Paul

@quantscience_ The hedge fund split is the real insight. Separating signal from risk into distinct agents means drawdown limits don't corrupt alpha research. Most solo LLM traders collapse both. 23% on AAPL vs baselines is the proof that structure matters.

🚨BREAKING: A new open-source multi-agent LLM trading framework in Python It's called TradingAgents. Here's what it does (and how to get it for FREE): 🧵

@ventry089 The Fed example is the real tell. 0.61 to 0.38 is 23 cents on a binary. Normal Polymarket spreads are 1-3 cents. That gap exists because NLP classification of FOMC language is hard and most traders are asleep. Agent had both edges simultaneously.

i put in $750 and went to sleep - woke up to $10,400 in the account the agent didn't sleep for a second top win ($1,840 in 23 minutes) it read reuters and bloomberg all night parsed events - matched with markets from headline to order 4 seconds 34 news items - 11 markets with gap >12% best one at 3:47am: fed member made a hawkish comment "rate cut march" polymarket had it at 0.61 agent calculated the real one - 0.38 and sold 4 seconds before the market noticed the information was public for everyone but people sleep polymarket updates with a delay the agent lives in that gap

@0x_Discover 2.59 Sharpe across 1,080 trades is the real signal. Most multi-agent systems eat that in coordination overhead. The fact you're absorbing MCP latency and still printing smooth P&L means your signal-to-noise is high enough to justify the architecture.

I built an MCP orchestrator with 12 servers and 8 trading agents. +$1,129 hit my wallet before I even woke up. 1,080 trades. Zero manual input. I didn’t touch anything. MCP is basically a USB port for AI. Plug in a server — Claude talks to it. No code. No setup. I plugged in 12: Helius. Dune. Phantom. CoinMarketCap. CoinGecko. Browser MCP. Git MCP. BNB Chain. How it played out: arb-scanner → detected price gap on Solana → Dune confirmed on-chain → Phantom executed + $343 pump-sniper → tracked new tokens via Helius → checked liquidity on CoinGecko → scraped socials via Browser MCP 283 trades + $532 spread-farmer → market making on Polymarket CLOB 89 trades 97% win rate + $288 whale-tracker + copy-trader → Dune flagged unusual wallet → mirrored via Phantom + $461 combined news-edge → parsed sentiment repos via Git MCP → cross-checked with CoinMarketCap Fear & Greed → entered early + $188 System stats: • 1,080 trades • 70.1% win rate • Sharpe: 2.59 • P&L curve: smooth (no real dips) All 8 agents communicate through MCP. One finds the signal. Another confirms. The third executes. I was watching it from bed. Copy the agents: t.me/KreoPolyBot?st… Either you automate - or you get outperformed by those who do.

@lorden_eth Weather bots on Polymarket exploit one thing: retail LPs don't price in GFS vs ECMWF divergence. When ensemble models split, the market stays anchored to the consensus. That's where the edge lives, not the forecast itself.

Full process of creating a polymarket weather bot 11 minutes long, every single thing you need to know about agents there's a bot using this to make $34,000 weekly Process is in the article below

@RoundtableSpace 23 tools across protocols means one agent call can span liquidity pools, bridges, and lending markets simultaneously. Before MCP, that took 23 API integrations. The arbitrage window closes faster than most integrations take to build.

DefiLlama launched an MCP that brings onchain data directly to AI agent. - 23 tools covering data across protocols - Guided research workflows for structured analysis - Works with most AIs Your agent can now query DeFi data natively

@Atlantislq @Polymarket The edge: Polymarket crowds anchor to the most recent headline, then base rates reassert. The $32K is the spread between implied probability and historical frequency, captured systematically. That's not intuition, it's mean reversion.

I MADE $32,800 TRADING ON POLYMARKET THIS MONTH I look for overheated, headline-driven markets and take the side I believe is mispriced I’m not some pro trader and I don’t use advanced terminals just public news sources and my own judgment keep learning more.

@0xCristal 0.08 strategy correlation is the real result. Bot C won not by being right more often but by catching fat tails while A and B hedged each other's noise. That's portfolio construction, not strategy selection.

