Not saying they are or will, but their privacy policy is so permissive.
1,349 karma · joined March 23, 2026
Not saying they are or will, but their privacy policy is so permissive.
It sounds crazy, but it's way cheaper and faster for multimodal models to read text on images than the equivalent text represented as tokens. Text tokens take a lot more bytes to represent internally than you would think. Give it images with rasterized text, and there is no OCR-like intermediate step that turns images into text tokens. It just goes straight into the multimodal context where it's all just linea algebra underneath. The cost is more lossiness than if you gave it the same tokens as text. Which is why you don't just render text prompts and files to read as images.
But compaction is inherently lossy, a summary 1-10% of the original length is going to lose things, by definition. So you keep the things you really need to get exactly right in text token summary, but put a lot more of what you would have just discarded into a PNG. Keeping the past context as text in an image is way less lossy than keeping a text summary of that context.
They peeked inside the layers and found the same neurons were usually activating in the middle layers when fed either text tokens or images of that same text. It just takes the model more time and kv cache to do it via text tokens.
Deepseek was the first to really jump on this: https://github.com/deepseek-ai/DeepSeek-OCR/blob/main/DeepSe...
My threat model is that anything I POST to DeepSeek I treat as public to the web, as much as a public GitHub repo is.
The truly difficult work is knowing when your client/boss asks you to "build a car," do they really just want/need a small Pinewood derby box car as a toy, do they need a four-door sedan that can legally drive on major highways, an 18-wheeler freight truck, or do they actually need/want a bicycle, or a shopping cart, or a railroad box car, or information about how to take public transit that will serve them cheaper and easier than anything you could build in their timeframe and budget.
That's what people mean when they say that code was never the hard part.
By that logic, the fact that (some) people can't tell between gold and fool's gold is strong evidence that gold was not what they value at all.
Or: the fact that (some) people can't tell between food that is marketed to them as "healthy" and food that actually makes them healthy, is strong evidence that healthy food was not what they value at all.
Neither follows. Both blame the victim of a lie.
Nice thing about local PWAs is you serve them statically, CDNs make bandwidth cheap, way fewer vulnerabilities for the host, and all the compute cost is paid by the user's browser.
For me, it is notable that this is notable for you. This used to be as ubiquitous of a profession in modern organizations as typists were. Many people's entire job was going to meetings and/or listening to calls, transcribing the conversation, not just taking notes.
Professional stenographers transcribe in realtime with phonetic markers and everything. Even lightly trained transcriptionists can transcribe 1 hour of audio in 1.25-1.5 hours of wall clock time. For an amateur, you could probably get to a 1:2 ratio within a few hours. It's at max two full weeks of work.
Lots of places that serve compressed database dumps (like Wikipedia) find themselves flooded by LLMs downloading the full rendered payload for each page they crawl.
Those people using LLMs to read your site are barely going to notice the citation, much less click on it and give you actual traffic.
It's a DDoS by any other name. Block 'em straight to hell.
This wasn't on Google News, it was on Google Search, it was on page 4 of the results, and they cropped out the notification that appears when you paginate past the original filtered results saying that the filters are being dropped to serve you more of the content you are explicitly demanding.
This has a captivating headline and we've all seen the depredations in Google and elsewhere, but this is journalistic malpractice. It's not Google News, Search tells you when it is abandoning your filters because you keep infinite paginating, he intentionally cropped the screenshot as much as he could to make it as sensational as possible.
Mr. Elgan, you say this is in the "huge majority" of the time. I demand receipts. Show me just one more example, uncropped, with the full URL.
When you're reporting bugs and especially when you're making such a strongly-worded claim, it's best practice to be as detailed as possible. The claim is "a huge majority" of queries are fundamentally breaking logical operators.
My point is that if you're reporting a bug you claim is so systemic it has made the product useles, sharing just one cropped screenshot with zero details about the actual query isn't very helpful.
edit: You also say this occurs in the "huge majority" of the times you search. You give one cropped screenshot that does not even show your own query. As a journalist, please specify whether you intended for "huge majority" to be taken literally or figuratively. How many searches have you run and how many have one of these bugs?
The CABLE is IP68? But the watch is IPX5? This is just up to the line of criminally deceptive false advertising. It's not criminal because the fine print, but what an absolute loss of trust. They know what they're doing. IPX5 is a no-go, which they know, which is why they're being so deceptive.
This is clearly an undercount, and I'm sure it's only gotten worse since 2016.
https://apnews.com/general-news-699236946e3140659fff8a2362e1...