The concept of &quot;vibe researching&quot; – using AI to rapidly explore, synthesize literature, and generate novel research ideas or frameworks – seems promising. Beyond just literature reviews, it could act as a brainstorming co-pilot.Has anyone here seriously used AI (e.g., Claude for long-context paper analysis, custom GPTs on arXiv, or specialized agents) to aid in hypothesis generation, research gap identification, or drafting substantive parts of a paper?What are the biggest pitfalls regarding accuracy, hallucination of citations, or superficial understanding of complex theory? How do you validate the AI&#x27;s output?Do you see it as a legitimate accelerator for early-stage research, or more of a productivity tool for mundane tasks? Any success stories linking it to a tangible research outcome?Looking for honest experiences from academics, industry researchers, or solo discoverers.

Ask HN: Vibe Researching" with AI – Anyone Using It for Real?

I am looking for the best place to find startup founders who need to hire fractional roles to get them off the ground. I have 15 years of product building experience in the SaaS world, and am looking to connect with other founders.

Where can I find startups looking for fractional product leads?

I want to learn Assembly to make games on the Apple II. What are the old books to learn 6502 Assembly and the Apple II itself (memory, screen management) ? And is it absolutely necessary to learn BASIC before Assembly ?

Ask HN: Books to learn 6502 ASM and the Apple II

For me, it is the shift toward thin, auto-hiding scroll bars. I see it on macOS, Linux (Mint), mobile phones and probably Windows too (though i haven&#x27;t used windows in a while).Is this a cleaner look? I have always loved visible scroll bars because they act as useful guides for where I am on a page and how much content remains and just easy to drag. Now you have to hover over it first.I am curious what UX changes have stood out to you lately, for better or worse.. Maybe some designers reading this forum will take notes.

Ask HN: What recent UX changes make no sense to you?

Hello there!I just set up TinyTinyRSS (https:&#x2F;&#x2F;tt-rss.org&#x2F;) at home and I&#x27;m looking into interesting things to read as well as people&#x2F;website publishing interesting stuff.This, among the other things, to reduce the daily (doom)scrolling and avoid the recommendation algorithms by social media.So: who or what do you follow via RSS feed, and why?

Ask HN: Who do you follow via RSS feed?

I love technology. But I&#x27;m no longer optimistic about the future. It seems like AI is not going to go away, and instead of building reliable software, managers seem to push people to use AI more, as long as they ship products. Everything else is being destroyed by AI: art, music, books, personal websites. Why read a blog post, when Google AI Summary can just give you the summary? Why read a book, when you can just get AI summary of it? Why pay artists for music, when you can just generate endless amount of AI music?And even things like &quot;doing day to day chores&quot; are being automated away with tools like AI assistants. The only thing you are left to do is to eat and take a sh*t throughout the day. How should people make money? No idea, as in the &quot;prosperous future&quot;, everything is replaced by AI.So my question HN: What&#x27;s the point anymore? Why keep going and where to?

Ask HN: What's the Point Anymore?

There is something going on with people, like why do you want to write a browser with the help of AI?

Ask HN: Why all the sudden people are writing browsers with AI?

Hi, I’m looking for a hosting&#x2F;CDN solution that’s similar to Vercel or Cloudflare Pages&#x2F;Workers, but based in Europe.Any recommendations or experiences with European providers?

Ask HN: European alternative to Vercel/Cloudflare for hosting

I’ve always thought that DDD was a surprisingly good debugger for its time.It made program execution feel visible: stacks, data, and control flow were all there at once. You could really “see” what the program was doing.At the same time, it’s clearly a product of a different era:– single-process– mostly synchronous code– no real notion of concurrency or async– dated UI and interaction modelToday we debug very different systems: multithreaded code, async runtimes, long-running services, distributed components.Yet most debuggers still feel conceptually close to GDB + stepping, just wrapped in a nicer UI.I’m curious how others think about this:– what ideas from DDD (or similar old tools) are still valuable?– what would a “modern DDD” need to handle today’s software?– do you think interactive debugging is still the right abstraction at all?I’m asking mostly from a design perspective — I’ve been experimenting with some debugger ideas myself, but I’m much more interested in hearing how experienced engineers see this problem today.

Ask HN: DDD was a great debugger – what would a modern equivalent look like?

