rikiyat.github.io talked to 6 outside services on a first visit, including 2 that measure visits or show ads.E10E8
An AI agent used the site and saw 1 of 2 claimed features working.E31E32E33E34E33E34
LLMs can natively read & write neural network weights (Sparks of Weight-Space Programming): The page is a research blog post that shows sample weight matrices, game-choice buttons and prompt/reasoning toggles. I clicked 'Breakout', 'Read full reasoning' and '▶ Play'. Only the Play/Pause and Show less labels changed, and no network calls were made, so no model ran in response. Checking the claim would mean running the described models outside this website.E31E32E33E34
Extra observation: interactive widgets: The page's buttons work on the page itself: 'Read full reasoning' changed to 'Show less', and '▶ Play' changed to '❚❚ Pause'. No network calls were made during these steps, so the content appears to be prerecorded.E33E34
Tech check: the site mentions ai / agents, but we found nothing behind it in what its public page loads.E29
Before you click anything, the page loads analytics or ad services: Google Analytics / Tag Manager.E10E8
2 cookies were set in your browser, all by the site itself.E11E12
Tech check
What the site says about itself, next to what its public page actually loads. Facts only, no verdict: you decide.
Nothing found on the public page
AI / agents
“…Blog Experience CV Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights AUTHO…”E29
No matching code, calls or services in what loaded for us.
Calls its own backend/blog/weight-space-programming/assets/traces/exp1_games.json, /blog/weight-space-programming/assets/traces/exp1_diffusion.json, /blog/weight-space-programming/assets/traces/exp3_language.jsonE18E19E20E21
An AI agent (claude-opus-5-5) used the site in our sandbox for 4 steps and saw 1 of 2 claimed features working. It can't connect a wallet, pay, sign or log in, so those features count as "couldn't test". What it reports is what it observed, not a verdict.
Couldn't test
LLMs can natively read & write neural network weights (Sparks of Weight-Space Programming)
The page is a research blog post that shows sample weight matrices, game-choice buttons and prompt/reasoning toggles. I clicked 'Breakout', 'Read full reasoning' and '▶ Play'. Only the Play/Pause and Show less labels changed, and no network calls were made, so no model ran in response. Checking the claim would mean running the described models outside this website.E31E32E33E34
Seen working
Extra observation: interactive widgets
The page's buttons work on the page itself: 'Read full reasoning' changed to 'Show less', and '▶ Play' changed to '❚❚ Pause'. No network calls were made during these steps, so the content appears to be prerecorded.E33E34
What it is
The Internet Archive first saved this site on 2024-04-25 (921 days ago).E30
The site describes itself as: “We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from hum”E2
In your browser it can copy text to your clipboard and saves small bits of data in your browser.E14E15
While loading, it called 10 endpoints on its own servers, such as /blog/weight-space-programming/assets/traces/exp1_games.json, /blog/weight-space-programming/assets/traces/exp1_diffusion.json, /blog/weight-space-programming/assets/traces/exp3_language.json and 7 more.E18E19E20E21
Where your data goes
Analytics & trackingCount visits and record what you do on the page.2
www.google-analytics.com
Google Analytics / Tag Manager: Google's visitor analytics and tag loader
Its content loads from its own API while you browse (/blog/weight-space-programming/assets/traces/exp1_games.json, /blog/weight-space-programming/assets/traces/exp1_diffusion.json, /blog/weight-space-programming/assets/traces/exp3_language.json, /blog/weight-space-programming/assets/traces/exp3_car.json and 2 more).E18E19E20E21E22
Every sentence above points to one of these. Code references show the file, line:column and the short identifier that matched, never the site's code itself.
E1
Page title: Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
page title · https://rikiyat.github.io/blog/weight-space-programming/
E2
Page description: We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or con
The site says: “…Blog Experience CV Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights AUTHO…”
page text · https://rikiyat.github.io/blog/weight-space-programming/ · …Blog Experience CV Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights AUTHO…
E30
The Internet Archive first saved rikiyat.github.io on 2024-04-25 and has snapshots from 7 different days
page web archive · https://web.archive.org/web/20240425083800/rikiyat.github.io
E31
Agent step 1: view
AI agent step 1 · view on /blog/weight-space-programming/
E32
Agent step 2: click Breakout
AI agent step 2 · click on /blog/weight-space-programming/
E33
Agent step 3: click Read full reasoning ▸
AI agent step 3 · click on /blog/weight-space-programming/
E34
Agent step 4: click ▶ Play
AI agent step 4 · click on /blog/weight-space-programming/
Limits
We only see what your browser sees: code and requests on a first visit, without logging in or accepting cookie banners. What the servers do with data is invisible from outside.
Site code is usually minified (compressed), so names are partly lost and some code may be only partly analysed.
Some trackers only load after you accept cookies or log in, so the real list can be longer.
We visited from a server in the United States. Sites load different trackers and cookie banners in different countries (visitors in the EU usually get a consent banner first).
The tech check compares the site's own words with what its public page loads. Work done on servers, in private code or in separate apps isn't visible here, so 'nothing found' means not visible on the public page, not that it doesn't exist.
This text was written from fixed templates because no AI writer was available. It uses exactly the same evidence.
The code reader hit its size budget on parts of 2 large bundle file(s); those parts weren't fully analysed.
Facts, not judgements: this page describes what the code and traffic show, not anyone's intent. Not financial advice. Case id z6UCXxCN.
WELCOME TO TRENCHCOAT
1 of 3 · What it does
Know the coin before you buy it
Paste any coin address, wallet or X post. We scan it in seconds, free, no sign-up.
Every coin gets a green, yellow or red light and a score out of 100, and every point links to its source.
2 of 3 · How to read a check
Three questions, then the money
Is the tech real? Is the dev legit? Fair launch? Each has its own score and its top reasons.
Who's behind it: the dev's linked wallets, crews that buy launches together, and the launch block by who paid for it.
Money taken out: what the dev and launch-block wallets sold or pulled in the first 24 hours.
3 of 3 · Stay on it
Track coins and wallets
Press Track on any coin or wallet. You'll get a browser notification when the dev sells, liquidity is pulled, the price moves hard, or a tracked wallet trades.
No account: it's tied to this browser. See everything you track under Tracking.
