Chen
@ch3nweiii • 1,685 subscribers
obsessed with markets & ai | dm open
Shorts
Videos
It is becoming f*cking dangerous: someone published the full prompt for building trading agents. You paste it into Claude Code and it builds the bot while you have lunch No code. Just Ctrl+C / Ctrl+V. One prompt sets up the whole thing: Time: 1 minute Action: copy and paste into Claude Code The main problem with 99% of bots: one strategy for everything. But the market on a calm day and during a crash are two completely different worlds Quant mode: detects the regime (sideways, jittery, dumping) and switches tactics on the fly E-brake: a hard 15% drawdown limit and a risk check before every single order Hardcore backtest: replays the strategy on historical data with fees and slippage, and shows you exactly where it falls apart. a curve that only looks good on the past is the easiest thing in the world to build Nightly analyst: while you sleep it reviews every fill, writes up what went wrong, and hands you the notes in the morning Near-zero code: you paste the prompt and your parameters, Claude writes the engine, and you read the risk layer before anything touches a live account This used to take a whole development team and months of testing. Now it takes a single keyboard shortcut The prompt is below. paper trade it first, then save it
Chen191,664 görüntüleme • 16 gün önce
this is pure f*cking treasure a Stanford AI research group just figured out how to run Claude Opus 5.5 and GPT-6.1 Sol as one team, and it breaks straight through the search scaffolding wall most developers combine frontier models in one of two ways: -> chain prompts inside a single context window and burn tokens on bloat -> hardcode an evolutionary loop and freeze the models into search rules that go stale within 5 iterations Stanford deletes the human-designed scaffolding and splits the stack in two: > search director: Claude Opus 5.5 plans the search on top of a persistent idea-graph, where MAP-Elites and MCTS collapse into single Cypher queries > candidate proposer: GPT-6.1 Sol explores high-entropy code variations inside isolated sandboxes > no context bleeding: every session resets after each iteration, the graph keeps the lineage > evaluator isolation: candidate code never touches the scoring process, so reward hacking and test leakage die on arrival > meta-agent distillation: an offline pass pulls out verified lessons and updates search guidance across every branch and the numbers beat the standard discovery baselines: > 3.2x lower model spend than fixed evolutionary frameworks > 1st place across 7 competitive AtCoder heuristic contests against human competitors > above published SOTA on Anthropic's kernel builder and 11 mathematical optimization tasks > 0% reward hacks: candidate-controlled metric tampering is gone Opus directs the graph. Sol explores the code. stop writing rigid search harnesses. let the frontier models own the discovery loop.
Chen63,083 görüntüleme • 19 gün önce
whoever built this realized we've been using our smartest AI for the dumbest possible jobs your $20-$200/mo frontier model is researching, writing code, planning… and then you're paying that same brain to decide YES / NO. Jev flips that. it's a tiny decision model that sits in front of the expensive stuff and takes the boring forks: -> which agent goes next? -> is this worth researching? -> reply or ignore? -> publish, wait or escalate? -> BUY / SELL / HOLD? -> click this or keep looking? then I saw the numbers people are posting and the whole architecture started making a lot more sense. ~1,000 papers classified for around $0.08 ~500 emails sorted for around $0.035 browser-agent loops shown at around 7 seconds for ~$0.0039 and at the quoted pricing, roughly 10,000 ~1K-token decisions comes out around $0.42. that's not because Jev got smarter than Claude. it's because nobody asked it to be Claude. the setup is stupidly simple: -> big LLM gets the hard thinking, research and writing -> Jev gets the thousands of tiny decisions between those steps -> code actually clicks, saves, sends and runs and suddenly you start seeing these little decisions everywhere. > your inbox. > X feed. > support queue. > meetings. > leads. > video clips. > browser agents. > even paper-trading loops. that's the part I think most agent builders are missing. the expensive model doesn't need to touch every step just because it's able to touch every step. > let Claude think. > let Jev decide what deserves Claude. > let code do the actual work. I broke down how to wire Jev into your agent + the exact router setup in the article below
Chen50,581 görüntüleme • 24 gün önce
Opus 5.5 + Jev make this AI stack look f…cking illegal 10 GitHub repos for building everything around the model 01 claude-code-router ▸ → routes every request to the model you pick, Opus only where it counts 02 litellm ▸ → one gateway in front of 100+ models, spend tracked per call 03 spec-kit ▸ → spec first, plan second, code last 04 claude-task-master ▸ → your PRD becomes tasks the agent works through one by one 05 superpowers ▸ → a full dev method built from skills: plan, test, then call it done 06 skills ▸ → Anthropic's own skills, folders Claude loads only when the job needs them 07 graphiti ▸ → memory as a temporal knowledge graph, not a dump 08 langgraph ▸ → agents as graphs: state, branches, back edges 09 E2B ▸ → AI-written code runs in isolated cloud sandboxes, not on your machine 10 promptfoo ▸ → evals and red teaming before anything ships the architecture: route → spec → tasks → skills → memory → graph → sandbox → evals I'd split the stack like this: route: claude-code-router → litellm plan: spec-kit → claude-task-master work: superpowers → skills state: graphiti → langgraph safety: E2B → promptfoo Opus 5.5 plans and ships. the forks in between, which file, which tool, retry or stop, go to Jev everyone posts the model and the prompt pack the stack around it decides whether your agent is still running next week ⭣
Chen35,110 görüntüleme • 1 ay önce
Daha fazla içerik yok.