🗣️ Post 1: Teen researcher: AI bug denied, later fixed without credit

As posted by: Anh_khoa  |  🔥 Points: 4

🔗 https://news.ycombinator.com/item?id=45194116

💬 Summary

I’m 14, based in Vietnam, and in July I discovered a vulnerability that exposed the system prompt of a major AI model.

I responsibly reported it via the official bug bounty program. Response: “Out of scope, just an AI issue.”

Weeks later, I checked again — the bug had been quietly patched. No acknowledgment, no credit.

If it’s “not a bug,” why fix it? And if it’s fixed, why dismiss the report?

Sharing here to hear thoughts from the security community.

🗣️ Post 2: Show HN: I built research agents because GPT cant accurately hilite txt in files

As posted by: ieuanking  |  🔥 Points: 2

🔗 https://app.ubik.studio/chat

💬 Summary

Hi HN,

We built Ubik because no available Chatbots, AI research platforms, or AI PDF analysis tools could highlight text down to the line level in the file.

Why is this important?

Researchers, academics, scientists, and students lack the tools to help produce usable research while elevating workflows without over-automating the steps between A-Z. They also need to trust the accuracy of output/analysis generated by the agents & models, and confidently gather quotes, sources, and citations for scholarly papers.

Key Features of Ubik:

• Search academic databases and convert open-access papers into AI-ready docs.
• Use enhanced OCR + PDF analysis tools.
• Make AI notes in PDFs that highlight text down to the line level.
• Reference files when prompting using the @ symbol (like Cursor), which minimizes hallucination and improves output.
• Generate high-level text with citations and use over 20+ models.

We are working on a custom EVAL suite and will publish our findings soon.

Please reach out with any comments, questions, or critiques.

🗣️ Post 3: AI's Path to Profits Is Being Driven by Consumers

As posted by: Bostonian  |  🔥 Points: 2

🔗 https://www.bloomberg.com/opinion/articles/2025-09-09/ai-s-path-to-profits-is-being-driven-by-consumers

💬 Summary

Markets naturally see through the lens of businesses. When tech stocks took a dive last month on concerns of an “AI winter,” investors were egged on by a study showing 95% of corporate AI pilot programs failed to deliver any gains in productivity or profit, making all this expensive AI start to look a little useless.

Consumers would beg to differ.

🗣️ Post 4: Show HN: AI image creation that works like chat threads

As posted by: Ronanxyz  |  🔥 Points: 1

🔗 https://hiflux.ai

💬 Summary

We've all been there: generate an image, hate the result, generate again. Or find a perfect photo but need one tiny change.

Current AI workflow is broken: • Generate image → switch to different tool → upload → configure parameters → single edit → download • Want another change? Start over. Different style? New tool. Iterate? Good luck.

Why can't image creation work like a conversation?

That's exactly what we built: https://hiflux.ai

The core insight: treat images like chat messages. Upload, generate, or quote any result to continue creating. No tool switching, no starting over.

Technical breakthrough was building "chat with images as threads":

*Multi-model architecture:* • FLUX.1 Kontext [pro]: unified generation + editing (Premium) • FLUX.1 Krea [dev]: opinionated text-to-image with distinctive aesthetics • FLUX.1 Kontext [dev]: context-aware editing that understands "change this, keep that" • Nano Banana (Gemini 2.5 Flash Image): multi-image fusion + character consistency

*Engineering challenges solved:* • Priority queue system: Free users ~20s, Plus ~6s, Premium ~2s • Pipeline optimization: parse → understand context → preserve → transform • 27-language support with semantic consistency • Zero-retention privacy (process in real-time, delete immediately)

Real workflow now: 1. "A sunset over mountains" → generates image 2. Click result → "make the sky more purple" 3. Quote previous → "add a lake in the foreground" 4. Continue iterating in one conversation thread

No account required. Completely free. Full commercial rights.

https://hiflux.ai

Most interesting technical details I'm happy to discuss: • How we maintain lighting consistency across edits • Multi-language prompt engineering • Building priority queues that feel instant • Why conversation threads work better than separate tools

What's your biggest frustration with current AI image tools?

🗣️ Post 5: Show HN: Narev – Rapid A/B tests for the LLM setup

As posted by: osquar  |  🔥 Points: 1

🔗 https://www.narev.ai/

💬 Summary

Hey! I want to get your feedback on our A/B testing platform for LLMs. Giving a 100 day Pro trial to anyone who signs up this week.

The premise: in 10 minutes compare a bunch of model/parameter/prompt combos to find the best point on the quality-latency-cost tradeoff.

Enter your data manually or sync from your tracing platform.

~300 models to select from.

Here is a 1 minute Loom: https://www.loom.com/share/39b7cd24166c4e1fafd5e7fbd12a9d4d?...

🎯 Final Takeaways

These discussions reveal how developers think about emerging AI trends, tool usage, and practical innovation. Take inspiration from these community insights to level up your own development or prompt workflows.