“Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning” Few-step diffusion and flow LMs usually require distilling a pretrained teacher, which adds training cost and caps the student at teacher quality. So this paper learns a direct path from noise to discrete tokens, generating the whole sequence in one step without a teacher model or tracking diffusion time. Then when you add extra steps, the model doesn’t just take smaller steps toward an answer. It revisits uncertain tokens and fixes them using the tokens it’s already confident about.

Discretionary traders should think carefully about instrument selection As in spot vs spot margin vs perps vs options and combinations therein The obvious components are cost, liquidity, and capital efficiency. The non-obvious components are drawdown tolerance, volatility tolerance, and trade management. From personal experience: Perps are excellent for intraday trading, but for swing trading and larger bets my performance was worse because I would over-manage those positions and be much more sensitive to drawdowns etc. Spot/spot margin are great for larger bets and higher time frame swing trades. I'd find myself less concerned with the tick-by-tick movements so I could actually hold the trade, but whenever I've tried to LTF trade those instruments, I would get complacent with trade management. The tempting default answer is perps for liquidity + to size up, but your bigger position isn't helpful if it means you sabotage yourself by managing it poorly I'm certain that you can think of trades that were great ideas but poorly executed - instrument selection may be the culprit

🧩 "Discretion on the outside, determinism on the inside." Prompt-only procedure following is brittle, and this approach fixes it by merging Skills with an embedded interpreter into executable code. Title: Building workflows for agents with Skills and Interpreter URL: 📝 Overview This post explains Interpreter Skills, which unify reusable behavior packages (Skills) with an embedded TypeScript runtime (Interpreter) that runs alongside the agent harness. A SKILL.md says when to apply a skill and an index.ts says how to execute it; the agent only decides whether the skill applies and what inputs to pass, while the module handles deterministic execution. ❓ Challenges Solved Agents are good at discretionary decisions but bad at deterministic procedures. Prompt-only procedure following is brittle, with agents skipping or reordering steps. On complex multi-step routines like processing 300+ items, making the model maintain coherence across context windows causes "context anxiety." 💡 Methodology & Proposed Approach ・Skills use progressive disclosure: agents scan a compact skill list, read full docs only for relevant ones, and keep skills as versioned, shareable units separate from the prompt ・The Interpreter restricts access by default; filesystem, network, tools, and subagents are usable only where explicitly exposed ・Skill modules programmatically spawn and manage subagents from code, orchestrating complex task graphs from code rather than through model-mediated steps ・Local operations like parsing, filtering, and grouping live in TypeScript code, keeping the tool surface lean and model-manageable 🎯 Use Cases It fits stateful multi-step workflows, like a GitHub triage that fetches issues, PRs, and discussions, spawns a subagent per item to summarize, and uses another to classify and cluster them. 📊 Evaluation and Significance ・Instead of "did it generally follow instructions?", you can ask the concrete question "did it call the expected function?", measuring whether procedures ran with correct inputs ・The model invokes once while the module orchestrates the whole workflow deterministically and returns a compact structured object ・The model keeps strategic control while critical procedures run in reviewable, testable code, shifting agent work toward software engineering practices: versioning, testing, and code review #AIAgents# #DevTools#