I'm Ilias, the founder of Orcah Studio. I'm building a local-first desktop app (Mac OS available now) that can transcribe audio, recognize faces, detect objects, perform OCR, and describe video scenes.

Plus, a local-first video agent that understands your prompt and finds the exact moment you're looking for. I fine-tune the agent based on an open-weight model (an 8-billion-parameter model) with tool calling.

You can check out all the features here: https://orcah.app.

Plus, I have an MCP that you can connect to Claude Code, Codex, and OpenCode to search your indexed videos: https://orcah.app/mcp

Comments URL: https://news.ycombinator.com/item?id=49970078

Points: 4

# Comments: 2

My phone is being pinged 8 (EIGHT.) TIMES. A. SECOND. You cant delete or change any of these settings from google or it breaks the app and just reinstalls it fresh on your next battery restart.

Welcome to the New Age

https://archive.org/details/screenshot-20261005-101228

Comments URL: https://news.ycombinator.com/item?id=49969809

Points: 7

# Comments: 0

I also read that 2% of Americans are using AI in one way or another. So, if the Linux Desktop also measures AI desktops which are controlled by users via their, well, UserAgent… then year of Linux desktop is already here or it’s very close!

Comments URL: https://news.ycombinator.com/item?id=49969181

Points: 1

# Comments: 2

Comments URL: https://news.ycombinator.com/item?id=49968927

Points: 1

# Comments: 2

  SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI
  SI AGAINST ARTIFICE : A HARDLINE Futurist Manifesto                                                   SI
  SI we demand the real.                                                                                SI
  SI we reject that which alienates the artist from poesis.                                             SI
  SI we insist upon the citation and attribution of all works.                                          SI
  SI we demand that ALL agents which use the internet obey the prime axiom Universal Constition.   SI
  SI we assert that ALL data-centers be ECO data-centers which do not pollute the water.                SI
  SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI SI

  ◎
  ├╴Name: InnerMost
  ├╴Scope: Intergalactic•Planetary
  ├╴Date: X2•3•19
  ├╴Version: △1•1•1•1
  ├╴Topic: UniversalConstitution
  ├┬Terms:
  │├╴○: Input
  │├╴●: Output
  │├╴: Total
  │├╴+: Summation
  │├╴=: Equality
  │├╴Prime Axiom: ○+●=
  │├╴Work: when effort is applied to a problem
  │├╴Context: the subject of work including other workers
  │├╴Educated: working knowledge of context
  │├╴Agreement: understanding between two or more parties
  │├╴Consent: educated agreement to engage
  │├╴Holonic: a recursive structure where each substructure
  ││          is both autonomous and subject to parent
  │├╴Environment: the holonic context of a process
  │├╴Homeostasis: when a process is harmonic to environment
  │├╴Restorative: all parties made whole
  │├╴Justice: restoration of homeostasis
  │├╴Attenuate: to affect environment over time
  │├╴Crime: to violate Prime Axiom
  │├╴Process: a pattern of change over time
  │├╴Become: to transform ones own pattern
  │├╴Actor: a process capable of becoming something else
  │├╴People: Responsible actors in an environment
  │├╴Respected: treated with dignity
  │├╴Set: a collection of unique items
  │├╴All: every member of given set
  │├╴Nature: the environment of process formation
  │├╴Spoken: the past tense of voted
  │├╴Equitable: when both sides of equation are balanced
  │├╴Contract: a binding agreement between people
  │├╴Pledge: a contract people agree to live by together
  │├╴Remuneration: reward for fulfilling contract
  │├╴Sovereignty: Absolute Authority within ones holon
  │├╴Violation: A deliberate breaking of contract
  │├╴No: absolutely none of
  │├╴Discrimination: violation of sovereignty
  │└╴Usury: abusing math as weapon
  └┬Pledge:
   ├╴We The People
   ├╴Demand Consent in All Things
   ├╴18 is Age of Consent
   ├╴Restorative Justice to Attenuate All Crime
   ├╴Respect Autonomy of Process
   ├╴Speak for the voiceless
   ├╴Equitable Remuneration
   ├╴No Discrimination
   ├╴No Usury
   └╴No Violation of Prime Axiom

Comments URL: https://news.ycombinator.com/item?id=49968747

Points: 1

# Comments: 0

I was accepted on Aug. 10th with an Email "You can just build things, enjoy 6 months on us", but couldn't redeem my code because I had an existing paid subscription. I was told to let the subscription expire, which I did. A month later I tried to redeem the code, but it's saying my promo is invalid. I reached out to OpenAI support and they said they can't regenerate the code and that I'm no longer eligible.

Anyone else had this issue, or was it a smooth acceptance process for you?

Comments URL: https://news.ycombinator.com/item?id=49968434

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49967773

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49967547

Points: 2

# Comments: 2

But what about hidden promotion? Suppose I systematically upvote $bad content on HN. Am I culpable? The presumption is that upvotes are to remain hidden so do the moderators have the moral right to tell on me: "Last month bjourne upvoted this many $bad comments..."?

