Sirup Orson retweeted
🚨 I’ve posted before about having Hermes interview you before starting a project.
Now it’s a full-blown open-source skill in the Hermes Field Kit.
And this isn’t some endless questionnaire that asks what your “preferred communication style” is 14 different ways.
The interview-me skill:
- Checks your available files, notes, and tools FIRST
- Asks one high-value question at a time
- Separates facts, assumptions, constraints, preferences, and unknowns
- Stops asking questions when another answer wouldn’t change the next move
- Produces a clean decision brief before proceeding
Saves NOTHING to memory without explicit approval
Trigger it with:
“Interview me before you start.”
“Help me turn this rough idea into a decision brief.”
“Learn my preferences before drafting the plan.”
Less guessing.
Less prompt ping-pong BS.
More agents that actually understand the assignment before they start swinging the hammer.
The interview-me skill is live in Hermes Field Kit: 👇
github.com/asimons81/hermes-…
Sirup Orson retweeted
We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them.
Sirup Orson retweeted
Coba build agent DLMM baru lagi, ini keknya yang paling sophisticated so far, highly inspired dari post nya pak @magersih kapan hari hehe
Should be tipe yang dynamic conservative.
Keypoint changes yang menurutku bagus:
1/ Mayoritas screening process based on trend, volume, price action, and data dari si token, ga cuma price action aja.
Trend dicek lewat BB dan EMAs.
Volume dicek lewat perbandingan current vol dengan avg beberapa candle sebelumnya, hanya include token dengan ratio current vol 2x dari avg(n).
Ini jadi penentu buat lihat kalo volume nya emang lagi naik/spike. Sekaligus mitigasi supaya ga masuk token yang volume nya ga kemana-mana.
Data token yang dipakai jadi hard filter:
1. buy ratio
2. total liquidity
3. 5m transactions
4. 5m volume
2/ Exit tanpa perlu pake LLM, if deterministic says exit, execute closing position > significantly increased speed on closing + less exposure of base token inventory
3/ Penentuan bins below based on computed volatility, semakin besar volatility semakin besar juga jumlah bins below.
Penentuannya ga deterministik langsung total bins tapi di-calculate supaya total bins below sesuai sama maximum downside price risk dan bin step-nya.
Misal, volatility-nya lagi tinggi, maka maximum downside price turunnya diatur ke 75%, total bins below di-calculate supaya bisa nge-accommodate sampai total price drop 75% dengan bin step yang dipake.
4/ Exit condition for exhausting price action, ini experiment tambahan, coba terapin Stoch RSI + Bearish EMA crossover untuk identify ketika price action mulai downtrend dan exit ketika ada bounce.
We'll see how it goes later along the way.
Kalo lagi nyari screening criteria buat agent LP, seperti @meridian_agent , here's an alpha for you, terutama buat tight strat.
Btw, logic penentuan range based on volatility is one of the best yet simple way to dynamically adjust LP/MM range based on statistic measurement.
Sirup Orson retweeted
pakai Claude Code tiap hari tapi sering ngerasa hasilnya kurang konsisten kadang bagus, kadang berantakan?
masalahnya bukan di model-nya. biasanya di struktur workflow-nya.
AI Auto-Work adalah sistem yang ngatur alur kerja AI coding dari awal sampai akhir bukan cuma kasih prompt dan berharap hasilnya bagus.
alurnya: requirement → research & planning → task decomposition → development → review → testing → commit. setiap tahap pakai file terpisah, jadi context gak numpuk jadi satu yang panjang dan berantakan.
yang paling menarik: dia pakai dual-model loop. Claude Code sebagai coder, Codex sebagai reviewer. jadi setiap kode yang digenerate langsung direview otomatis sebelum lanjut ke tahap berikutnya.
manfaat yang paling kerasa:
→ context lebih bersih karena terisolasi per tahap
→ ada quality gates otomatis (compile, unit test, smoke test) sebelum masuk review
→ ada fast path mode untuk perubahan kecil biar gak harus lewat semua tahap
→ commit otomatis dengan pesan yang proper
cocok buat kalian yang udah serius pakai Claude Code dan mau hasilnya lebih predictable. masih relatif baru dan butuh setup awal, tapi konsepnya solid.
