancel each other out.
Step 3: trade only if the backtest beats the market.
Step 4: every night, tweak the strategy and test again.
Give it a strategy with zero edge. Within 30 nights, it approves it 72% of the time ↓
━━━━━━━━━━
The regime switch and the kill switch are solid. The bug is how steps 3 and 4 interact.
A backtest Sharpe ratio is a noisy estimate. On 3 years of daily data, a strategy with a true Sharpe of 0 still scores around ±0.58 (one standard deviation).
Say the bar is Sharpe > 1. A pure-noise tweak clears it about 4% of the time.
One try: 4%. But the night shift never tries just once.
━━━━━━━━━━
The formula
P(approved) = 1 − (1 − p)^N
p = chance one useless tweak passes (4%)
N = nights of tweaking
N = 1 → 4%
N = 7 → 26%
N = 30 → 72%
N = 90 → 98%
Nothing got smarter. It rerolled the dice every night until the dice agreed.
And the winner won't look marginal. After 90 tries, the luckiest noise strategy shows a backtest Sharpe of about 1.4. Good enough to put real money behind.
━━━━━━━━━━
The fix: 3 lines to add to the prompt
1. "Hold out the last 12 months. The night shift never sees them. Test on them once."
2. "Log every tweak you try, including the rejected ones."
3. "Raise the bar with every try." That's the Deflated Sharpe Ratio (Bailey & López de Prado, 2014): compare the result to the best Sharpe luck alone would produce after N tries, not to zero.
━━━━━━━━━━
Automation doesn't remove overfitting. It industrializes it.
I break down the math inside viral money posts, with the code. Follow for the Python version.
(Math, not financial advice. Simulated zero-edge strategies.)
SOMEONE LEAKED THE PROMPT TO CREATE TRADING AGENTS. YOU PASTE IT INTO CLAUDE CODE AND IT SETS UP THE BOT WHILE YOU GRAB A COFFEE.
No programming required. Just paste. A single prompt does it all:
Time required: 1 minute
Actions: copy and paste into Claude Code
The mistake that wipes out the accounts of 99% of traders with bots: they use the same strategy for everything.
But the market isn't the same on a quiet Tuesday as it is in the middle of a crash
1. Read the market like a quant: detect if it's calm, sideways, jittery or in freefall and switch strategies for each one
2. Handbrake: maximum drawdown of 15% and a kill switch that's checked before EVERY order
3. Test before touching the money: backtest the strategy period by period with real data and refuse to trade if it doesn't beat the market
4. Night shift: while you sleep, it analyzes what went wrong and leaves ready improvements. You decide if they're applied
5. Zero code: you paste, fill in your data and Claude handles the rest
Before, this meant a team of PhDs, mathematicians, thousands of lines of code and weeks of work.
Now it's just Ctrl+V.
Save it before they delete it.
@akshay_pachaar The lock-in math is brutal: k agent stacks wired point-to-point = k(k−1)/2 integrations. 4 stacks: 6. 8 stacks: 28. A shared layer like this makes it k. That's the real case for keeping coordination separate from any one provider
@akshay_pachaar Cool idea. The number I'd love to see next: rerun spread. At ~50% pass on 100 tasks, two runs of the same harness differ by ~7 pts (1σ). So +11.8 likely holds with 150+ tasks, while +4.5 needs ~1,000 unpaired. Any seeds or ± in the paper?
Harness² gained +11.8 points in one round. One number is missing.
The claim: one round of self-edits beat the original harness. Up to +11.5 on WorkBuddy-Bench, +4.5 on JobBench, +11.8 pp on LAB's partial pass rate.
The missing number: how much the score moves when you simply run it again ↓
Agent runs are noisy. Same harness, same tasks, a different score every time. A gain is real only when it's bigger than that wobble.
For a pass rate p on n tasks, two runs differ by noise of about
σ = √(2p(1−p)/n)
To clear 2σ, you need
n ≥ 8p(1−p) / Δ²
At p = 0.5 (the worst case):
→ +11.8 pp needs ≥ 144 tasks
→ +4.5 needs ≥ 988 tasks (if it's a pass rate)
On 100 tasks, rerunning the exact same harness "gains" +4.5 or more in about 1 run out of 4. Zero real change.
The README lists no task counts, seeds, base scores or error bars. From the page alone, you can't tell a better harness from a lucky run.
(Paired runs on the same tasks need fewer. You still need the variance to know how many.)
One more line from the repo is worth reading twice:
"Public benchmark rubrics may be accessible online."
