Hacker News

Article URL: https://mina86.com/2019/srgb-xyz-conversion/

Comments URL: https://news.ycombinator.com/item?id=47214478

Points: 2

# Comments: 0

LTX-2 is an open-source diffusion model that combines video and audio.

Visually it's not at the level of Seedance 2.0, Veo 3.1, or Sora 2, but it’s open-weights, so anyone can play with it.

I wanted to see how good it is at generating video from just audio.

Off-the-shelf, it's not very good, but I found that if you run the audio through Gemini to generate a prompt, then feed that into LTX-2, in addition to the audio, the output matches the audio much more often.

Foley sounds work particularly well, and one fun use case is uploading audio of yourself to see what AI thinks you look like.

Limitations:

- Doesn't know real people, so a famous person's voice just gets a generic person

- Sometimes gets gender wrong if the voice is more androgynous

- In dialogue with similar voices, it can render the same person saying both lines

Comments URL: https://news.ycombinator.com/item?id=47214472

Points: 9

# Comments: 0

# Elevator Pitch

*Show HN: L-Rep – Geometry engine, GA multivectors, dynamic trees in a single integer*

Imagine every SDF/CSG scene graph, every Conformal Geometric Algebra (CGA) motor, every lock-free octree/BVH node, every piecewise-Chebyshev approximation with provable error bound, and both forward + reverse autodiff — all living inside one single mixed-radix Python int (or future 1024-bit hardware word).

We built L-Representation: a tiny ISA + L-VM + L-ALU micro-architecture that treats this integer as a universal register.

JIT-compiles your operator tree straight into fused L_FIELD_OP + L_GA_MUL primitives

Native L_GA_MUL does dense 5D CGA multiply in a handful of wide-field MACs (hardware estimate: 8–64× faster than current GA libs on FPGA/ASIC)

Dynamic trees are encoded as B-chunked single-integer pointers with atomic L_ALLOC/L_INSERT/L_SPLIT_MERGE — zero pointer chasing, perfect for lock-free multi-thread or hardware

Transcendentals are auto-replaced by piecewise Chebyshev with compiler-guaranteed ε-total error; autodiff is just dual-field layout or a compact tape of the same primitives

Prototype (single-file, included in the .tex) already shows 10–100× per-query speedup on localized SDF raymarching and GA ops vs naive decode-eval-encode.

This is not another math paper. This is an architecture-level unification — the same way RISC-V or LLVM became the substrate for everything else. If anyone with serious FPGA/ASIC budget builds the L-ALU tile, every graphics, robotics, physics sim, and scientific-computing stack suddenly gets a free 10-100× geometry accelerator with provable correctness baked in.

Repo: Paper + full single-file prototype + reproducible benches: + https://github.com/nahhididwin/L-Representation/blob/main/ma...

+ https://github.com/nahhididwin/L-Representation

Comments URL: https://news.ycombinator.com/item?id=47214471

Points: 3

# Comments: 0

I've been experimenting with long-running OpenClaw agents on dedicated ClawHost instances and wanted to share what's possible when you give an AI agent full server access and let it run autonomously.

One of my agents manages a complete blogging pipeline. It writes articles about OpenClaw, generates images using Nano Banana 2, handles the full git workflow (branch, merge, deploy), triggers Vercel rebuilds, and notifies me on Telegram for every action it takes. All of this runs 24/7 without manual intervention.

What makes this work is giving the agent a real environment to operate in. Full SSH access, no sandbox restrictions, full control over git, APIs, and deployment pipelines. That's the difference between a chatbot and an actual autonomous agent.

The interesting part: this agent is now contributing to the very platform that hosts it. The platform deploys the agent, and the agent builds the platform. That recursive loop is where things start to feel like a shift in how we think about AI infrastructure.

I'm curious how others are approaching long-lived autonomous agents. How do you handle reliability, monitoring, and the trust boundary when an agent has real access to production systems?