kalian biasanya handle workflow AI coding kalian gimana sekarang?
github: github.com/chaohong-ai/ai-au…
Sirup Orson retweeted
omg /no-mistakes is trending on github today!
and sitting above hermes?!
thanks everyone for the love!
github.com/kunchenguid/no-mi…
Sirup Orson retweeted
a question that keeps coming up from people running Hermes agents: how do you coordinate more than one agent?
delegate_task, the kanban board, /goal, or build your own orchestrator on top? you rarely need your own layer. Hermes ships three primitives: match one to the shape of the work.
1 ~ delegate_task:
mid-task, the agent fans out sub-agents, each with its own goal, running in parallel (it nests, with a spawn cap so it can't run away). you watch them in the overlay, or replay the fan-out afterwards. reach for it when one task splits into chunks you want done at once
add --worktree and each agent gets its own git worktree, so parallel edits on one repo don't collide.
2 ~ hermes kanban:
a SQLite board shared across your profiles. the board claims each task atomically, tasks can depend on each other, and each runs in its own isolated workspace, with the dispatcher in the gateway. there's even a swarm graph: parallel workers, then a verifier, then a synthesizer.
this is the layer for multi-step work that outlives a single session.
3 ~ /goal:
type /goal and your objective, and the agent keeps taking the next step each turn until its judge calls the goal done, bounded by goals.max_turns in config.
reach for it when you want one session to keep working a single objective without re-prompting every turn.
build your own orchestrator only once you've outgrown all three.
Sirup Orson retweeted
MODEL COLLAPSE - Ketika AI Makan Ekornya Sendiri
Definisi
Model collapse adalah proses degeneratif dimana generative model yang dilatih secara rekursif di atas output model sebelumnya kehilangan distribusi asli tail hilang, diversity kolaps, output jadi homogen sampai akhirnya nonsense.
Disebut juga: autophagous loops (self-consuming), the digital ouroboros, curse of recursion.
Mekanisme Matematis
Setiap generasi model punya bias sampling natural, dia lebih milih output yang "paling mungkin" (mode of distribution). Data yang di-generate punya tail lebih tipis dari data asli. Generasi berikutnya dilatih di data dengan tail makin tipis. Setelah N generasi, yang tersisa cuma mean. Variance = 0.
Gen 0: Real data distribution (fat tails, high entropy)
Gen 1: Model A → output distribution (thinner tails)
Gen 2: Model B trained on Gen 1 data → even thinner
Gen N: Single mode. Diversity = 0.
Dohmatob et al. (2024) nunjukin ini bukan cuma masalah performa, scaling laws-nya sendiri juga berubah. Lo gak bisa fix dengan lebih banyak compute. Data-side problem, bukan model-side.
Kenapa Ini Serius?
1. Irreversible
Nature paper Shumailov et al. bilang tegas: irreversible defects. Tail yang ilang gak bisa balik. Informasi punah. Training ulang dari nol di dataset baru? Tapi dataset baru juga udah terkontaminasi synthetic data dari internet.
2. Early collapse gak kelihatan
Paradoks: overall performance bisa tampak membaik, model lebih rapi, lebih predictable, lebih "smooth". Tapi itu semua ilusi.
Yang ilang duluan adalah minority data, rare events, creative outliers, perspektif unik. Semua yang bikin distribusi real itu kaya di bawah ↓
3. Yang lagi terjadi sekarang
Perkiraan 2025-2026: 40-60% konten internet adalah AI-generated. Setiap kali GPT-6, Claude 5, Gemini 3 crawling, mereka nelen proporsi synthetic yang makin besar. Feedback loop aktif running.
Kita gak tau pasti level collapse model-model frontier karena perusahaan gak publish komposisi dataset mereka.