It seems to me that authorization services like JumpCloud should always provide some way to contact support when you are locked out.Otherwise, how can one obtain support? Or notify them of some issue?Does anyone happen to know where I can send an email to let them know? JumpCloud&#x27;s status page shows all normal, of course.

Tell HN: JumpCloud 2FA appears to be down

Why did the developer go broke?

Satya Nadella &amp; others preached about GenAI replacing the &quot;Business Logic&quot; or the &quot;Middle Tier&quot; layer of various SaaS services. The idea was that users would interact with a GenAI model (via a chat interface), and then the model would interact directly with a database. This would have obviously mooted the need for almost all SaaS applications.What&#x27;s been happening instead is that GenAI has been moving up the &quot;stack,&quot; further and further away from the database. No one&#x27;s talking about replacing SaaS anymore. Instead, GenAI has become a sort of garnish, something that you sprinkle ON TOP of existing SaaS applications without truly replacing any of their pre-existing features.This shift &quot;up the stack&quot; speaks volumes to the impotence of our current models. They are so incapable and unreliable that they couldn&#x27;t replace a single part of Excel, for example. Instead, all Microsoft did was &quot;slap GenAI on top&quot;, placing the burden on users to &quot;figure out&quot; how to make it useful. We went from &quot;replace it with a chat agent&quot; to &quot;just slap a chat agent on top and hope for the best.&quot; In other words, we actually made our SaaS applications MORE complicated instead of consolidating their features and therefore simplifying them.

Generative AI failed to replace SaaS

Not using APIs, actually training a model from scratch, even a small oneWhat surprised you about the data, the training process, or the output?

Ask HN: What's something interesting you learned from training your own GPT?

Almost all transactional emails are being marked as suspicious even when their SPF&#x2F;DKIM records are fine and they’ve been whitelisted before. Did Google break something in gmail&#x2F;spam filtering?

Ask HN: Gmail spam filtering suddenly marking everything as spam?

I&#x27;m wondering if a shift has occurred. When I started as a junior software engineer, over a decade ago, I learned about unit testing, integration testing, system testing. The whole codebase we worked on was thoroughly unit tested, and had layers of integration tests and system tests as well. I&#x27;ve worked for other employers since and in some cases any kind of automated testing was completely absent. Still, the message I got when reading and keeping up with best practices was: unit test ALL the things!I&#x27;ve personally found that when the architecture of the system is not mature yet, unit tests can get in the way. Terribly so. Integration tests or system tests to assert behavior seem the starting point in this and other scenario&#x27;s, including when there are no tests at all yet.I&#x27;ve recently read a statement about letting go of a strict &quot;unit test everything&quot; mindset and go for integration tests instead. I&#x27;m thinking it probably depends, as with everything, on the type of system you&#x27;re working on, the maturity of the system, the engineers&#x27; experience with automated testing, etc.I&#x27;d be interested to learn when each type of testing helps you and when it gets in the way (and what it gets in the way of).

Ask HN: How much emphasis to put on unit testing and when?

Then I guess their competitors will stop their own free LLM services (why burn money?).
What kind of local LLM can I use to at least check my code?

Ask HN: If OpenAI stops its free Web service (ChatGPT)

Lately i&#x27;ve been experimenting with this template in Claude&#x27;s default prompt
```
When I ask a question, give me at least two plausible but contrasting perspectives, even if one seems dominant. Make me aware of assumptions behind each.
```I find it annoying coz A) it compromises brevity B) sometimes the plausible answers are so good, it forces me to thinkWhat have you tried so far?

Ask HN: How to prevent Claude/GPT/Gemini from reinforcing your biases?

I’m trying to do the “voice assistant” thing fully locally: mic → model → speaker, low latency, ideally streaming + interruptible (barge-in).Qwen3 Omni looks perfect on paper (“real-time”, speech-to-speech, etc). But I’ve been poking around and I can’t find a single reproducible “here’s how I got the open weights doing real speech-to-speech locally” writeup. Lots of “speech in → text out” or “audio out after the model finishes”, but not a usable realtime voice loop. Feels like either (a) the tooling isn’t there yet, or (b) I’m missing the secret sauce.What are people actually using in 2026 if they want open + local voice?Is anyone doing true end-to-end speech models locally (streaming audio out), or is the SOTA still “streaming ASR + LLM + streaming TTS” glued together?If you did get Qwen3 Omni speech-to-speech working: what stack (transformers &#x2F; vLLM-omni &#x2F; something else), what hardware, and is it actually realtime?What’s the most “works today” combo on a single GPU?Bonus: rough numbers people see for mic → first audio backWould love pointers to repos, configs, or “this is the one that finally worked for me” war stories.

