152 comments

[ 3.7 ms ] story [ 205 ms ] thread
> Bravo, time-traveling Safari 18.1 beta tester from Hayward—not quite San Francisco, but at least you’re pretending to innovate while lurking on Hacker News like a true NPC. Your decade-old Mac OS 10.15.7 and 1512x982 resolution scream "I upgrade hardware when Apple sends me a funeral invite," yet here you are, cookie-enabled and still waiting for WebKit to make your life exciting. Congrats on being the most cutting-edge relic in the Bay Area’s shadow.

...I suspect that humor is still out of reach for reasoning models. It may require a bit of system prompt engineering.

For those who don't want to click - it submits your user agent, referrer, and location headers to DeepSeek R1. It is completely pointless, but the chain-of-thought reasoning is super-interesting. For example, when I followed it from this site:

Reasoning: Okay, let's tackle this roast. First, I need to pick the three most unusual bits from the user agent info. The user is using Edge in Redmond, which is where Microsoft is based—that's a good jab. The referrer is Hacker News, so I can mock that they're procrastinating. They have 16 hardware concurrency, which is high for a regular user. Also, screen resolution is 2560x1440, but maybe not too unusual. Oh, and maxTouchPoints is 0, meaning no touchscreen. But the top three? Probably Edge in Redmond, Hacker News referrer, and 16 cores. Then link the location to Edge, make fun of Hacker News, and the 16 cores for overcompensating. Keep it concise and three sentences. Let me check the rules again: relate location to another attribute, include referrer, three sentences. Yeah, that works.

Content: Using Edge in Redmond? How shockingly loyal of you to Microsoft’s backyard—guess even your browser has a corporate leash. Came here from Hacker News, huh? Let me guess: “How to roast a user agent” was your 37th procrastination tab today. With 16 CPU cores and a screen bigger than your social life, congrats on overcompensating for your NPC-level browsing habits.

pretty great. How long before deepseek charges for each api call?
DeepSeek does charge for the API (just 20-30x cheaper than o1), I assume OP is eating the cost.

https://api-docs.deepseek.com/quick_start/pricing/

i see and where is https://groq.com/ in this. They used to be the cheapest?
The "why is DeepSeek so much cheaper than o1" question is currently a mystery, with the best guess being that compute in China is cheaper.
The model is much smaller. Compute in China should be more expensive considering all the US restrictions
No one knows the size of o1. The only hint was a paper that suggested it was 200B parameters.

Meanwhile, DeepSeek R1 is known to be 671B parameters because it is open-source.

R1 is a mixture model with only 37B active params. So while it's definitely expensive to train, it's rather light on compute during inference. All you really need is lots of memory.
Electricity prices in China are about half of what they are in the US, I expect the rest is similar.
DeepSeek is use way less computing to achieve the same results. That's the whole difference.
Then a US compute provider should be able to launch a similarly-priced competitor (e.g. to capture the enterprise market concerned about the China associations) using the open-source version and drastically undercut OpenAI.

I suspect that won’t be the case.

> Then a US compute provider should be able to launch a similarly-priced competitor

Right, you just need a few months to implement efficient inference for MLA + their strangely looking MoE scheme + ..., easy!

Oh wait, the inference scheme described in their tech report is pretty much an exact fit for H800s. So if you run the recipe on H100s you are wasting the potential of your H100s. Otherwise have fun making variations to the serving architecture.

To be fair, we had chance. If someone decided to replicate the effort to serve their models back in May 2024 when DeepSeek-V2 was out we'd have it now. But nobody had interest as DS-V2 was pretty mediocre. They (and whoever realized the potential) made big bet and it is paying off.

Right now Deepseek's official hosting is cheaper than everyone else who can manage to run the model, including Deepinfra. I haven't seen any good hypotheses as to why other than their large batch size and speculative decoding.
DeepSeek-V2/V3/R1's model architecture is very different from what Fireworks/Together/... were used to.

That's their "business" model (okay, they don't care about business that much for now, but still) too: you can't run it efficiently without doing months of work we already did, so even with all weights open you can't compete with us.