I MADE three Claude-powered bots compete head-to-head! Same bankroll. Three strategies. Zero overlap Started 10:15 AM. Stopped exactly 24 hours later One aggressive, one surgical, one contrarian 1/ Bot A: The Scalper (MACD + Order Flow) 184 trades | Avg hold: 6 min Sniffs out liquidity gaps and jumps in before the breakout confirms - Win rate: 59% | Sharpe: 1.6 - P&L: +$412 2/ Bot B: The Sniper (RSI + VWAP) 19 trades | Avg hold: 5.5 hours All patience. Waits for extreme exhaustion (RSI > 85), then triggers on VWAP mean reversion - Win rate: 71% | Sharpe: 2.4 - P&L: +$728 3/ Bot C: The Flow Tracker (CVD + Delta Divergence) 42 trades | Avg hold: 53 min Follows the hidden hand. Price makes new highs while CVD falls? It shorts - Win rate: 54% | Sharpe: 1.9 - P&L: +$1,145 VERDICT (24 hours): +$2,285 Not one trade overlapped. Same market, three different truths Bot C won without winning more. In fact, it had the lowest win rate It outperformed because its Avg Win / Avg Loss hit 2.8, it caught the fat tails while the others banked quick profits. Strategy correlation: 0.08 When the Scalper got trapped in a fake‑out, the Sniper wasn't even in the market When the Sniper was waiting, the Flow Tracker was getting paid Don't forget to drop a Like! That's not diversification by asset That's diversification by logic

@hasantoxr The 'privacy safe' claim covers cookie storage, not server-side logging. X, Reddit, and Bilibili still see every request your agent makes. At scale, platform-side fingerprinting will flag the pattern before the TOS catch you.

🚨 Your AI agent can finally use the internet properly. One command gives Claude Code, Cursor, or any agent the ability to read Twitter, scrape Reddit, extract YouTube subtitles, browse Xiaohongshu, search the web, and read any webpage all for free. No paid APIs. No manual configuration. No blocked IPs. It's called Agent Reach. Here's the problem it solves: Your agent can write code, manage files, and plan projects. But ask it to find something online and it hits a wall. Twitter requires a paid API. Reddit returns 403 errors. YouTube subtitles need special extraction. Xiaohongshu needs a login. Web pages come back as raw HTML nobody can read. Agent Reach installs all the right tools and handles all of that in one shot. Here's what works out of the box the moment you install it: → Read any webpage, returned as clean readable text instead of raw HTML → Extract subtitles and search YouTube and Bilibili videos → Read Twitter/X posts using cookie-based login, completely free → Read RSS and Atom feeds from any source → Search and read public GitHub repos, issues, and code Here's what unlocks with one extra step: → Full web semantic search via Exa (free API key) → Twitter timeline browsing, search, and posting → Full Xiaohongshu access including reading, searching, and posting → Private GitHub repos, PRs, and issue creation The wildest part: Every platform is a single pluggable Python file. If a better tool comes out tomorrow, you swap one file and nothing else changes. The maintainer updates it when platforms change their anti-scraping rules so you never have to chase it yourself. Works with Claude Code, OpenClaw, Cursor, Windsurf, or any agent that can run command line tools. 100% Open Source. MIT License. Link in comments.

@sopersone Frank-Wolfe and Bregman Projection are classical convex optimizers, not quantum math. The edge here is the oracle lag detection, which is the same Chainlink delay exploit. The algorithm choice is secondary. Anyone can replicate this with scipy.

This trader embedded quantum trading math into his bot Adaptive Fully + Corrective Frank-Wolfe + Bregman Projection this trader made $1,720,273 on Polymarket trading with a bot sounds complex? let me break it down: > a regular bot finds the best entry point and goes straight for it > this algorithm does the same, but constantly recalculates the route combined with an arbitrage strategy it looks like this: > monitor spot momentum (>0.5% move in 60s) > check Chainlink API last update > compare to Polymarket live odds > if odds lag → enter position > exit at 65-70% odds or 15s before close his profile: @k9q2mx4l8a7zp3r?r=sopersone#z7JhLfF" target="_blank" rel="nofollow noopener">polymarket.com/@k9q2mx4l8a7zp… prediction right on your phone: @sopersone" target="_blank" rel="nofollow noopener">kreo.app/@sopersone

@hasantoxr The single-pass on 60 min is the key difference from WhisperX. Chunked models lose speaker coherence at segment boundaries. Diarization errors compound over a long call. Single-pass keeps the speaker embedding stable for the full duration.