And the executor runs with unrestricted network access. An agent that edits its own setup and can browse is a good reason to switch the network off before trusting a score.
None of this says Harness² doesn't work. A harness that improves itself is a real idea, and the code is open, so anyone can run the check:
→ 3+ seeds per arm
→ mean ± spread, not one number
→ network off
Would you trust a +4.5 without error bars?
I break down the math behind AI and money claims, with the code. Follow for more.
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: https://t.co/i00WLeeFpw
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
The ledger in Python:
deals = [(186_420, 0.055), (97_236, 0.045), (63_880, 0.060)] # illustrative
buyer = sum(round(l * s) for l, s in deals) # +18,462 (5.31%)
vendor = -buyer # -18,462
R = round(37.5 * 68) # haggling hours nobody does × $/h
print(buyer, vendor, buyer + vendor + R) # 18462 -18462 2550
Source: https://t.co/dgRW7H68dz
We are putting a lot of effort into a specialized Hagglebot for business use, as we think it could improve enterprise cost-efficiency by 5% or more.
Applied to the global GDP, this could be up to $6 trillion/year in savings, delivering an immediate and measurable advantage to any company using Grok @Bot.
I had Opus 5.5 audit Elon's $6T Haggle Bot math. One line is missing.
Elon: Hagglebot could make companies 5%+ more cost-efficient. "Applied to the global GDP, this could be up to $6 trillion/year in savings."
The math behind it: 5% × ~$120T world GDP = $6T ↓
For one company, it's real.
Saving = s × S
s = the discount Haggle Bot wins (5%)
S = the spend it can negotiate
Three contracts worth $347,536, haggled down 4.5–6% each: $18,462 back (5.31%). Immediate and measurable. Elon is right about that part.
For the world, every haggle has two sides.
Your $18,462 discount is your vendors' $18,462 of lost revenue. Same dollars, opposite signs. Money moved; the needle didn't.
World gain = Σ saved − Σ lost + R
The first two terms cancel. What's left is R:
→ hours nobody spends negotiating anymore
→ things nobody buys once the bot shows they weren't needed
→ vendors forced to get cheaper to run, not just to earn less
So the honest version:
$6T is how much money Haggle Bot could move between companies. The gain to the world is R, and nobody has measured R yet.
Still a huge market. Just a different number.
Same trap as "AI saves $X in salaries": count both sides of the ledger before you add it to GDP.
I break down the math behind viral money posts, with the code. Follow for more.
(Math, not financial advice. Deals in the video are illustrative.)
We are putting a lot of effort into a specialized Hagglebot for business use, as we think it could improve enterprise cost-efficiency by 5% or more.
Applied to the global GDP, this could be up to $6 trillion/year in savings, delivering an immediate and measurable advantage to any company using Grok @Bot.
For the company using it, 5% is real money. For the world it's mostly a transfer: every haggled dollar is a vendor's lost revenue. What's left is the waste that actually disappears (hours, spend nobody needed). Still big. Just not 5% of GDP
Last Seen Users on Sotwe
hakan keskin
Seen from Germany
belanja stw
Seen from United States
Lady Unicornio
Seen from United States
فحل سوداني
Nicola Porcella
Seen from Turkey
Cockold หาชายเดี่ยวช่วยเมีย
Seen from Thailand
LOLI (Part-Time Sahibe)
Seen from Turkey
Türk İfşa - Nisa
Seen from Netherlands
نوف
binor
Seen from Indonesia
Most Popular Users
Elon Musk 
@elonmusk
241.5M followers
Barack Obama
@barackobama
118.8M followers
Cristiano Ronaldo
@cristiano
114.6M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Narendra Modi
@narendramodi
107.1M followers
Rihanna
@rihanna
98.7M followers
NASA
@nasa
92.3M followers
Justin Bieber
@justinbieber
91.8M followers
KATY PERRY
@katyperry
90M followers
Taylor Swift
@taylorswift13
84M followers
Lady Gaga
@ladygaga
75.5M followers
Virat Kohli
@imvkohli
73.5M followers
Kim Kardashian
@kimkardashian
70.9M followers
YouTube
@youtube
68.8M followers
Neymar Jr
@neymarjr
66.5M followers
Bill Gates
@billgates
65.3M followers
Selena Gomez
@selenagomez
63.1M followers
The Ellen Show
@theellenshow
62.2M followers
CNN
@cnn
61.7M followers
X
@x
60.7M followers