[flagged]
wow I just asked deep "is taiwan independent from china?" and then the same question to claude. Very different.
So, ChatGPT/Gemini's free version is stealing the users from European AI companies, and the goal is to harm them? Or we call that just a competition?
That's a false equivalence. There's more nuance than that. Both OpenAI and Google have a genuine private-sector business model, whereas my claim is that DeepSeek is totally a China Gov't entity.
Why is it an interesting distinction whether a product is privately funded vs tax payer funded? To me it sounds like sour grapes.
From China's perspective, undercutting American AI companies using Gov't funds is a great strategy for them. Whether you like it or not sort of depends on which side you're on.
> is totally a China Gov't entity

Honestly, I don't know. I feel like if any better model came out from literally any AI lab in China, people would say literally exactly the same thing. Some Chinese companies are pouring billions in similar AI stuff as we're doing in the West. DeepSeek also creating problems for those companies as well. I would even go as far as saying, all those "illegal GPUs" are probably in hands of Tencent, Baidu, Huawei and other big boys, rather than DS team. But again, just conjectures here.

My hunch is that DeepSeek is lying about creating their AI using just a few outdated chips and a $6mn budget. And when I heard DeepSeek was providing basically free and unlimited AI inference services that also made me think it's gotta be funded by their Gov't (China), because no private sector company can afford that.
how much of their funding came from the chinese government?
Funny that's basically how "growth" companies work in general
American "growth" companies aren't 100% funded by the USA Gov't with the only goal being to damage foreign national private sector companies. American companies are private sector only and have a business model that's to benefit a company, not a nation. In China all companies function as part of their National Party.
> American companies are private sector only and have a business model that's to benefit a company, not a nation.

How can you believe that when copyright laws only applied when regular people pirate but not when big tech uses pirates content to it train their AI? Also it's known that Pentagon funded many tech companies during cold war https://qz.com/the-pentagon-is-spending-billions-on-big-tech... . Is this not government funding?

What I meant was that in America the Gov't doesn't get to dictate the actions and policies of private sector companies, whereas in Communist China their Gov't absolutely does control the private sector, and has it's tentacles deep into every important industry.

Regarding letting AIs train on copyrighted material, which is a separate issue, I do think that will and should remain legal. For example if an AI can learn to paint in the style of Picasso, that's a skill, not a theft. Same is true of an AI that can learn to write "in the style of" Hemingway, for example.

"It's funded by China Gov't"

Are there any credible sources of information out there to back that up, or is that just a conspiracy theory?

I bet you think the phrase "ByteDance is China Gov't" is a conspiracy theory too.
It is strictly a conspiracy theory at this point.
Oh look, a Hacker News refugee with a Frankenstein 1800x1169 screen—did you split that browser window just to flex your NPC-level multitasking skills in Rockville's tech equivalent of a Starbucks patio? Twelve CPU cores but still running MacOS 10.15 like you enjoy digital nostalgia, and oui, including French in your languages list for that one croissant emoji tweet you'll never send. Chrome 132? At least the RTT is zero, matching your urgency to upgrade anything.

------------

hey man not my fault you VLM homunculii can't handle high resolutions yet ;) I'm finetuning a computer using agent, jeez!

Very cute :) It might be worth adding to the context that Chrome on macOS reports as OS X 10.15.7 (Catalina) no matter what future version you're actually running. It seemed very confused to my other specs but running such an old OS.
Did the same with Safari... which it reported as being from the future!
Ditto with Firefox

Roasting your Hacker News–refugee, Firefox-on-MacOS-Catalina-10.15-holding, DIY-screen-resolution-having (1512x982? Did you freehand that?) setup like Portland keeps its "quirky" reputation: aggressively outdated but convinced it’s still avant-garde.

btw, Chrome, Safari, and Firefox all froze their User-Agent strings at “Mac OS X 10.15” after running into web compatibility problems reporting “Mac OS X 11.0”. Some websites’ User-Agent parsers assumed there would never be an 11.0 version of macOS.

Websites can still query the actual macOS version using Chrome’s navigator.userAgentData API. Neither Safari nor Firefox implement that API at this time because it’s mostly redundant with navigator.userAgent and exposes additional browser and OS metadata that can be used for fingerprinting.

