Popular Posts

Popular Content

Powered by Blogger.

Search This Blog

Blog Archive

Follow on Google+

Recent Posts

About us

December 29, 2025

OP here.

Most Deep Learning approaches for TSP rely on pre-training with large-scale datasets. I wanted to see if a solver could learn "on the fly" for a specific instance without any priors from other problems.

I built a solver using PPO that learns from scratch per instance. It achieved a 1.66% gap on TSPLIB d1291 in about 5.6 hours on a single A100.

The Core Idea: My hypothesis was that while optimal solutions are mostly composed of 'minimum edges' (nearest neighbors), the actual difficulty comes from a small number of 'exception edges' outside of that local scope.

Instead of pre-training, I designed an inductive bias based on the topological/geometric structure of these exception edges. The agent receives guides on which edges are likely promising based on micro/macro structures, and PPO fills in the gaps through trial and error.

It is interesting to see RL reach this level without a dataset. I have open-sourced the code and a Colab notebook for anyone who wants to verify the results or tinker with the 'exception edge' hypothesis.

Code & Colab: https://github.com/jivaprime/TSP_exception-edge

Happy to answer any questions about the geometric priors or the PPO implementation!

Comments URL: https://news.ycombinator.com/item?id=46420670

Points: 4

# Comments: 0



from Hacker News: Front Page https://ift.tt/jQL9orS
Continue Reading
December 25, 2025

I wanted to share this fun craft activity for the holidays that I've been doing with my family over the last few years. I came up with these while cutting up some cans trying to make an aluminum version of paper spinners.

There are a variety of shapes that work, but generally bigger+lighter spinners are better. Also incandescent bulbs are the best, but LEDs work too.

They remind me of candle carousels I would see at my grandparents' house during Christmas. Let me know what you think!

Comments URL: https://news.ycombinator.com/item?id=46379145

Points: 12

# Comments: 0



from Hacker News: Front Page https://ift.tt/Kzi6j4o
Continue Reading
December 20, 2025

I was looking for some fun project to play around with the latest Gemini models and ended up building this :)

Enter your username and get:

- Generated roasts and stats based on your HN activity 2025

- Your personalized HN front page from 2035 (inspired by a recent Show HN [0])

- An xkcd-style comic of your HN persona

It uses the latest gemini-3-flash and gemini-3-pro-image (nano banana pro) models, which deliver pretty impressive and funny results.

Give it a try and let me know what you think :)

[0] https://news.ycombinator.com/item?id=46205632

Comments URL: https://news.ycombinator.com/item?id=46336104

Points: 8

# Comments: 1



from Hacker News: Front Page https://ift.tt/QGUItvs
Continue Reading
December 18, 2025

Hi HN, We are the OpenDCAI group from Peking University. We built Paper2Any, an open-source tool designed to automate the "Paper to Slides" workflow based on our DataFlow-Agent framework. The Problem: Writing papers is hard, but creating professional architecture diagrams and slides (PPTs) is often more tedious. Most AI tools just generate static images (PNGs) that are impossible to tweak for final publication. The Solution: Paper2Any takes a PDF, text, or sketch as input, understands the research logic, and generates fully editable PPTX (PowerPoint) files and SVGs. We prioritize flexibility and fidelity—allowing you to specify page ranges, switch visual styles, and preserve original assets. How it works: 1. Multimodal Reading: Extracts text and visual elements from the paper. You can now specify page ranges (e.g., Method section only) to focus the context and reduce token usage. 2. Content Understanding: Identifies core contributions and structural logic. 3. PPT Generation: Instead of generating one flat image, it generates independent elements (blocks, arrows, text) with selectable visual styles and organizes them into a slide layout. Links: - Demo: http://dcai-paper2any.cpolar.top/ - Code (DataFlow-Agent): https://github.com/OpenDCAI/DataFlow-Agent We'd love to hear your feedback on the generation quality and the agent workflow!

Comments URL: https://news.ycombinator.com/item?id=46315047

Points: 5

# Comments: 0



from Hacker News: Front Page https://ift.tt/zVnuaSE
Continue Reading
December 16, 2025

Hi HN, I’m Andrew, Founder of Zencoder.