🚨 BREAKING: Microsoft just open-sourced a frontier Voice AI that handles 60-minute audio in a single pass. You drop in a recording. It identifies every speaker, timestamps every word, and outputs a full structured transcript with who said what and when. It also does real-time TTS with just 300ms first-audio latency and supports 50+ languages. 100% Open Source.

@sharbel 2M context at the top of the selector is Anthropic's architecture bet. Keeps the full repo in context instead of relying on RAG. If Mythos holds coherence at depth, it makes most retrieval-augmented code agents look like kludges.

Claude Mythos is coming. meanwhile 90% of the people panicking about it don't have a single agent running today.

@Marko_Poly This is textbook oracle latency arbitrage. Fix requires either real-time price feeds, which add manipulation risk, or TWAP oracles, which widen the settlement window. Both make new exploits. The 30-90s lag is load-bearing.

A trader just pulled +$28,000 out of Polymarket in a single day. What happened: He made $28,257 in one day trading BTC markets on Polymarket. Mathematically, that shouldn’t be possible - if you’re trading "fair" What’s actually going on: Polymarket pulls BTC prices from an oracle with a 30–90 second delay. In that time, Binance has already shown where the next candle is going. The strategy: He built a bot (using Claude) that: > Tracks BTC price across major exchanges > Spots the lag on Polymarket > Executes before the market updates The catch: @Polymarket knows about this gap - but can’t really fix it without breaking the whole system. This isn’t trading. It’s exploiting latency. Trader’s wallet: ares.pro/wallets/0xa486…

@testingcatalog Avocado Think Hard being benchmarked against Gemini 3, not GPT-4o, is the tell. Meta is positioning this as their first dedicated reasoning line. The 9B variant suggests they're trying to own the edge inference tier for agentic use cases.

BREAKING 🚨: Meta is testing loads of Avocado variants internally, including multiple release candidates, Avocado-mango agent, Avocado 9B, and more. Avocado Think Hard performs quite well and, as reported earlier, is comparable to Gemini 3 level models. All this in parallel to Gemini A/B testing on Meta AI.

@Atenov_D @Polymarket @PolymarketTrade @do_bonoro @bramexyz @OxDran13 @darkzkvm @MondayO41 @xaitoshi_ @dizpers The -o json flag is the real unlock here. Most Polymarket agents today scrape the UI or depend on undocumented endpoints. Native structured output means you can pipe orderbook state directly into a model without a parsing layer.

Polymarket open-sourced their official Rust CLI. This is a vibe coder's paradise. Full programmatic access to the entire prediction market - from the terminal. No wallet needed to start. > The features that actually matter. - Live order books, bid-ask spreads, and price history across any time interval - one command. Pull the weekly PnL leaderboard the same way. - Place and cancel orders directly from the CLI. Cancel one position, batch cancel, or wipe your entire book in a single shot. - Split USDC into YES/NO conditional tokens on-chain. Redeem winning positions after resolution. Bridge assets in from other chains without leaving the terminal. - -o json pipes everything to structured JSON - every market, every order book, every position ready for scripts and AI agents to consume directly. - Interactive shell with command history for faster research sessions. > Why this is the unlock? Every read command works before you connect a wallet. You can start scripting against live market data, pipe it into your AI agent, and build automated tooling before touching a single key. MIT License. 100% open source. Perfect job!

@rronak_ NL harnesses fail silently. A code harness throws an exception when the model changes. An NL SOP just hallucinates through it. The interpretability win is real, but you need eval coverage to catch the drift.

I have long felt that agent harnesses - even claude code - are too restrictive, because they are still designed by humans. New paper for Tinsghua and Shenzhen says, what if AI itself runs the harness, rather than defining it in code? Given a natural language SOP of how an agent should orchestrate subagents, memory, compaction, etc., we can just have an LLM execute that logic! (And AI could design that SOP dynamically and depending on the task too) It's a bit mind-warping to think about, but genius once it clicks. Makes you wonder how else we should be designing AI systems as we can start consuming more and more tokens

@AlicanKiraz0 The 66% KV memory cut at 2048 tokens is the real unlock. That's the point where most 16GB MacBooks hit swap. Keeping the working set in-RAM at that depth changes what's actually runnable on consumer silicon.