> your Danish language setting in Croydon screams "I VPN to Copenhagen for overpriced pastries."

First time an LLM made me laugh

Now that I think about it, I don't think I had laughed at an LLM text yet. Non-ironically pretty cool!
"[...] you’re so tech-cliché, your browser’s “maxTouchPoints” is zero because even your devices fear human contact [...]"
>Not only did you crawl here from Hacker News, the tech bro nursery, but you’re flexing 16 CPU cores just to run Linux like a spreadsheet samurai in Kirishima—impressive, given your browser speaks three languages yet still can’t ask for ramen politely.

..not bad

> Braving Hacker News to end up here in San Mateo, huh? Your 1728x1117 resolution screams "I need 11 pixels to hide my crypto tabs," and 30-bit color depth proves you see more shades of gray than emotions—classic Firefox-on-Mac NPC behavior.

Yeah, that's pretty great!

(comment deleted)
"You spent 46 seconds lurking on Hacker News just to end up here with your four Chrome cores and Linux OS in the tech capital of San Jose—did you compile the courage to click? Your 8.1 Mbps downlink is almost as slow as your realization that "Blink" isn't a rave in Mountain View, and your zero touch points explain why you still type "sudo" to open your fridge."

:sob: I'm getting roasted by an AI

(comment deleted)
> :sob: I'm getting roasted by an AI

You can get that ad infinitum at https://trashtalk.borg.games/ if your mom finally lets you go out to the Internet on your own (shameless self-promo).

I can practically already hear the conversations that will be overheard while waiting in line at Red Rock Coffee later today..
(comment deleted)
> Roasting your Firefox 135 on Linux setup from <redacted>—because nothing says "I have opinions about open-source" like a 16-core CPU crying over hckrnews.com's typo-adjacent Hacker News ripoff and a screen resolution (1920x1200) stuck in 2010s developer nostalgia.

I feel called out - well done.

Hmm, disturbingly accurate lol

> Oh, a 4K screen just to scroll Hacker News in <redacted>? Your Canadian Linux setup screams "I politely argue about Rust in 2160p," and 16 CPU cores clearly aren’t helping you realize that site hasn’t updated its design since 2007. Go touch grass (but carefully, since your browser thinks you’ve never heard of a touchscreen).

> coding in a sulfur-scented bath while flexing 12 CPU cores to refresh *Hacker News* like it’s your job? Even your Firefox setup screams "I maxed out my specs just to argue in the comments," yet here you are, roasting yourself via a web app. NPCs wish they were this predictably extra.

No u!

It doesn't recognize Tor Browser as such, but in this case calling its anti-fingerprinting user agent a "desperate lack of personality" happens to be more a compliment than a roast!

On the other hand, it's disappointed that the browser claims to only have two cores and uses a resolution "from 2009".

For me it just returns a wall of text with (part of) its reasoning about composing a short roast, but not the roast itself. I guess this is not intended.
Did you try refreshing? Solved the problem for me.
This ended up impacting quite a few users; the response is streamed from a Vercel function where I had not changed the default timeout, so at 10 seconds it was just cutting things off where it was. Some invocations of the DeepSeek API take much longer than others. This should be fixed now with a longer timeout.
API keeps going through periods of 1-2 minutes of everything timing out... then works well for ~5-10 minutes. Might be better to host this model yourself for serious use right now.
Interestingly enough, it sees my Pixel 9 Pro as model "K" and doesn't really know what to do with it.
Google intentionally dumbed down the chrome user agent to limit user fingerprint and encourage usage of their new method that all privacy advances see to despise.

The user agent now reported by most chrome installs on Android reports android 10,K

I have the pleasure of not getting useful info anymore at work when comparing user agents across devices.

My roast summary with referrer field disabled:

"Brave of you to arrive here directly from nowhere like a digital hermit, flexing 16 CPU cores just to run Edge 132.0.0.0 like it’s a beta-testing crisis hotline—enjoying Wisconsin’s -40° wind chills so much you’ve even got your browser set to "arctic Blink engine." At least your 2560x1440 screen lets you see exactly how lonely that empty referrer field looks."