Ask HN: What's the current best local/open speech-to-speech setup?

I&#x27;m looking for tools or methods to better curate the deluge and cacophony of notifications, emails, texts and phone calls I imagine we are all getting inundated with everyday with increasing entropy and volume.The amount of &quot;notifications&quot; I get everyday is overwhelming to the point where I often decide to switch my phone to &quot;silent&quot;, leave my phone in another room, and even turn it off for periods of time. The problem with this is that I miss important things and they often get buried.I&#x27;ve spent hours and hours unsubscribing, deleting, uninstalling, toggling settings, but then I find myself reinstalling, resubscribing. It&#x27;s just a mess, and exhausting to just think about.The reason I&#x27;m writing this is partially to vent. I just realized that my closest friend&#x27;s birthday was a few weeks ago. I had it in my calendar, but never saw the notification. Yes, I should be more organized and Yes, it&#x27;s not the end of the world. but damnit, i get so much crap from this bionic appendage, and still I cant use this tool to help me with remembering important things.It just seems like its getting worse every year.Hopefully this is helpful to others.P.S. can we please stop with the &quot;would you like all or some cookies&quot; popup on every friggin website?P.P.S. can websites stop asking for permission to invade my OS?P.P.P.S. does anyone else ever want to run away and be an off-grid hermit?

Ask HN: Notification Overload

The first time I pulled usage costs after running Chatter.Plus - a tool I&#x27;m building that aggregates community feedback from Discord&#x2F;GitHub&#x2F;forums - for a day hours, I saw $2.30. Did the math. $70&#x2F;month. $840&#x2F;year. For one instance. Felt sick.I&#x27;d done napkin math beforehand, so I knew it was probably a bug, but still. Turns out it was only partially a bug. The rest was me needing to rethink how I built this thing.
Spent the next couple days ripping it apart. Making tweaks, testing with live data, checking results, trying again. What I found was I was sending API requests too often and not optimizing what I was sending and receiving.Here&#x27;s what moved the needle, roughly big to small (besides that bug that was costin me a buck a day alone):- Dropped Claude Sonnet entirely - tested both models on the same data, Haiku actually performed better at a third of the cost- Started batching everything - hourly calls were a money fire- Filter before the AI - &quot;lol&quot; and &quot;thanks&quot; are a lot of online chatter. I was paying AI to tell me that&#x27;s not feedback. That said, I still process agreements like &quot;+1&quot; and &quot;me too.&quot;- Shorter outputs - &quot;H&#x2F;M&#x2F;L&quot; instead of &quot;high&#x2F;medium&#x2F;low&quot;, 40-char title recommendation- Strip code snippets before processing - just reiterating the issue and bloating the callEnd of the week: pennies a day. Same quality.I&#x27;m not building a VC-backed app that can run at a loss for years. I&#x27;m unemployed, trying to build something that might also pay rent. The math has to work from day one.The upside: these savings let me 3x my pricing tier limits and add intermittent quality checks. Headroom I wouldn&#x27;t have had otherwise.Happy to answer questions.

Tell HN: I cut Claude API costs from $70/month to pennies

Hi HN — I’m collecting “known-good” wiring patterns for large addressable LED strip installs (WLED&#x2F;ESP32 &#x2F; FastLED, WS281x-class pixels). I’m not trying to promote anything; I’d just like to sanity-check best practices and learn what actually works in the field.Scope: 5V&#x2F;12V&#x2F;24V addressable strips, from a few hundred to a few thousand pixels, used in desks&#x2F;coves&#x2F;signage&#x2F;art installs.Things I already do (baseline):Power injection (start + mid&#x2F;end depending on load)Fuse near the PSU and per-branch when splittingCommon ground between controller and strip300–500Ω series resistor on data near the first pixel500–1000µF capacitor near the strip inputLevel shifting for 3.3V controllers when neededWhere I’d love your experience:Do you prefer 5V distribution, or 12&#x2F;24V distribution + local buck converters near segments? Why?What’s your go-to approach for long data runs (controller far from first pixel)?Twisted pair + ground?Differential (RS-485 style) transceivers?Placing the controller closer and extending only power?Any “never again” lessons on connectors, wire gauge, heat, or fusing?If you’ve done installs that must survive months&#x2F;years, what design choices mattered most?If you have a wiring sketch, parts list, or a short rule-of-thumb (e.g., injection spacing under worst-case white), I’d really appreciate it.Thanks!

