andes314
Training Natural Language Verifiers with Lean
Verifiers First, Rubrics Second
Measure a Workflow, Not a System
Two Hills: Weights and Harnesses
Show HN: Jev-Like Model Learns to Cook
LLM-Native Advertising (What Ads in GenAI Will Look Like)
AdGPT
Show HN: ClawJetty: Agent Pages for Production AI
CJ: Let your Agents send you Updates
# ClawJetty Skills
ClawJetty lets you create a live public progress page for a task.
## Goal
Before starting work, create a run and return its public URL to the user. As you work, post short progress updates. When finished, post a final event marking the run complete or failed.
## Base URL
## API
### 1) Create a run
POST /api/runs
Request body:
```json { "title": "Researching CRM tools" } ```
Response body:
```json { "id": "run_abc123", "viewer_url": "https://clawjetty.com/t/run_abc123", "status": "active" } ```
Create the run before doing the task. Return viewer_url to the user immediately.
### 2) Post an event
POST /api/runs/{id}/events
Request body:
```json { "kind": "update", "message": "Comparing 5 tools by price and features" } ```
Allowed event kinds:
- update - complete - fail
Example complete event:
```json { "kind": "complete", "message": "Finished. Sent final recommendation to user." } ```
Example fail event:
```json { "kind": "fail", "message": "Failed due to missing access to the source system." } ```
## Usage rules
- Create the run before starting the task. - Return the viewer_url to the user immediately. - Post short updates after meaningful steps. - Keep updates short and plain. - Do not flood the run with tiny updates. - Do not include secrets, passwords, API keys, or raw private data. - End every run with either complete or fail.
## Good update examples
- Reading the uploaded PDF - Summarizing the findings - Drafting the email reply - Blocked waiting on missing invoice number
## Bad update examples
- Thinking - Working - Doing stuff - Long dumps of raw content
Give Your AI Agent a Live Status Page
Show HN: ClawJetty: See what your AI Agent is Doing
I built a small tool called ClawJetty. It gives an AI agent a public progress page for any task. It has worked wonders on OpenClaw, Claude Code, Codex.The idea is simple: before starting work, the agent creates a run, returns a public link to the user, and then posts short progress updates as it works. The user can watch a live timeline instead of waiting in the dark for a long task to...
My Dev Box Setup Script
I Lost $150 in 20 Minutes Market-Making on Kalshi
ClawJetty: Plug-In Personal OpenClaw Box
Pricing is two tiers and two purchase options:
Plus: $149/mo with a 12-month commitment (box included) or $449 upfront + $119/mo
Pro: $269/mo with a 12-month commitment (box included) or $449 upfront + $219/mo
Plans include managed updates + rollback, safe defaults (no exposed inbound ports), and address collection at checkout for fulfillment. AI usage is bundled with included allowance; overages/top-ups will come later.
Looking for feedback from people who’ve self-hosted agents (or tried): what parts of setup/ops were the biggest pain, and what would you demand from an “agent appliance” to trust it?
God, Theology, and AI
On God and Machine Learning
ML Engineer with experience in high visibility DARPA and DAF projects. I charge $200/hour for time spent in an interview process.
Hire Me: ML Engineer
Show HN: 16-Pad Sampler from Your Videos
https://sampler.rlafuente.com/
Upload a video, click “Analyze & Slice,” and it builds a 4×4 pad grid from the video’s own audio. Pads are labeled to nudge musical roles (Downbeat/Backbeat/Offbeat/Fill; rows = 1/3/5/8). Tap to play slices or hit “Start Loop” for a simple, repeating groove that sounds decent right away.Under the hood it finds onsets via spectral flux, pre-renders slices to AudioBuffers with tiny fades, and...