big_toast

"Mr. Musk or his affiliates may become aware, from time to time, of certain business opportunities ... and may direct such opportunities to other businesses in which they have invested."

"Under our charter, Mr. Musk and his affiliates are not restricted from owning assets or engaging in businesses that compete directly or indirectly with us"

Pg. 56

I think this part is interesting considering Tesla shareholders seem to have lost out on developing (x)AI (AGI?) internally.

Is there any risk to SpaceX that the Musk brand pulls the market cap too far ahead now?

It's not a risk factor I see in the prospectus but seems plausible to me.

Just like with the AI company vesting, I imagine a scenario where a company seeds its own competition by realizing the monetary gains before the work is done. Maybe there's precedent in the dot com bubble. Certainly people were able to sell before the dip a la Cuban and broadcast.com. But I'm thinking more more specifically inducing competitive space ventures.

The paper references some threat models they considered. They suggest someone might "possess paired information (both original and watermarked content)" and therefore be able to undo watermarking. Presumably it's fairly easy to get identity operations out of image APIs that would result in this situation. I'm not sure that addresses echelon's main concerns though.

I guess the SynthID-Image paper from Oct 2025[0] was an encoder-decoder for which they tested checking a flag or a 136 bit payload in 512x512 images and the watermark's robustness after various transformations.

Presumably the deployed version is meaningfully different.

[0]:https://arxiv.org/html/2510.09263v1

What information is included in the metadata or SynthID? How many bits can be encoded in a SynthID?

Can it be used to create something like nutritional labels for synthetic content? 10% synthetic text, 30 synthetic images.

Your reality was 15% synthetic today (75% mega corp, 25% open-weight neocloud).

Musk leaves the board in 2018 I think. And something happens in DX-754 where they've pivoted to AI in SpaceX around then too. I had a lot of trouble telling what "AI" meant in late 2017 at Tesla.

---

Sept 1, 2017 DX-669: Funding paused confirmation. Elon is still on the board for a while. DX-707 specifies the board as of Sept 26, 2017, and even suggests adding Shivon, Jared, Sam Teller.

Jan 31, 2018 DX-748: Elon is still discussing things with Greg. Elon: "The only paths I can think of are a major expansion of OpenAI and a major expansion of Tesla AI. Perhaps both simultaneously"

Feb 3, 2018 DX-754: Sam Teller says Elon "just suggested we use SpaceX email for AI stuff so switching over to that"

Feb 4, 2018 DX-755: Sam Teller and Shivon Zilis discuss disabling Openai

Feb 20, 2018 DX-770: Elon officially leaves board (first document I see specifying)

I love Robin's newsletters (and the books etc) and love "opinionated operator decisions make the internet go round".

Maybe I'm missing something but it seems like if i go gmail -> via a Robin Sloane Newsletter -> webpage, it seems generally useful to know Robin sent me. Like a backlink that's been liberated?

Is something different happening?

People need a mental bucket for 'stochastic software' for a while. Or hot mess, a fast food meal that you can expect to mostly be bad in some sense, but serves a purpose, and can be really good in that case.

Conflating the new style of agent-driven/vibe coded software with the old more predictable software leads to applying wrong heuristics/expectations.

People have a pretty good mental model of different types of meals they'll have in a year, and modulate their expectations by context. I think there's room for a new type of software that operates on different principles. Peter has mostly been clear what type of software he's developing. And if it ever converges to bug free, that's great, but I think some of his motivation is to figure out what this new software is. While not giving the users food poisoning.

Ya! So many posts with clear presentations of css/svg/canvas.

The Daily Sketch series or 'CSS Animation with offset-path' are equally fun.

The AI bothered me less, but I got a little frustrated with less than substantive comments on the front page.

Oddly I made an extension* to use the site more the way I wanted and now I find it a little easier to get a higher SNR past the front page and am enjoying that. I didn't really get past post rank 60 for two decades and now generally get much further.

*(It's basically vim-keys support for basically two functions. A function to "highlight" stories/comment threads I think will be promising and then hide function for the rest.)

I think one of the main goals of NASA QueSST program[0] is contributing to making commercial supersonic air travel happen.

And there's a lot of private companies contributing[1]. But makes me wonder where Boom/Hermeus/Astro Mechanica etc fit in.

[0]:https://www.nasa.gov/quesst-the-mission/

[1]:https://www.nasa.gov/wp-content/uploads/2019/05/companies-co...

This footage includes a landing this time. And I guess something akin to the cockpit HUD?

A post[0] from a couple days ago says they've reach ~0.95 Mach, ~43k feet, 10 flights, 11.7 hours of flight time with the longest flight at 1.7 hours.

[0]:https://www.lockheedmartin.com/en-us/news/features/2026/X-59...

I've been aware of Bbedit but never really used it. Are there things it does better than more recent editors or is it more of a devil you know? Like a WordPerfect situation.

The page about the earnings[0] is interesting. The Amazon share for kindle edition is higher than I expected.

[0]:https://vivianvoss.net/print/integrated-by-design

spacecowboy runs a 'For You' bsky feed powered by matching your likes to similar user likes.

He added a way to traverse posts via the like graph[0]. (The backing architecture was posted previously at [hn].)

[0]:https://bsky.app/profile/spacecowboy17.bsky.social/post/3mjr...

[hn]:https://news.ycombinator.com/item?id=47870458

"The MQ-25A Stingray is the U.S. Navy's first operational, carrier based unmanned aircraft system, designed primarily to serve as an aerial refueling tanker."

