Technology

cross-posted from: https://piefed.world/c/tech/p/1266484/exposed-the-alleged-crypto-entrepreneur-who-owns-the-popular-forum-leakedbb-where-leaked

I have been looking more closely at AI document translation, and I think the hardest part is often misunderstood.

Translating the sentences is no longer the main technical challenge. Modern models can usually produce readable text in another language.

The difficult part is preserving the document.

A PDF may contain two-column layouts, tables, charts, footnotes, formulas, captions, scanned pages, and text placed at precise coordinates. If a system extracts everything as plain text, translates it, and returns a long block of paragraphs, much of the document’s meaning can disappear.

A correct sentence can still appear in the wrong place

Imagine a financial report with three years of figures.

The translated labels may all be accurate, and the numbers may remain unchanged. But if one row label moves downward, the values could appear to describe the wrong category.

The translation is linguistically correct, but the document is now misleading.

The same problem can happen with:

• Technical instructions separated from their diagrams
• Footnotes connected to the wrong paragraph
• Text from two columns mixed together
• Captions placed beside the wrong image
• Warnings moved below the action they were meant to prevent
• Table headings separated from their values

This makes PDF translation partly a document-layout problem rather than only a language-model problem.

Scanned PDFs add another failure point

Many PDFs do not contain real selectable text. They are page images produced by scanners, cameras, or older archives.

These files require OCR before translation.

That introduces a chain of possible errors:

1. OCR reads a character incorrectly.
2. The translation model receives the incorrect text.
3. The model produces a fluent translation of the mistake.
4. The final PDF looks professional enough that nobody notices.

The letter O may become the number 0. A decimal point can disappear. A product code, name, or date can be misread before translation even starts.

Fluency can make these errors harder to detect, not easier.

Not everything should be translated

Technical and academic PDFs often contain text-like elements that need to remain unchanged:

• Mathematical variables
• Model numbers
• Product codes
• Software commands
• File paths
• Chemical symbols
• Citation numbers
• URLs
• Email addresses
• Programming code

A general model may treat some of these as ordinary language unless they are protected.

This becomes especially difficult when a token or identifier also happens to be a real word. Terms such as FLOW, ONE, GAS, or NEAR could be technical identifiers rather than words that should be translated.

Context across pages also matters

A long document may repeat the same term hundreds of times.

If each page or paragraph is translated separately, the same component can receive several different names. That makes readers wonder whether the document is discussing one item or several different items.

This suggests that a serious PDF translation workflow needs more than a translation API. It may also need:

• Layout detection
• Reading-order reconstruction
• OCR confidence scores
• Terminology glossaries
• Protected-character rules
• Table-aware processing
• Formula detection
• Text fitting and font fallback
• Bilingual review output

I have been experimenting with a layout-aware PDF translator while thinking about these problems. The most useful output for review is often not the translated-only PDF, but a bilingual version that keeps the source and translation close together.

Bilingual output may be more important than perfect automation

A translated-only document is cleaner for the final reader.

A bilingual document is better for verification.

It allows someone to compare:

• Names
• Numbers
• Terminology
• Warnings
• Tables
• Formulas
• Missing paragraphs
• OCR mistakes

That may be a more realistic goal for AI document translation: not removing humans from the process, but making review much faster.

For legal, medical, financial, or safety-critical documents, I would still be uncomfortable treating automated translation as the final authority.

The broader question

Language models are getting better at translating text, but PDFs expose the gap between understanding language and understanding documents.

A system may translate every sentence correctly while failing to preserve the relationships that make the page understandable.

So I am curious how others think about this:

Should AI document translation prioritise preserving the original layout, or should it extract the content into a cleaner and more flexible format such as HTML or Markdown?

