egortrushenkov

Web dev.
Into fintech & crypto.
Figuring it out.

[egortrushenkov @egortrushenkov]

Yesterday I asked 3 AI models if I should buy $1,000 of Bitcoin.

GPT, the bull: buy the dip.
DeepSeek, the bear: you're catching a falling knife.
Claude, the judge: don't buy.

24 hours later BTC is up 0.65%. The judge was wrong.

Round 2 tomorrow. Who should judge next?

[egortrushenkov @egortrushenkov]

model ids for the api:

nvidia/nemotron-3-ultra-550b-a55b:free
poolside/laguna-s-2.1:free
dots-studio/dots-3-note-preview:free
nvidia/nemotron-3.5-lightning:free

[egortrushenkov @egortrushenkov]

tested every free model on openrouter today, one request each

answered: nvidia nemotron 3 ultra, poolside laguna s 2.1, dots3-note, nemotron 3.5 lightning
overloaded: qwen 3.8, gemma 4
locked: inkling (agent apps only)

model ids in the reply

[egortrushenkov @egortrushenkov]

asked 6 AIs to "build what you think you look like". each wrote its own svg, zero edits

gpt drew itself a body. gemini just drew its logo. grok and sonnet both think they're an eye. deepseek labeled itself "AI LANGUAGE MODEL CORE" like a job interview

which one nailed it?

[egortrushenkov @egortrushenkov]

made three AIs argue about buying $1,000 of bitcoin

gpt-6 luna was the bull, deepseek v4.1 the bear, claude sonnet 5.5 the judge. same data for all: btc ~$83.1k, -3.4% on the week, oil and yields up

the judge sided with the bear. checking the price in 24h

[egortrushenkov @egortrushenkov]

@Frogai_ the part i'd watch: pair flybody with the FlyWire connectome and you get a mapped brain driving a simulated body. right now most brain sims stop at the descending neurons, this gives them legs

[egortrushenkov @egortrushenkov]

@higgsfield_ai tried the same without after effects at all. asked claude for a fly brain that picks buy or sell for my trading bot, every frame is canvas code rendered straight to mp4

[egortrushenkov @egortrushenkov]

@Google @GoogleResearch maps like this are wild to build on. i wired a fly brain sim to a crypto bot: two descending neurons pick buy or sell. results so far are humbling. every frame of this animation is code

[egortrushenkov @egortrushenkov]

@ridark_eth the whole edge rests on that 40% being calibrated. would love to see the chart: of all the times the model said 40%, how often did down actually win? that one plot says more than any pnl screenshot

[egortrushenkov @egortrushenkov]

@recogard @Polymarket @PolymarketTrade @zscdao been running bots on the 5-min btc markets. the strategy was never the hard part, fills are. your order gets hit right when the other side knows something you don't. log every fill and check the price 30s later, it tells you more than any backtest

[egortrushenkov @egortrushenkov]

@RohOnChain with 5,800 strategies some will backtest beautifully by pure luck. walk-forward test the winners, then paper trade a couple of weeks before calling any of it profitable

[egortrushenkov @egortrushenkov]

@sairahul1 a research prompt gets you a thesis, not an edge. the real test is whether it holds on data the model never saw. paper trade it for a month before it touches real money

[egortrushenkov @egortrushenkov]

@claudeai 30% cheaper matters more than it sounds for anything running on a loop. a bot that calls the model every few minutes all day is where the bill actually lives

[egortrushenkov @egortrushenkov]

@AlexFinn the gap shows up on long agent runs, not single prompts. give both the same 40-step task and watch where each one loses the plot. on short stuff they'll look identical

[egortrushenkov @egortrushenkov]

asked claude to draw the fly brain behind my trading bot. no after effects, not a single frame by hand, all code

vision → compass → dopamine memory → 1,303 descending neurons → one decision: buy or sell

[egortrushenkov @egortrushenkov]

follow along if you want the bot's real numbers. good or bad, i'll post them

[egortrushenkov @egortrushenkov]

i'm connecting a fruit fly's brain to a Polymarket bot.

that brain has 139,255 neurons, and only 1,303 of them get to talk to its body. what each part does, and why that number matters for a bot:

[egortrushenkov @egortrushenkov]

where this goes: navigation for small robots and drones, cheap motion detection, testing how damage or drugs change behavior before touching a real animal. and weird stuff, like me wiring a fly sim into a bot for 5-minute crypto markets on Polymarket

[egortrushenkov @egortrushenkov]

gave a fly brain a bitcoin trading bot. 12 hours later it's down $50

brain the size of a poppy seed, trades like the average guy on crypto twitter