And youtube promo video[0] of the flight.

[0]:https://www.youtube.com/watch?v=b9WqSR_XYG0

Interesting history. From the wiki:

"A Senate bill was introduced in February 1947 to create the National Science Foundation (NSF) to replace the OSRD. This bill favored most of the features advocated by Bush, including the controversial administration by an autonomous scientific board. The bill passed the Senate and the House, but was pocket vetoed by Truman on August 6, on the grounds that the administrative officers were not properly responsible to either the president or Congress."

Also mentions the preceding organization OSRD (Office of Scientific Research & Development) and that they had tried to exempt it from conflict of interest regulations.

Having a "Recovery Mode"/"Safe Boot" flag to disable our configurations (or progressively enable) to see how claude code responds would be nice. Sometimes I get worried some old flag I set is breaking things. Maybe the flag already exists? I tried Claude doctor but it wasn't quite the solution.

For instance:

Is Haiku supposed to hit a warm system-prompt cache in a default Claude code setup?

I had `DISABLE_TELEMETRY=1` in my env and found the haiku requests would not hit a warm-cached system prompt. E.g. on first request just now w/ most recent version (v2.1.118, but happened on others):

w/ telemetry off - input_tokens:10 cache_read:0 cache_write:28897 out:249

w/ telemetry on - input_tokens:10 cache_read:24344 cache_write:7237 out:243

I used to think having so many users was leading to people hitting a lot of edge cases, 3 million users is 3 million different problems. Everyone can't be on the happy path. But then I started hitting weird edge cases and started thinking the permutations might not be under control.

Do you have additional information about this? I'm having trouble finding reporting or discussion on it.

Ya. I could imagine a browser extension performing some form of verification loop for simpler webpages. Maybe too niche.

I found a link that said they were discussing the testing schedule with media, but haven't found a timeline for when they might go supersonic.

First X-59 post on hn was 8 years ago!

"The X-59’s thin, tapered nose accounts for almost a third of its length, and will break up the shock waves that would ordinarily result in a supersonic aircraft causing a sonic boom." [1]

The video with the wheels up must be really satisfying for the people who worked on it. Beautiful.

[1]:https://www.nasa.gov/news-release/nasa-lockheed-martin-revea...

Ya, I don't understand the calculus. If it's important enough for a big website, why not promote the site for a week first. Too many applications seems more desirable than missing the right people who are 'heads down building' or just not terminally online.

That on top of Direct Hire Authority.

I can see NASA Force[1] is part of the US Tech Force[2] push and has been talked about for the last several months.

[1]:https://www.meritalk.com/articles/nasa-opm-kick-off-drive-fo...

[2]:https://meritalk.com/articles/opm-launches-us-tech-force-to-...

Why does the application window last four days?

Charitably they're moving fast, but without already having people in mind for the roles or having created the hiring pipeline, how do you reach a sufficiently large audience. Is there an explanation I'm missing? Was this announced a while ago?

Makes it feel like they already know who they want for the roles/preferential selection. On a longer or recurring timescale, seems like a cool way to reach out to potential hires.

No, I don't have any good ideas. Just hoping someone else does, or that I'm missing something.

I think it's in the hands of browser vendors.

The agent review a la socket.dev probably doesn't address all the gaps. I think you're already doing about as much as you reasonably can.

URL data sites are always very cool to me. The offline service worker part is great.

The analytics[1] is incredible. Thank you for sharing (and explaining)! I love this implementation.

I'm a little confused about the privacy mention. Maybe the fragment data isn't passed but that's not a particularly strong guarantee. The javascript still has access so privacy is just a promise as far as I can tell.

Am I misunderstanding something and is there a stronger mechanism in browsers preserving the fragment data's isolation? Or is there some way to prove a url is running a github repo without modification?

[1]:https://sdocs.dev/analytics

Page 38 has an interesting comparison saying a session of 8 "medium length" messages uses roughly 10Wh, the equivalent of ~2 phone charges. The data dashboard[1] I think is defining medium length as 1000 tokens in, 1000 tokens out. And the page before suggests GPT-5 mini (medium) used 3x the energy of GPT-4o (Aug) which I guess I assumed would be the opposite ratio.

---

I guess these estimates are based on calculations for gpt-4o (Aug) from Jegham et al[2] in November 2025.

[1]:https://app.powerbi.com/view?r=eyJrIjoiZjVmOTI0MmMtY2U2Mi00Z...

[2]:https://arxiv.org/pdf/2505.09598

"The deal was negotiated without the input of the trial team, catching even the lead counsel by surprise." Is trial team a technical term? How would you negotiate without the people people most familiar with the case.

I remember it being worse a while ago. But most of the time I can drag a binary into Settings->Privacy & Security->Full Disk Access or other things (Accessibility API). Maybe other issues come up.

I feel like it should still be much easier, but the general sandboxing model seems directionally functional. (My understanding is containerization isn't a silver bullet security-wise, still requires fiddling, and would be a resource hog ram-wise if not CPU?)

I wish I could pick a parent folder/file and get a box to control everything (network/disk/folders/peripherals/accessibility).

I feel like I can mostly use containers on macOS. Is there a different sense that people are using containers on *nix? Or are you referring to all the macOS specific software footguns?

I would like to be able to run arbitrary code with gradual/granular privilege escalation. (e.g iOS/android with more affordances and escape hatches. macOS is getting there, but it's been a pretty bumpy/potholed road). Right now if I download a random github repo, I'd put it in a docker container and give it ports/volumes/etc.