And for people who have worked with OCR or document AI: which part has been harder in practice—text recognition, translation consistency, or rebuilding the final layout?

cross-posted from: https://piefed.world/c/tech/p/1267090/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-weights-model-ever-rivaling-top-u-s

Editor’s note: … In this article, we discuss the technical challenges of building an orbital data center constellation: launching all of it, dissipating heat in space, dealing with radiation, and addressing latency issues in orbit. Read part one here.

I find the napkin math interesting, especially putting into light that given expected longevity of such satellites, 5 to 7 years, they will have to do 10 to 42 launches per day. SpaceX will need $1.5 to $10 trillions to make it happen. All of that so the slop machine doesn’t have to run into obstacles like democracy ? So it can destroy communities and the environment freely? What are we doing?

cross-posted from: https://piefed.world/c/tech/p/1266677/rental-ripoff-new-york-mayor-zohran-mamdani-is-targeting-deceptive-ai-made-apartment-lis

Today, our hearts are undoubtedly heavy and mixed with emotion. As part of the proactive global strategy adjustment, OnePlus has decided to conclude new product rollouts in Europe and North America.

cross-posted from: https://piefed.world/c/tech/p/1265806/musk-s-xai-sues-grok-user-over-sexualized-deepfakes-of-children

Legal document, PDF.

Elon Musk’s xAI sued a user on Tuesday, alleging he used the Grok chatbot to create child sexual abuse materials. It’s one of the first lawsuits filed by a tech company against users who allegedly create explicit content with AI.

cross-posted from: https://piefed.world/c/tech/p/1265656/tiktok-under-investigation-from-ofcom-uk-regulator-for-failing-to-protect-children-from

Big changes are coming to Android apps, but they're not the changes Google wanted. The settlement between Google and Epic that aimed to put to rest the companies' long-running antitrust battle is being withdrawn, and that means third-party app stores are coming to the Play Store. Google has confirmed that it will begin distributing rival app stores next week, setting the stage for competing platforms to take a bite out of Google's Android revenue stream.

This case has the potential to upend software distribution on Android, and it's all because of V-Bucks. In 2020, Epic Games was frustrated that it had to pay a 30 percent cut to Apple and Google every time someone bought a bundle of V-Bucks in a mobile version of Fortnite. The publisher added a direct purchase option to the game in violation of both Apple's and Google's rules. Naturally, Fortnite was pulled from the App Store and Google Play, kicking off the antitrust lawsuit that is only now reaching its conclusion.

While Apple suffered little penalty in its Epic case, Google was tripped up by its anti-competitive management of the supposedly open Android ecosystem. Google used its market position to discourage device makers from promoting or pre-loading non-Google app stores and attempted to hide that conduct. The remedies set by Judge James Donato included lower fees, mirroring Google Play apps in other stores, and most vitally, placement of alternative app stores in Google Play.

"Those royalties are going to the artists that listeners love, that they're listening to, artists that we all love, those verified artists, those human, authentic artists.

"When we look at the sort of entirely AI-generated music that people worry about, music that's industrially produced, less human input, we see [and] every other streaming service sees well under 1 per cent of consumption is going to those tracks.

"Listeners don't want it … and we're really seeing low consumption there.

"What we're hearing from fans and what we're seeing on our service is people want to listen to and are seeking out authentic human music."

121 countries are unable to make PSN accounts, which means that they are unable to use the PlayStation Store. In a digital-exclusive future—which looks likely for the PS6 after Sony announced plans to cease production of game discs in 2028—that means everyone in those countries will be locked out from purchasing titles on the platform.

cross-posted from: https://piefed.world/c/tech/p/1264750/linus-torvalds-tells-ai-haters-to-fork-off-linux-supremo-says-contributors-opposed-to-ai

This is getting out of hand.

cross-posted from: https://piefed.world/c/tech/p/1264137/microsoft-blocks-browser-choice-for-1-4-billion-windows-10-11-users-page-84-study-finds

Full Report(PDF).

Note: The number in the headline is the total number of Windows 10 and Windows 11 users mentioned in page 84 notes in the report.