Positive HackerNews

Incident Report: May 19, 2026 – GCP Account Suspension

Wed, 20 May 2026 08:37:55 -0000

Previous thread: Incident Report: Railway Blocked by Google Cloud [resolved] - https://news.ycombinator.com/item?id=48201484

SBCL: the ultimate assembly code breadboard (2014)

Wed, 20 May 2026 15:39:48 -0000

No Text

Qwen3.7-Max: The Agent Frontier

Wed, 20 May 2026 10:35:02 -0000

No Text

Saying Goodbye to Asm.js

Wed, 20 May 2026 12:01:56 -0000

No Text

Map of Metal

Wed, 20 May 2026 10:47:20 -0000

No Text

Victory: Tennessee man jailed 37 days for Trump meme wins $835,000 settlement

Wed, 20 May 2026 14:30:47 -0000

No Text

How fast is N tokens per second really?

Mon, 18 May 2026 02:04:38 -0000

No Text

Testing distributed systems with AI agents

Wed, 20 May 2026 14:40:42 -0000

No Text

Everything in C is undefined behavior

Wed, 20 May 2026 06:07:22 -0000

No Text

Formal Verification Gates for AI Coding Loops

Wed, 20 May 2026 15:25:45 -0000

No Text

Gemini 3.5 Flash

Tue, 19 May 2026 17:43:45 -0000

https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flas...

When Fast Fourier Transform Meets Transformer for Image Restoration (2024)

Mon, 18 May 2026 14:10:47 -0000

No Text

Stable Audio 3

Wed, 20 May 2026 15:10:05 -0000

No Text

FiveThirtyEight articles on the Internet Archive

Wed, 20 May 2026 01:34:19 -0000

No Text

Show HN: Hocuspocus 4 – self-hosted Yjs collaboration backend

Wed, 20 May 2026 14:51:00 -0000

Hi HN! I'm Philip, one of the founders of Tiptap. Next to our open-source rich text editor framework, we started developing Hocuspocus about five years ago and open-sourced it too, to solve one of our biggest challenges back then: real-time collaboration in web editors. We found Yjs by Kevin Jahns, a CRDT library that handles concurrent edits without conflicts. Basically, Yjs merges changes from users without conflicts and in real-time. Hocuspocus is the WebSocket server built on top of Yjs. It handles real-time sync, presence/awareness, persistence, and Redis-based scaling.

While we use Hocuspocus at Tiptap as the collaboration backend for our cloud services, it also works with any Yjs client (Slate, Quill, Monaco, ProseMirror, or your own setup), and Yjs documents aren't limited to text at all. You can sync any structured data through them, and in the meantime we see projects that rely on Hocuspocus without using the Tiptap editor.

We released Hocuspocus v4 under the MIT license a few weeks ago, and the biggest change is that it's no longer tied to Node. The previous versions depended on the ws package, which meant you couldn't run Hocuspocus on Bun, Deno, or Cloudflare Workers. We moved to crossws, a universal websocket adapter, so the same server now runs on Node, Bun, Deno, Cloudflare Workers, and Node with uWebSockets. That also lets you run collaboration at the edge.

The other changes are smaller but are important if you're using Hocuspocus in production:

1. Every core class and hook payload takes a generic Context type now, so the auth/session shape you build in onAuthenticate flows through every other hook with full type safety (defaults to any so existing code doesn't break).

2. Document updates are now processed sequentially per connection through an internal queue, which fixes a correctness bug where async hooks could cause CRDT updates to apply out of order under load.

3. Transaction origins are structured objects now with a source field instead of raw values and there's an isTransactionOrigin() helper for narrowing.

4. Hook payloads use web-standard Request and Headers instead of Node's IncomingMessage.

5. The wire protocol is backward compatible in both directions, so you can roll out servers and providers independently.

If you want to test Hocuspocus: npm install @hocuspocus/server @hocuspocus/provider

Docs at: https://tiptap.dev/docs/hocuspocus

Source at: https://github.com/ueberdosis/hocuspocus

Because running real-time collaboration on Workers or Durable Objects is new in v4, that's the use case we'd most like to hear your questions and feedback on.

Show HN: Lance – image/video generation and understanding in one model

Wed, 20 May 2026 15:45:32 -0000

The model has 3B active parameters. We put the code, homepage, paper and model links here:

- Code: https://github.com/bytedance/Lance

- Homepage: https://lance-project.github.io/

- Paper: https://arxiv.org/abs/2605.18678

- Model: https://huggingface.co/bytedance-research/Lance

p.s. Lance is a research project, not a polished product. The model was trained using fewer than 128 GPUs.