While building our IDE extensions and cloud agents, we ran into the same issue many of you likely face when using coding agents in complex repos: agents getting stuck in loops, apologizing, and wasting time.

We tried to manage this with scripts, but juggling terminal windows and copy-paste prompting was painful. So we built Zenflow, a free desktop tool to orchestrate AI coding workflows.

It handles the things we were missing in standard chat interfaces:

Cross-Model Verification: You can have Codex review Claude’s code, or run them in parallel to see which model handles the specific context better.

Parallel Execution: Run five different approaches on a backlog item simultaneously—mix "Human-in-the-Loop" for hard problems with "YOLO" runs for simple tasks.

Dynamic Workflows: Configured via simple .md files. Agents can actually "rewire" the next steps of the workflow dynamically based on the problem at hand.

Project list/kanban views across all workload

What we learned building this

To tune Zenflow, we ran 100+ experiments across public benchmarks (SWE-Bench-*, T-Bench) and private datasets. Two major takeaways that might interest this community:

Benchmark Saturation: Models are becoming progressively overtrained on all versions of SWE-Bench (even Pro). We found public results are diverging significantly from performance on private datasets. If you are building workflows, you can't rely on public benches.

The "Goldilocks" Workflow: In autonomous mode, heavy multi-step processes often multiply errors rather than fix them. Massive, complex prompt templates look good on paper but fail in practice. The most reliable setups landed in a narrow “Goldilocks” zone of just enough structure without over-orchestration.

The app is free to use and supports Claude Code, Codex, Gemini, and Zencoder.

We’ve been dogfooding this heavily, but I'd love to hear your thoughts on the default workflows and if they fit your mental model for agentic coding.

Download: https://zencoder.ai/zenflow YT flyby: https://www.youtube.com/watch?v=67Ai-klT-B8

Comments URL: https://news.ycombinator.com/item?id=46290617

Points: 9

# Comments: 4



from Hacker News: Front Page https://ift.tt/rgSF0o5
Continue Reading
December 11, 2025

Hey HN, Waleed here. We're building Sim (https://sim.ai/), an open-source visual editor to build agentic workflows. Repo here: https://github.com/simstudioai/sim/. Docs here: https://docs.sim.ai.

You can run Sim locally using Docker, with no execution limits or other restrictions.

We started building Sim almost a year ago after repeatedly troubleshooting why our agents failed in production. Code-first frameworks felt hard to debug because of implicit control flow, and workflow platforms added more overhead than they removed. We wanted granular control and easy observability without piecing everything together ourselves.

We launched Sim [1][2] as a drag-and-drop canvas around 6 months ago. Since then, we've added:

- 138 blocks: Slack, GitHub, Linear, Notion, Supabase, SSH, TTS, SFTP, MongoDB, S3, Pinecone, ...

- Tool calling with granular control: forced, auto

- Agent memory: conversation memory with sliding window support (by last n messages or tokens)

- Trace spans: detailed logging and observability for nested workflows and tool calling

- Native RAG: upload documents, we chunk, embed with pgvector, and expose vector search to agents

- Workflow deployment versioning with rollbacks

- MCP support, Human-in-the-loop block

- Copilot to build workflows using natural language (just shipped a new version that also acts as a superagent and can call into any of your connected services directly, not just build workflows)

Under the hood, the workflow is a DAG with concurrent execution by default. Nodes run as soon as their dependencies (upstream blocks) are satisfied. Loops (for, forEach, while, do-while) and parallel fan-out/join are also first-class primitives.

Agent blocks are pass-through to the provider. You pick your model (OpenAI, Anthropic, Gemini, Ollama, vLLM), and and we pass through prompts, tools, and response format directly to the provider API. We normalize response shapes for block interoperability, but we're not adding layers that obscure what's happening.

We're currently working on our own MCP server and the ability to deploy workflows as MCP servers. Would love to hear your thoughts and where we should take it next (:

[1] https://news.ycombinator.com/item?id=43823096

[2] https://news.ycombinator.com/item?id=44052766

Comments URL: https://news.ycombinator.com/item?id=46234186

Points: 5

# Comments: 0



from Hacker News: Front Page https://ift.tt/M28UGFK
Continue Reading
Subscribe to: Comments (Atom)