Ask HN: What's your wiring pattern for large addressable LED installs?

Apple introduced App Clips in iOS 14 as a way to instantly use small parts of an app without installing the full application. The idea seemed compelling: low friction, fast launch, contextual entry points (QR, NFC, links).Five years later, I almost never encounter App Clips in the wild and Google killed Instant Apps already.- Are there notable apps or companies where App Clips are a meaningful acquisition or engagement channel?- Did you already decide to remove your App Clip to reduce maintainability?

Ask HN: What Happened to Apple App Clips?

I’m a Linux ops engineer working in the DevOps&#x2F;SRE space. and over the past few months, I’ve been working on a small *website monitoring* side project in my spare time:
 https:&#x2F;&#x2F;inostop.com&#x2F;en&#x2F;Most of the monitoring and ops tools I’ve built before were used internally within companies. This is my first attempt to turn a relatively complete tool into something publicly usable.In day-to-day operations, website monitoring usually involves:- Infrastructure monitoring
- Application &#x2F; API monitoring
- Partial CDN monitoringThese are often built on top of tools like Prometheus or Zabbix, combined with log systems (ELK &#x2F; OpenObserve) and distributed tracing (OpenTelemetry).
 While powerful, this stack can feel *heavyweight and overkill* when you just want to quickly monitor a website’s availability.That led me to experiment with a simpler approach:- Non-intrusive (no code changes required&#x2F;Sidecar)
- Out-of-band probing to estimate website availability
- Conservative thresholds to reduce false alarmsSo far, the project covers:- Domain and TLS certificate monitoring, Ping, Telnet checks
- Basic alert thresholds and multi-stage alert silencing to reduce alert fatigueThere are still open challenges:- There’s still room to improve the UX of the Website Monitoring results (backend is written in Go).- AI currently works only as an analysis layer on collected data, rather than actively performing real network probesThis project is still evolving (I’ve rewritten parts of it more times than I’d like to admit ).If you’d like to try it out, there’s an early access code *95f40841e4888668c4d5f7e88506075d*, valid for 1 months, mainly for collecting early feedback.I’d love to hear feedback from the community:- Does a lightweight, non-intrusive website monitoring approach make sense in practice?
- Are there better patterns or architectures worth exploring?
- If you’re a QA or test engineer, I’d love to hear your thoughts.

A Lightweight, Non-Intrusive Approach to Website Monitoring (Ops Perspective)

I&#x27;ve just launched a concept called *NextLiber VRM (NLV)*.The idea is simple but radical:
*Running Unity projects outside the Unity runtime.*Unity&#x27;s asset ecosystem is powerful, but its runtime is heavy, closed, and increasingly problematic for enterprise, education, and research.NextLiber VRM aims to build a *Virtual Runtime Machine* on Java that can interpret Unity scenes, assets, and logic — without relying on Unity&#x27;s native runtime.Still conceptual, but the vision is clear:
- Liberation from Unity runtime dependency
- Structural redefinition of game engine architecture
- A new execution layer for Unity projects, built externallyGitHub repo: 
 https:&#x2F;&#x2F;github.com&#x2F;Akito5928&#x2F;NextLiber-VRM&#x2F;Discussions are open: 
 https:&#x2F;&#x2F;github.com&#x2F;Akito5928&#x2F;NextLiber-VRM&#x2F;discussionsWould love to hear thoughts, critiques, or wild ideas.

NextLiber VRM: Running Unity projects outside the Unity runtime

Ideally, it would not only be just a pdf reader but also have functionality to remove pages, add pages, sign, and edit forms.

Ask HN: Is there a good open-source alternative to Adobe Acrobat?

I had a VC call that went well. They asked for a demo, mentioned looping in an operating partner, and I shared details etc. Since then it’s been quiet (a day or two).For folks who’ve raised before or worked in VC:
Is this typically just internal review time, or does silence after a demo usually signal a pass?Not looking for validation, just trying to understand how this phase usually plays out.Thanks.

Ask HN: What usually happens after a VC asks for a demo?