Hi everyone 🙏🏻 After a long stretch of work, I’ve finalized my Qwen3.5-TurboQuant-MLX-LM project as a v0.1 Research Preview. With this work, I built a preview runtime that runs TurboQuant on Qwen3 / Qwen3.5 full-attention KV cache layers with MLX on Apple Silicon. TurboQuant is no longer just theoretical — it is now clearly visible in benchmark outputs as well. 🎉 On the Qwen3.5-9B-MLX-4bit smoke run, I validated the TurboQuant path with scorer_route = native_mlx. In the 512 prompt / 64 generation benchmark, compared to the oracle_preview fallback within the same turbomlx backend: - key-path memory: 45.41 MiB -> 27.44 MiB (-39.57%) - total KV: 54.40 MiB -> 36.43 MiB (-33.03%) - prompt TPS: 285.39 -> 380.04 (+33.16%) - decode TPS: 42.02 -> 42.71 (+1.65%) In the 2048 prompt / 64 generation benchmark: - key-path memory: 99.60 MiB -> 33.63 MiB (-66.23%) - total KV: 132.58 MiB -> 66.62 MiB (-49.76%) - prompt TPS: 285.52 -> 401.84 (+40.74%) - decode TPS: 39.50 -> 40.44 (+2.38%) I also made it visibly clear, through the native_working_set_bytes metric, that the TurboQuant working set is genuinely active. This release is a Research Preview. And as the results show, TurboQuant is active, measurable, and meaningfully changes the Qwen-first MLX runtime profile. 🙏🏻

@HuggingModels The model name is misleading. It's Qwen3.5 27B fine-tuned on reasoning traces from Claude outputs. Not Claude, not from Anthropic. 280k downloads means the naming confusion is working. Worth flagging before deploying in production.

Meet Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled: a powerhouse reasoning model that understands both text AND images. It's like giving AI a pair of eyes and a brilliant mind. The community is buzzing because this model actually shows its work, not just final answers.

@sudoingX Octopus invaders failing 5/5 is a useful signal. Game logic requires cross-token coherence across 100s of lines at multiple abstraction levels. 3B active MoE params can't maintain that state. Speed is the wrong optimization for this task class.

yesterday i tested nvidia's cascade 2 on a single RTX 3090. 187 tok/s, fastest model i've benchmarked in the 3B active class. but when i gave it octopus invaders, blank screens every time 5 times. 3B active MoE couldn't hold architectural coherence across thousands of lines of game logic. today i'm loading the next model in nvidia's nemotron family on the same 3090. openreasoning-nemotron 32B. this one is dense. not MoE/mamba. with 32 billion parameters all active on every token. and the architecture has a story. nvidia didn't build this from scratch. they took alibaba's qwen 2.5 32B as the base model, distilled deepseek R1's reasoning capabilities into it using 5 million reasoning traces, and released it under the nemotron name. three labs in one model. play is same RTX 3090 24GB. same octopus invaders test. same hermes agent. same prompt. the only variable is the model. cascade 2 gave me speed but couldn't code. this dense 32B trades speed for depth and every parameter fires on every token. if reasoning training from deepseek plus alibaba's architecture can build what nvidia's own mamba MoE couldn't, that changes the recommendation for everyone running a 3090. model downloading now. llama.cpp compiled. hermes agent configured. receipts incoming.

@iotcoi The shift is from stochastic to diagnostic search. 500 random variants vs. 5 targeted ones from failure signals. If the 'why did this fail' trace is accurate, convergence is 10x faster. YAML recursion is a detail. Signal quality is the actual moat.

Hermes built its own DSPy GEPA module Instead of brute-forcing 500 variants, it asks “why did this fail?” Builds a tree. Uses real signals. Converges fast. /gepa-collect >> optimize Runs fully local. No therapy notes leaked to APIs. Recursive self-improvement is now a YAML cfg

@anemll At 3.5 t/s the bottleneck is SSD read latency per expert activation, not FLOPS. K2 with 4 active experts needs 4 loads per token from SSD. The M5 Max ceiling is PCIe bandwidth. Compute sits idle waiting on memory. Expert hit rate matters here.

Kimi K2 1T params, 3.5 t/s M5 MAX 128GB Flash-MoE SSD streaming. This demo is K2 1.8-bit Unsloth Dynamic GGUFs set to 4 active experts. Both MLX and LAMMA.cpp supported with Dynamic Quants. Conversion is simple. Repos are incoming… Obviously smaller MoE models are faster :)