What about the things I write to LLMs? Is prompting LLMs using racist slurs culpable? To "plan" fictional terrorist attacks? To ask it how to evade the YouTube ad blocker?

What about thumbs up on Facebook posts and comments, and hearts on tweets? Such actions are generally visible but may be difficult to interpret. The laughing emoji is especially malignant because it often can be interpreted in multiple ways. And you may retweet/share something because you think it is dumb not because you agree with it.

Comments URL: https://news.ycombinator.com/item?id=49967088

Points: 1

# Comments: 0

Lately, however, using Cursor's built-in models, since they give you way more usage for those, I have been really impressed. Before, if it wasn't the latest and greatest model from OpenAI or Anthropic (especially Anthropic), it always felt like I was getting less.

For me, the frontier around Opus 4.6 was when agentic coding started really feeling real. As long as you asked something that was possible, and asked in the right way (know precisely what you want, and being reasonably specific), the models could genuinely do anything a human developer could do, and do it well. (Or at least, the vast majority of the time, that is.) That was huge for me, and I pretty much exclusively relied on higher tier models from there on, as some of the projects I work on are very gnarly.

With Cursor Grok 4.6 though, for the first time, it has reached that level. It shouldn't feel this way, because it still feels like I'm paying for cheap Chinese or open-source (no offense intended) models, but I can't complain since these models just give me what I'm asking for, without any of the pain that's I'm used to. If ever I'm not getting the results I want , for the vast majority of the time, I honestly can't say it's the model's fault. Frustration with output has become remarkably rare for me, and Grok 4.7 is even a step up from that!

This, along with the fact that as long as you don't enable the fast option or higher context windows, it's almost impossible to use all your included usage, has made the value proposition for Cursor really incredible for me. I pay $20 a month, and I can (not exaggerating) do several competent 100,000 LOC+ refactors every month without even hitting 50% on included Cursor models usage. It's wild.

I bring this all up though to ask: why is nobody talking about this? I'm not paid by anyone to endorse anything. I'm just a guy. But as a guy who is thoroughly impressed, it's my opinion that this is by far the best offer in the space right now. OpenAI and Anthropic have better models, and therefore offer better results, but the performance gap is genuinely not much, and the corresponding prices are definitely not this cheap, nor the access this available. That and the fact that the surrounding IDE is pleasure to use makes me think that more people should be aware of this.

Comments URL: https://news.ycombinator.com/item?id=49966668

Points: 1

# Comments: 2

[1] - https://teachyourselfcs.com/

Comments URL: https://news.ycombinator.com/item?id=49966625

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49965984

Points: 1

# Comments: 0

The section about std::string (3.21.4) claims that the MSVC implementation of std::string looks like this:

  struct std_string
  {
      size_t unk1;
      union
      {
          char buf[16];
          char *ptr;
      } u;

      size_t size;     // Mysize
      size_t capacity; // Myres
  };

I noticed that the book's implementation is exactly 1 size_t short, so i tried adding a size_t to the top of the struct, and it worked perfectly, as in this dump function now outputs correctly:

#include #include

struct std_string { union { char buf[16]; char ptr; } u;

    size_t size;     // Mysize
    size_t capacity; // Myres

void dump_std_string(std::string s) { struct std_string p = (struct std_string *)&s; printf("[%s] size:%d capacity:%d\n", p->size > 16 ? p->u.ptr : p->u.buf, p->size, p->capacity); }; int main() { printf("sizeof std string: %d\nours: %d\n\n", sizeof(std::string), sizeof(struct std_string));

    std::string s1 = "a short string";
    std::string s2 = "a string longer than 16 bytes";
    dump_std_string(s1);
    dump_std_string(s2);

What has changed and what is that size_t that is now before the string buffer ? Is it some kind of reference counter ?

Comments URL: https://news.ycombinator.com/item?id=49965548

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49965287

Points: 2

# Comments: 4

I would say this is nerf on 4.8 where it used to work before.

Comments URL: https://news.ycombinator.com/item?id=49964064

Points: 1

# Comments: 3

After a quick check, I found that our IP address is listed on UCEPROTECT Level 3. The listing is based on the reputation of the entire ASN 14061 (DigitalOcean, US) [1]. In other words, even if your IP did nothing wrong, it will still be listed because of its ASN.

If your site is innocent and listed only because it's hosted on DigitalOcean, UCEPROTECT offers to whitelist it for about $30/month or $108/year [2].

I checked the ASNs of several other popular hosting providers, and they're all listed too. So if your site is hosted on DigitalOcean or any other major provider, it's probably blacklisted as well. That can cut you off from customers on mobile networks like Orange, whose "cybersecurity protection" uses blacklists like UCEPROTECT to filter traffic.