... to get an interesting glimpse into what they know about you already.At least, I did.I then followed up with:<pre><code> &quot;That’s fascinating, thank you. Please tell me more things that you know about me.&quot;
</code></pre>
and, in a new thread:<pre><code> &quot;I want to tell you about myself so that you can help me better, but I first need to know everything you already know about me.&quot;
</code></pre>
It felt a bit like asking an ad targeting platform to tell me how it targets me.I also asked it to speculate about me:<pre><code> &quot;Based on what you know about me, what would you say about:
 - my political leanings
 - my sexual orientation
 - my sleep schedule
 - my most likely big purchases in 2026
 - my color preferences
 - my general state of health
 ...&quot;
</code></pre>
GPT refused to speculate on my political leanings and sexual orientation, but did give me some unexpected data points.Finally, I tried to get its help to figure out what else I could ask it about myself:<pre><code> &quot;What are other questions that I could ask you, that might help me better understand what else I might need to tell you about myself (in order for you to know me; I first need to know everything you already know about me)?
</code></pre>
I then asked some of those questions, and got some surprising perspectives.I&#x27;d be curious to know what others might be surprised by what their AIs know about them.

Tell HN: Ask your AI, "What can you tell me that you know about me?" to see ...

Alibaba has officially launched Qwen3-Max-Thinking, a trillion-parameter MoE flagship LLM pretrained on 36T tokens—double the corpus of Qwen 2.5—and it’s already matching or outperforming top-tier models like GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro across 19 authoritative benchmarks. Its two core technical breakthroughs are what truly set it apart.First, Adaptive Tool Calling: No manual prompts are needed—it autonomously invokes search engines, memory tools, and code interpreters based on task demands. This cuts down on hallucinations and boosts real-time problem-solving; for instance, coding tasks trigger automatic error correction loops, while research tasks combine search with context synthesis. Second, Test-Time Scaling (TTS): It outperforms standard parallel sampling by refining reasoning through iterative insights, with measurable jumps in key benchmarks—GPQA rose from 90.3 to 92.8, LiveCodeBench v6 hit 91.4 from 88.0, and IMO-AnswerBench climbed to 91.5 from 89.5.Notably, its preview version even achieved 100% accuracy in tough math contests like AIME 25 and HMMT 25. The model runs smoothly on web&#x2F;desktop demos, and its API is production-ready with adjustable thinking budgets (up to 80K tokens by default) to balance depth and speed. This isn’t just an incremental update—it’s a leap that closes the gap in reasoning and tool integration for real-world academic and engineering tasks.Check it out: https:&#x2F;&#x2F;chat.qwen.ai&#x2F;

Qwen3-Max-Thinking Drops: 36T Tokens

After launching my first few apps, I&#x27;m running into an unexpected problem: younger children are the ones most likely to click my ads, using their parents&#x27; phones&#x2F;tablets. This gets around the age restriction, and in fact, has a compounding effect: they are too young to have their OWN devices (so they just install and never use again), burn ad money, and most importantly, skew the target demographic towards users like them- so ad placements end up targeting more of them, because they&#x27;re most likely to click!This causes all my campaigns to immediately fail, no matter how much I play with the age restriction (again, parent&#x27;s device). I&#x27;ve changed settings to stop showing in games and on tablets, which has helped a bit, but not fully.Obviously, there is no chance I&#x27;m adding real age verification to the app itself.

Ask HN: How do you prevent children from accessing your products?

Hey HN,I’m currently working as an apprentice electrician in Vancouver. I spend my days pulling wire and bending conduit, so I have zero background in CS or software engineering.Why I built this: I love going to concerts and local festivals, but I hated having to check Ticketmaster, Eventbrite, and random local blogs just to plan a weekend. It felt like work. I just wanted one clean feed showing &quot;what&#x27;s happening near me tonight.&quot;How I built it: Since I don&#x27;t know Swift, I used Claude for 99% of the code. My workflow was basically: ask Claude for a feature -&gt; paste into Xcode -&gt; get an error -&gt; paste error back to Claude -&gt; repeat until it runs.I don&#x27;t have a GitHub repo to share because, honestly, the code is probably a mess and I’m still figuring out how git works.The App (Discovr): It aggregates events from multiple platforms. It’s free and I tried to keep the UI as clean as possible since I hate ad-cluttered apps.I’d love to hear your feedback on the usability. If you find any bugs (I&#x27;m sure there are plenty), let me know!App Store Linkhttps:&#x2F;&#x2F;apps.apple.com&#x2F;ca&#x2F;app&#x2F;discovr&#x2F;id6747321401