You can read the articles: https://clawhost.cloud/blog

Comments URL: https://news.ycombinator.com/item?id=47214461

Points: 2

# Comments: 0

Article URL: https://theartificialq.github.io/2026/02/28/strix-first-impressions.html

Comments URL: https://news.ycombinator.com/item?id=47214440

Points: 1

# Comments: 1

Article URL: https://www.fivewhys.ai/

Comments URL: https://news.ycombinator.com/item?id=47214254

Points: 1

# Comments: 0

Article URL: https://www.bbc.com/future/article/20260226-how-ai-can-read-your-thoughts

Comments URL: https://news.ycombinator.com/item?id=47214250

Points: 1

# Comments: 0

Article URL: https://www.letitsimmer.co

Comments URL: https://news.ycombinator.com/item?id=47214228

Points: 1

# Comments: 1

Article URL: https://github.com/tfatykhov/nous

Comments URL: https://news.ycombinator.com/item?id=47214204

Points: 1

# Comments: 0

Article URL: https://www.youtube.com/watch?v=ShusuVq32hc

Comments URL: https://news.ycombinator.com/item?id=47214185

Points: 2

# Comments: 0

The more I used Claude Code, the messier my ~/.claude got.

I started out thinking one global config was fine — just put the skills and MCP servers everyone needs at the user level and call it done. But over time I was wearing two hats: a frontend developer at my day job, and a fullstack developer on side projects. The configs started bleeding into each other.

The specific frustrations that pushed me over the edge: - Project-level settings had to be redone for every new repo - Marketplace and plugin-based configs needed manual toggling depending on which context I was in - Trying an experimental setup, then cleaning it up afterward, was tedious every time

Beyond config mess, the bigger realization was about personas. As I started doing more AI agent development, my developer identity kept splitting. More roles, more personas — and a single global Claude Code environment can't cleanly represent all of them.

That's when I thought: nvm and pyenv let you switch environments by profile. Why not Claude Code? So I built clenv.

clenv manages multiple Claude Code profiles. Each profile is an isolated ~/.claude directory (CLAUDE.md, MCP servers, hooks, agents, skills) backed by its own git repository. Free and open source (MIT).

clenv init # backs up ~/.claude, creates default profile clenv profile create work --use # create + switch instantly clenv profile create agent-prod --from agent-dev # clone from existing clenv commit -m "add GitHub MCP server" clenv diff HEAD~1..HEAD clenv log --oneline clenv revert abc123f clenv tag v1.0 -m "production agent config" Teams can export a baseline and let members layer personal changes on top:

clenv profile export team-standard -o team.clenvprofile clenv profile import team.clenvprofile --use MCP API keys are automatically redacted during export.

Per-directory auto-switching works like .nvmrc:

clenv rc set work # pin profile to this directory clenv rc show Written in Rust, statically linked, zero runtime deps. macOS and Linux.

brew tap Imchaemin/clenv && brew install clenv cargo install clenv GitHub: https://github.com/Imchaemin/clenv

Would especially love feedback from people doing AI agent development — that's the use case where environment isolation feels most important.

Comments URL: https://news.ycombinator.com/item?id=47214169

Points: 1

# Comments: 0

Article URL: https://www.theartsjournal.org/index.php/site/article/download/261/190/709

Comments URL: https://news.ycombinator.com/item?id=47214156

Points: 1

# Comments: 0

Article URL: https://addyosmani.com/blog/factory-model/

Comments URL: https://news.ycombinator.com/item?id=47214150

Points: 1

# Comments: 0

Article URL: https://huggingface.co/blog/alexshengzhili/aiscientist

Comments URL: https://news.ycombinator.com/item?id=47214141

Points: 1

# Comments: 0

Article URL: https://old.reddit.com/r/paypal/comments/1rifn18/paypal_apparently_does_not_understand_how_gmail/

Comments URL: https://news.ycombinator.com/item?id=47214127

Points: 1

# Comments: 1

Article URL: https://globe.gl/example/satellites/index.html

Comments URL: https://news.ycombinator.com/item?id=47214119

Points: 1

# Comments: 0

Article URL: https://github.com/actionbook/actionbook

Comments URL: https://news.ycombinator.com/item?id=47214116

Points: 1

# Comments: 0

Article URL: https://globe.gl/example/moon-landing-sites/index.html

Comments URL: https://news.ycombinator.com/item?id=47214110

Points: 1

# Comments: 0

Skills for OpenAI Codex, Claude Code, Gemini CLI, Cursor and more.

Comments URL: https://news.ycombinator.com/item?id=47214068

Points: 1

# Comments: 0

Article URL: https://blog.tohojo.dk/2026/02/the-inner-workings-of-tcp-zero-copy.html

Comments URL: https://news.ycombinator.com/item?id=47214064

Points: 2

# Comments: 0