Autoregressive next token prediction and KV Cache in transformers

Sun, 17 May 2026 20:07:14 -0000

No Text

Handling the great code forge fragmentation

Sun, 17 May 2026 15:31:03 -0000

No Text

Smartmedia Card Spec Opened, available free (2000)

Sun, 17 May 2026 19:22:42 -0000

No Text

Goodbye Visa and Mastercard: 130M Europeans switching to sovereign payment

Wed, 20 May 2026 13:02:30 -0000

No Text

OpenAI Is Preparing to File for an IPO Soon

Wed, 20 May 2026 16:24:42 -0000

No Text

Show HN: Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks

Tue, 19 May 2026 12:23:07 -0000

Hi HN, I'm Antoine Zambelli, AI Director at Texas Instruments.

I built Forge, an open-source reliability layer for self-hosted LLM tool-calling.

What it does:

- Adds domain-and-tool-agnostic guardrails (retry nudges, step enforcement, error recovery, VRAM-aware context management) to local models running on consumer hardware

- Takes an 8B model from ~53% to ~99% on multi-step agentic workflows without changing the model - just the system around it

- Ships with an eval harness and interactive dashboard so you can reproduce every number

I wanted to run a handful of always-on agentic systems for my portfolio, didn't want to pay cloud frontier costs, and immediately hit the compounding math problem on local models. 90% per-step accuracy sounds great, but with a 5-step workflow that's a 40% failure rate. No existing framework seemed to address this mechanical reliability issue - they all seemed tailor-made for cloud frontier.

Demo video: https://youtu.be/MzRgJoJAXGc (side-by-side: same model, same task, with and without Forge guardrails)

The paper (accepted to ACM CAIS '26, presenting May 26-29 in San Jose) covers the peer-reviewed findings across 97 model/backend configurations, 18 scenarios, 50 runs each. Key numbers:

- Ministral 8B with Forge: 99.3%. Claude Sonnet with Forge: 100%. The gap between a free local 8B model on a $600 GPU and a frontier API is less than 1 point.

- The same 8B local model with Forge (99.3%) outperforms Claude Sonnet without guardrails (87.2%) - an 8B model with framework support beats the best result you can get through frontier API alone.

- Error recovery scores 0% for every model tested - local and frontier - without the retry mechanism. Not a capability gap, an architectural absence.

I'm currently using this for my home assistant running on Ministral 14B-Reasoning, and for my locally hosted agentic coding harness (8B managed to contribute to the codebase!).

The guardrail stack has five layers, each independently toggleable. The two that carry the most weight (per ablation study with McNemar's test): retry nudges (24-49 point drops when disabled) and error recovery (~10 point drops, significant for every model tested). Step enforcement is situational - only fires for models with weaker sequencing discipline. Rescue parsing and context compaction showed no significance in the eval but are retained for production workloads where they activate once in a while.

One thing I really didn't expect: the serving backend matters. Same Mistral-Nemo 12B weights produce 7% accuracy on llama-server with native function calling and 83% on Llamafile in prompt mode. A 75-point swing from infrastructure alone. I don't think anyone's published this because standard benchmarks don't control for serving backend.

Another surprise: there's no distinction in current LLM tool-calling between "the tool ran successfully and returned data" and "the tool ran successfully but found nothing." Both return a value, the orchestrator marks the step complete, and bad data cascades downstream. It's the equivalent of HTTP having 200 but no 404. Forge adds this as a new exception class (ToolResolutionError) - the model sees the error and can retry instead of silently passing garbage forward.

Biggest technical challenge was context compaction for memory-constrained hardware. Both Ollama and Llamafile silently fall back to CPU when the model exceeds VRAM - no warning, no error, just 10-100x slower inference. Forge queries nvidia-smi at startup and derives a token budget to prevent this.

How to try it:

- Clone the repo, run the eval harness on a model I haven't tested. If you get interesting results I'll add them to the dashboard.

- Try the proxy server mode - point any OpenAI-compatible client at Forge and it handles guardrails transparently. It's the newest model and I'd love more eyes on it.

- Dogfooding led me to optimize model parameters in v0.6.0. The harder eval suite (26 scenarios) is designed to raise the ceiling so no one sits at 100%. Several that did on the original suite can't sweep it - including Opus 4.6. Curious if anyone finds scenarios that expose gaps I haven't thought of. Paper numbers based on pre v0.6.0 code.

Background: prior ML publication in unsupervised learning (83 citations). This paper accepted to ACM CAIS '26 - presenting May 26-29.

Repo: https://github.com/antoinezambelli/forge

Paper: https://www.caisconf.org/program/2026/demos/forge-agentic-re... https://github.com/antoinezambelli/forge/blob/main/docs/forg...

Dashboard: https://github.com/antoinezambelli/forge/docs/results/dashbo...

Infomaniak transitions to a foundation model to protect user data privacy

Wed, 20 May 2026 05:43:51 -0000

No Text