Railway Blocked by Google Cloud

<a href="https:&#x2F;&#x2F;ai.google.dev&#x2F;gemini-api&#x2F;docs&#x2F;models&#x2F;gemini-3.5-flash" rel="nofollow">https:&#x2F;&#x2F;ai.google.dev&#x2F;gemini-api&#x2F;docs&#x2F;models&#x2F;gemini-3.5-flas...</a>

Gemini 3.5 Flash

Ben Welsh made an index of all FiveThirtyEight articles on the Internet Archive

I’ve built a virtual museum with nearly every operating system you can think of

<a href="https:&#x2F;&#x2F;www.nytimes.com&#x2F;2026&#x2F;05&#x2F;19&#x2F;business&#x2F;google-seach-bar-ai-gemini.html?unlocked_article_code=1.jlA.95yh.ptfBUHf-rBtB&amp;smid=url-share" rel="nofollow">https:&#x2F;&#x2F;www.nytimes.com&#x2F;2026&#x2F;05&#x2F;19&#x2F;business&#x2F;google-seach-bar...</a>, <a href="https:&#x2F;&#x2F;archive.ph&#x2F;XI1sQ" rel="nofollow">https:&#x2F;&#x2F;archive.ph&#x2F;XI1sQ</a><a href="https:&#x2F;&#x2F;techcrunch.com&#x2F;2026&#x2F;05&#x2F;19&#x2F;google-search-as-you-know-it-is-over&#x2F;" rel="nofollow">https:&#x2F;&#x2F;techcrunch.com&#x2F;2026&#x2F;05&#x2F;19&#x2F;google-search-as-you-know-...</a><a href="https:&#x2F;&#x2F;www.theverge.com&#x2F;tech&#x2F;932970&#x2F;google-search-ai-update-io-2026" rel="nofollow">https:&#x2F;&#x2F;www.theverge.com&#x2F;tech&#x2F;932970&#x2F;google-search-ai-update...</a>

Google changes its search box

Remove–AI–Watermarks – CLI and library for removing AI watermarks from images

Hi HN, I&#x27;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 numberI wanted to run a handful of always-on agentic systems for my portfolio, didn&#x27;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&#x27;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: <a href="https:&#x2F;&#x2F;youtu.be&#x2F;MzRgJoJAXGc" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;MzRgJoJAXGc</a> (side-by-side: same model, same task, with and without Forge guardrails)The paper (accepted to ACM CAIS &#x27;26, presenting May 26-29 in San Jose) covers the peer-reviewed findings across 97 model&#x2F;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&#x27;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&#x27;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&#x27;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&#x27;t think anyone&#x27;s published this because standard benchmarks don&#x27;t control for serving backend.Another surprise: there&#x27;s no distinction in current LLM tool-calling between &quot;the tool ran successfully and returned data&quot; and &quot;the tool ran successfully but found nothing.&quot; Both return a value, the orchestrator marks the step complete, and bad data cascades downstream. It&#x27;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&#x27;t tested. If you get interesting results I&#x27;ll add them to the dashboard.- Try the proxy server mode - point any OpenAI-compatible client at Forge and it handles guardrails transparently. It&#x27;s the newest model and I&#x27;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&#x27;t sweep it - including Opus 4.6. Curious if anyone finds scenarios that expose gaps I haven&#x27;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 &#x27;26 - presenting May 26-29.Repo: <a href="https:&#x2F;&#x2F;github.com&#x2F;antoinezambelli&#x2F;forge" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;antoinezambelli&#x2F;forge</a>Paper: <a href="https:&#x2F;&#x2F;www.caisconf.org&#x2F;program&#x2F;2026&#x2F;demos&#x2F;forge-agentic-reliability&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.caisconf.org&#x2F;program&#x2F;2026&#x2F;demos&#x2F;forge-agentic-re...</a> <a href="https:&#x2F;&#x2F;github.com&#x2F;antoinezambelli&#x2F;forge&#x2F;blob&#x2F;main&#x2F;docs&#x2F;forge_ieee_preprint.pdf" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;antoinezambelli&#x2F;forge&#x2F;blob&#x2F;main&#x2F;docs&#x2F;forg...</a>Dashboard: <a href="https:&#x2F;&#x2F;github.com&#x2F;antoinezambelli&#x2F;forge&#x2F;docs&#x2F;results&#x2F;dashboard.html" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;antoinezambelli&#x2F;forge&#x2F;docs&#x2F;results&#x2F;dashbo...</a>

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

The Mercury logic programming system

Mistral AI acquires Emmi AI

Apple unveils new accessibility features

Minnesota becomes first state to ban prediction markets

Growing Neural Cellular Automata

The two oldest printing presses

Tool mapping 90 companies in the photonics and CPO supply chain

Copy Fail, Dirty Frag, and Fragnesia kernel vulnerabilities

HTML-in-Canvas Demos

Disney erased FiveThirtyEight

Hanoi’s humble beer glass and the memory of a nation

Dumb ways for an open source project to die

Gemini Omni

Architect of the UK Online Safety Act Calls for Its Complete Repeal

AI, "Humanity", and Dr. Manhattan Syndrome: A Communications Intervention

The Setup:<a href="https:&#x2F;&#x2F;i.imgur.com&#x2F;o0hgybh.jpeg" rel="nofollow">https:&#x2F;&#x2F;i.imgur.com&#x2F;o0hgybh.jpeg</a><a href="https:&#x2F;&#x2F;i.imgur.com&#x2F;mcNiomp.jpeg" rel="nofollow">https:&#x2F;&#x2F;i.imgur.com&#x2F;mcNiomp.jpeg</a><a href="https:&#x2F;&#x2F;i.imgur.com&#x2F;vIjw6pc.jpeg" rel="nofollow">https:&#x2F;&#x2F;i.imgur.com&#x2F;vIjw6pc.jpeg</a><a href="https:&#x2F;&#x2F;i.imgur.com&#x2F;nzOwmSC.jpeg" rel="nofollow">https:&#x2F;&#x2F;i.imgur.com&#x2F;nzOwmSC.jpeg</a>

4:34:17 AM WEDNESDAY MAY 20, 2026

THE HACKER NEWS