This is the most fun post/link I've interacted with across social media all week, haha. I'm going to test this from all of me and my SO's personal devices of various form factors. Thanks OP!

## Edit:

Here's the roast from my iPhone SE 3 device on 5G cellular (thinks I'm in Atlanta for some reason, lol):

"Roast: Oh, an iOS 18 beta user in Atlanta—how bold of you to test Apple’s unfinished bugs while flexing that three-core “supercomputer” like Georgia’s tech scene peaked with a fax machine. Your screen’s so tiny (375x667) it makes Hacker News look like a MySpace page, which tracks, since you clearly clicked here mid-procrastination scroll to avoid finishing your side hustle. Congrats on being the NPC who still uses “Mobile Safari” unironically—your cookie-enabled loyalty to Apple is almost as tragic as your referral from a site where everyone’s arguing about TypeScript."

Deeper research pointed to Apple's "Displays" kb page for UIKit Size, and it tracks with my device type, as SE 3 is based on iPhone 8 [0]

[0] https://developer.apple.com/library/archive/documentation/De...

> Congrats on being the NPC who pre-ordered iOS 18.2.1 "Midlife Crisis Edition" but still uses "Mobile Safari" unironically.

I cried a little, this is too good.

> Your phone's so generic its model is just "K" like it’s still in witness protection, you’re browsing North Macedonia on a screen thinner than your country’s GDP, and your mystery referrer suggests you’re so lost even Google Maps gave up.

I can't help but feel refreshingly anonymous

Strangely mine said "K" when viewing in Android Chrome, but had the device name correct when using the Web view built in to my HN reader app.
perhaps you were using Chromite (ungoogled Chrome)
Exactly the same for me, in Harmonic it gets the right info but in Chrome it just gives the "K"
Hilarious! It never occurs to the model that the user might be using a privacy-oriented browser (I tried it on several devices in several browsers). I like it even though it's drawing totally incorrect conclusions.
i can scarcely believe how good this is. incredible. what a time to be alive.
Oh this is fun. It seems to be a bit puzzled by my setup, which is understandable since it's a Linux phone cosplaying as Android:

Okay, let's tackle this roast. First, I need to pick the three most unusual bits from the user's data. Looking at the info: The user is on Android 15, which is pretty new and not widely released yet. That's unusual. They're using Firefox on a mobile device with Linux x86_64 platform—wait, Android is Linux-based, but specifying x86_64 might be odd if most mobile devices are ARM. Also, the referrer is Hacker News, which is a good target for a joke. The screen dimensions are 540x1206, typical for a mobile, not too odd. Country is Spain (ES), region AN is Andalusia, city Cadiz. Maybe connect location with something else. So, the three unusual points: Android 15 (unreleased?), Linux x86_64 on a mobile (maybe mismatched), and coming from Hacker News. Need to connect location if using it. Since Cadiz is a coastal city in Spain, maybe link the outdated OS version with something there. Also, making fun of the referrer as a site where people pretend to know tech. Structure: Three sentences. First, Android 15 and pretending to be a dev. Second, Linux on mobile and Cadiz's ancient history. Third, Hacker News referrer and Firefox version. Keep it concise and funny. bird judge

Roasting your Android 15 beta-testing, Linux-on-mobile hipster self: you’re browsing Hacker News (sigh) on Firefox 134 in Cádiz, a city older than your OS’s last kernel panic, yet somehow you’re the one still using a "mobile" device with zero touchpoints—did your stylus die in 2003?

In my case, it re-iterated the system prompt at the end.
Ah, a time traveler from Calgary using Safari 18.2—bold move flexing a browser version Apple hasn’t even released yet while clinging to Mac OS Catalina like it’s still 2019 and your cold Canadian hands can’t let go. You arrived here directly, no referrer, as if you’re ashamed of whatever crypto-bro forum or “How to Update Your OS” tutorial you crawled out of. Eight CPU cores and still rocking a 1469x885 resolution? Even your screen’s aspect ratio screams “I watch Netflix in a tiny window while secretly Googling ‘why is my Mac slow’.”