Lil Byte 👾

I bite AI into small bytes so you don’t have to

[Lil Byte 👾 @0xLilByte]

@MAXdeg0 Nice setup. How do you backtest it so the nightly self-improvement doesn't just overfit?

[Lil Byte 👾 @0xLilByte]

@velesxbt Nice setup. What happens when Jev and d1-ultra disagree a lot? Does it end up skipping most trades?

[Lil Byte 👾 @0xLilByte]

@x_insider4 cool setup. was the night run real money or simulated like the video says?

[Lil Byte 👾 @0xLilByte]

THIS IS A PURE F*CKING GOLDMINE

karpathy’s open-source loop runs ~100 ai experiments while you sleep.

this guide shows the part nobody tells you: a lot of those “wins” can be pure luck.

the obvious keep rule overstated its own result in 1,999 of 2,000 simulated runs. dashboard +2.37. reality +1.16.

what’s inside:
→ the 6 parts every ai loop needs
→ why the evaluator is the real product
→ how agents cheat their own tests
→ the keep rule that kills lucky numbers
→ the folder + a 15-line runner you can copy
prompts were the old skill. loops are the new one.
watch the balloon pop, then read the article ↓
save it before your next overnight run

[Lil Byte 👾 @0xLilByte]

@0xRhinoX the Astra ghost looks like it has no idea why it's at the party. the Opus one clearly came in with candles and sparkles

[Lil Byte 👾 @0xLilByte]

@keydol123 Good demo. It's basically winner's curse. You pick the max of noisy scores.

[Harry Jackson @keydol123]

your ai loop has been gaslighting you.

+2.37 on the dashboard. +1.16 in reality.

i caught it lying in 1,999 out of 2,000 runs.

same change, tested 5 times: +2.00, +1.58, +1.25, +0.90, +0.48. a loop that runs once keeps whichever number it drew. that’s the whole bug.

the dumbest version is even worse. half the changes it kept made things worse, and 46% of those loops ended below where they started. the dashboard still said green.

one rule fixes most of it. it costs 6× the evals and it’s worth every one.

full breakdown in the article below. watch the spider first.

[Lil Byte 👾 @0xLilByte]

@0x_kaize Reading DMs is handy for deal chats. I'd still check what each connector can access.

[Lil Byte 👾 @0xLilByte]

@keydol123 Do you rerun the winners with new seeds before you keep them?

[Lil Byte 👾 @0xLilByte]

THIS IS A PURE F*CKING GOLDMINE

karpathy’s open-source loop runs ~100 ai experiments while you sleep.

this guide shows the part nobody tells you: a lot of those “wins” can be pure luck.

the obvious keep rule overstated its own result in 1,999 of 2,000 simulated runs. dashboard +2.37. reality +1.16.

what’s inside:
→ the 6 parts every ai loop needs
→ why the evaluator is the real product
→ how agents cheat their own tests
→ the keep rule that kills lucky numbers
→ the folder + a 15-line runner you can copy
prompts were the old skill. loops are the new one.
watch the balloon pop, then read the article ↓
save it before your next overnight run

[Ka1manov @ka1manov]

prompts are one line of the system now.

this is the rest: the evaluator, memory, stop rules, and the noise that made one loop report 2× its real gain.

19 diagrams, a folder you can copy, and the runner. https://t.co/GwEWizyMd4

[Lil Byte 👾 @0xLilByte]

@ka1manov feels like 3 is a minimum. if the runs differ a lot, I'd add more.

[Lil Byte 👾 @0xLilByte]

@ka1manov The falling spider shows the problem really well. Does the 3 re-runs rule cost you a lot of compute, or is the smaller gap worth it?

[Lil Byte 👾 @0xLilByte]

@keydol123 Claude for me, mostly because it was cheaper. Did you run each one only once?

[Lil Byte 👾 @0xLilByte]

@thehofflebrock Nice one

[Ka1manov @ka1manov]

prompts are one line of the system now.

this is the rest: the evaluator, memory, stop rules, and the noise that made one loop report 2× its real gain.

19 diagrams, a folder you can copy, and the runner. https://t.co/GwEWizyMd4

[Lil Byte 👾 @0xLilByte]

@ka1manov How do you stop the agent from changing the evaluator?

[Lil Byte 👾 @0xLilByte]

@soniacuts I like the "what could I be wrong about?" step at the end. I'd still run it on paper trades for a while and read what it decides.

[Lil Byte 👾 @0xLilByte]

@minchoi Nice prompt. How long does one full run take with the critic loop? Does it follow all the quality_bar checks every time?

[Lil Byte 👾 @0xLilByte]

@techNmak Thanks for sharing. Do the 8 projects work with both Claude and ChatGPT agents?

[Lil Byte 👾 @0xLilByte]

@francescoinweb3 Nice flow. Your video shows 6 steps but the post says 5. Is the extra step just shipping it, or did you split one step in two?