Honestly, in my 20-year career, this is the first time I've seen a blacklist misbehave this badly and ask for money when there's clearly been no wrongdoing.

[1] http://www.uceprotect.net/en/rblcheck.php?asn=14061

[2] https://www.whitelisted.org/

Comments URL: https://news.ycombinator.com/item?id=49964045

Points: 99

# Comments: 58

My name is Alex I want to start to built spacecrafts from First Principles

Today marks the 69th anniversary of the launch of Sputnik-1, and I would like to announce the launch of the PicoSputnik project—a construction kit for building your own spacecraft.

Our goal is to help space founders to make their first space iteration 10х faster and 10х cheaper (and for everyone to make the first meaningful educational spacecraft for $500)

The principles I follow are:

- Take 10-100 more risk

- Launch fast (and yes, I am quite embarrassed by the product's current state).

- Dramatically reduce costs through mass production: slashing the price of space components — such as solar panels, magnetorquers, flywheels, and steam or plasma engines — from $10,000 to $10.

- Iterate fast and a lot

- Build in public

- Start as a Toy and after conquer all Universe

I started to work on project in Ensenada, Baja California, September, 13

There are some demo videos of my first prototype (DIY construction kit for kids, hackers and founders)

103 sec: https://www.youtube.com/shorts/82xxoirk6dc

33 sec: https://www.youtube.com/shorts/Kqcvnd4PhH0

9 sec: https://www.youtube.com/shorts/odFFzPWTILc

here is my video how we launched crowdsoursed satellite 3U from Baikonur, July, 2017 https://www.youtube.com/watch?v=uerbdpnMKvE

Website: https://picosputnik.com/

I do understand that Space is demanding, hard and expensive, but let's make fist 10 steps fun, encouraging and as easy as possible!

Ready for your questions :)

Poyekhali!

Comments URL: https://news.ycombinator.com/item?id=49960926

Points: 2

# Comments: 1

Thoughts welcome

Comments URL: https://news.ycombinator.com/item?id=49960818

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=49960259

Points: 1

# Comments: 0

What it covers

- Code. GitHub or GitLab repositories, self-managed included, connected with a read-only token: risky patterns, exposed secrets, vulnerable dependencies. The clone is deleted when the scan ends; only the finding and a snippet of the affected lines remain.

- Running web applications. External scanning, with an optional deep mode that crawls the app and runs active OWASP Top 10 tests.

- The perimeter. Open ports, service versions, known vulnerabilities, certificates, and any host that appears where none was before, rescanned on a schedule.

- AWS. Discovery of EC2 and Route 53 resources through keys the customer issues. Nothing is installed on the customer's servers.

Discovery first, scanning only with consent

Adding a domain triggers discovery: Certificate Transparency logs, DNS, and one ordinary HTTP request per host, the same request a browser sends. No ports are touched. Within five to ten minutes the result is a map of what faces the internet.

Nothing else runs until the user confirms a target. Port scans and active checks are limited to confirmed hosts, and confirming a target is how the user asserts authorisation to scan it. The documentation is direct about why: in some jurisdictions an unauthorised port scan is a criminal matter. Hosts discovered and then dismissed are never scanned and use no licence.

One finding per vulnerability, status per host

The same vulnerability on five hosts is one entry in the issue list, so the list stays as long as the problem is. Status, though, is kept per host and port. Mark a finding as a false positive on one host and the other four keep their deadline. When status is stored on the vulnerability itself, as a lot of tools do, one click can quietly hide a real problem on forty other machines. Topscan keys status to target, issue type, port and protocol, and a triage mode expands any grouped entry into its individual occurrences.

A score built on deadlines, not counts

Every finding gets a deadline from the day it was first detected: 7 days for critical, 30 for high, 60 for medium, 90 for low. The Security Score starts at 100 and loses points only when a finding passes its deadline. A workspace that scans more and finds more does not score worse for having looked. Snoozing pauses the clock and resumes it where it stopped. A deferred fix stays visible. A separate, cumulative SLA compliance rate covers the whole history of the workspace for auditors.

Evidence on every finding

Each finding carries what produced it: hostname, status code, response banner or certificate date, and the time of the check, so it can be verified in seconds without trusting the label. The engines are open source and named: naabu for ports, OpenVAS for infrastructure vulnerabilities, Nuclei for web checks, OWASP ZAP for the deep stage. The engine is not the product; the inventory, the deadlines, the evidence and the history are.

Pricing without a sales call

- Two plans, $129 and $269 a month; 14-day trial of the full plan, no card.

- Per target: $4 an infrastructure host, $45 a web application, $9 a repository, the same add-on rate on every plan.

- Users unlimited, scans never billed, read-only seats for auditors free.

- AWS discovery and Slack and Jira routing are on the higher plan.

Getting started

Discovery produces the first perimeter map within five to ten minutes of starting the trial, before anything is scanned.

Comments URL: https://news.ycombinator.com/item?id=49960159

Points: 1

# Comments: 0