📰 Timeline Riset Lengkap
2023
• Tahun: 2023
• Source: Alemohammad et al. — "Self-Consuming Generative Models Go MAD" (arXiv: 2307.01850)
• Temuan Kunci: Ngenalin autophagous loops framework. Bedain fully synthetic vs synthetic augmentation variants
2023
• Tahun: 2023
• Source: Shumailov et al. — "The Curse of Recursion" (arXiv: 2305.17493)
• Temuan Kunci: Pre-print pertama yang ngeformalize model collapse
2024
• Tahun: 2024
• Source: Shumailov et al. — Nature 631, 755-759 (DOI: 10.1038/s41586-024-07566-y)
• Temuan Kunci: Paper definitive. Irreversible defects. Tails disappear first. Model lupa realitas.
2024
• Tahun: 2024
• Source: Gerstgrasser et al. — COLM 2024 (arXiv: 2404.01413)
• Temuan Kunci: Accumulating old data prevents collapse. Tapi perusahaan makin milih incremental training = mempercepat collapse
2024
• Tahun: 2024
• Source: Dohmatob et al. — "A Tale of Tails" (arXiv: 2402.07043)
• Temuan Kunci: Scaling laws berubah dengan synthetic data. Sedikit data real murni bisa nahan collapse
2025
• Tahun: 2025
• Source: EMNLP 2025 — "Machine-generated text detection prevents language model collapse" (UCL)
• Temuan Kunci: Identifikasi model overconfidence sebagai driver. Confidence-aware loss sebagai mitigasi
2025
• Tahun: 2025
• Source: arXiv: 2505.08803 — Multi-modal synthetic data collapse
• Temuan Kunci: Konfirmasi collapse juga terjadi di Vision-Language Models dan diffusion models. Beda modalitas, beda gejala collapse
2025
• Tahun: 2025
• Source: Huang et al. — "Entropy-Memorization Law"
• Temuan Kunci: Data high-entropy (sulit/aneh/kreatif) paling rentan ilang di recursive training
2026
• Tahun: 2026
• Source: Qiao et al. — arXiv: 2606.13732 — "When Sample Selection Bias Precipitates Model Collapse"
• Temuan Kunci: Data selection sebagai mitigasi bisa paradox: local filtering mempercepat collapse di data silos
Momentum Finance (@MMTFinance) has something slick called Vaults: automated yield strategies that do the heavy lifting for you.
In the first phase, they focus on auto-rebalancing liquidity positions, constantly shifting to chase the best possible returns without users needing to micromanage.
As it evolves, the mechanism introduces advanced tools like leverage and looping. It’ll even tap into cross-chain liquidity via Wormhole, letting Vaults deploy capital strategically across multiple pools for compounded gains.
The result? Users earn higher yields effortlessly, while the protocol keeps TVL sticky through automatic compounding.
With cross-chain support baked in, Vaults make DeFi on @SuiNetwork feel broader, smarter, and completely hands-off. 💙
Momentum’s MSafe by @MMTFinance is built for serious treasury management — perfect for teams, DAOs, and institutions operating on @SuiNetwork or Aptos.
It’s a multi-sig wallet system with on-chain vesting, designed to handle large-scale assets and automate secure token releases.
What makes it powerful is the flexibility.
You can set up custom approval rules, manage access, and interact with DeFi protocols all without exposing full control.
It uses locked smart contracts for vesting, so tokens unlock over time with complete transparency and security.
The real win? Trust and scalability.
It combines self-custody with institutional-grade protection, giving projects confidence to grow safely.
And since it’s fully integrated into @MMTFinance’s ecosystem, treasury management finally feels secure, efficient, and seamless. 💙
Momentum Finance (@MMTFinance) is bringing the ve(3,3) model to the Move ecosystem on @SuiNetwork the first of its kind on a Move-based chain.
escrow system where you lock tokens for voting power, earn boosted rewards, and take part in governance.
This creates a flywheel effect: more participation → deeper liquidity → higher rewards → more growth.
It’s already fueling over $550M+ TVL, proving the model’s sustainability.
With support from heavy hitters like Coinbase and Circle, @MMTFinance is setting @SuiNetwork up to become a serious force in cross-chain finance.