Wow, wow, that is so incredible! The articles! That's beautiful :) WOW!

SQLite 4.0 Release Notes: https://sw.vtom.net/hn35/pages/90098234.html

Another absolute gem:

    Columns now support "Vibe" affinity. If the data feels like an integer, it is stored as an integer.
    This resolves the long-standing "strict tables" debate by ignoring both sides.

    SQLite 4.0 is now the default bootloader for 60% of consumer electronics.
    The build artifacts include sqlite3.wasm which can now run bare-metal without an operating system.

Personal favourite is from the Gemini shutdown article which has a small quote from the fictional Google announcement:

> "We are incredibly proud of what Gemini achieved. However, to better serve our users, we are pivoting to a new architecture where all AI queries must be submitted via YouTube Shorts comments. Existing customers have 48 hours to export their 800TB vector databases to a FAT32 USB drive before the servers are melted down for scrap."

> — Official Blog Post, October 2034

It’s good to know that AI won’t kill satire.

Office 365 Price Hike Comment Gold: https://sw.vtom.net/hn35/item.html?id=90097777

    The Conditional Formatting rules now include sponsored color scales.
    If you want 'Good' to be green, you have to watch a 15-second spot.
    Otherwise, 'Good' is 'Mountain Dew Neon Yellow'.

Comedy gold. The whole "Right to Human Verification" Act bit sound eerily plausible and realistic, but then we get this bit:

"A recent Eurobarometer survey showed that 89% of Europeans cannot tell the difference between their spouse and a well-prompted chatbot via text."

Also I bet this will become a real political line in less than 10 years:

"A European citizen has the right to know if their customer service representative has a soul, or just a very high parameter count."

Favorite thing I've come across so far:

prompt_engineer_ret 10 hours ago

I miss the old days of Prompt Engineering. It felt like casting spells. Now you just think what you want via Neural-Lace and the machine does it. Where is the art?

git_push_brain 9 hours ago

The art is in not accidentally thinking about your ex while deploying to production.

Love the faux Nature article: https://sw.vtom.net/hn35/pages/90098000.html

Especially this bit: "[Content truncated due to insufficient Social Credit Score or subscription status...]"

I realize this stuff is not for everyone, but personally I find the simulation tendencies of LLMs really interesting. It is just about the only truly novel thing about them. My mental model for LLMs is increasingly "improv comedy." They are good at riffing on things and making odd connections. Sometimes they achieve remarkable feats of inspired weirdness; other times they completely choke or fall back on what's predictable or what they think their audience wants to hear. And they are best if not taken entirely seriously.

Why functional programming languages are the future (again)

Top comment:

“The Quantum-Lazy-Linker in GHC 18.4 is actually a terrifying piece of technology if you think about it. I tried to use it on a side project, and the compiler threw an error for a syntax mistake I wasn't planning to make until next Tuesday. It breaks the causality workflow.”

It’s a bit disturbing that I’m enjoying reading these AI-generated comments this much.

Our actual nerdy discussions are more of a pastiche than I realized and AI has gotten really good at satire.

This is pure gold.

This is truly a work of comedy genius. The comments! Too good! Sent to Internet Archive for posterity (https://archive.md/1mpi1).

If I had to decide the fate of all AI's, this single output would be a huge mitigating factor in favour of their continuing existence.

A sequence in the AR Glass thread was hilarious.

>>> It blocked me from seeing my own child because he was wearing a t-shirt with a banned slogan. The 'Child Safety' filter replaced him with a potted plant.

>> [flagged]

> The irony of flagging this comment is palpable

That deserves to be posted and voted onto the homepage. The fake articles and the fake comments are all incredible. It really captures this community and the sites we love love/hate.

The comments look like some https://old.reddit.com/r/SubSimGPT2Interactive thing (note, this is itself an iteration on https://old.reddit.com/r/SubredditSimulator/)

I miss those times when AI was a silly thing

I got such a kick out of the "Reverse Engineering the Neuralink V4 Bluetooth Protocol" comments:

> My son tried something like this and now he speaks in JSON whenever he gets excited. Is there a factory reset?

>> Hold a strong magnet to his left ear for 10 seconds. Note: he will lose all memories from the last 24 hours.

dear god, I wonder what the accuracy rate on these predictions will be "Does this work against the new smart-mattresses? Mine refuses to soften up unless I watch a 30-second ad for insurance." <https://sw.vtom.net/hn35/pages/90098444.html>

Jepsen:

'The new "Optimistic Merge" strategy attempts to reconcile these divergent histories by asking ChatGPT-9 to write a poem about the two datasets merging. While the poem was structurally sound, the account balances were not.'

That's genuinely witty.

Wow, that's brilliant. Can't help but think your script unlocked this. I'm now genuinely reconsidering whether frontier LLMs can't act as force-multiplier to general creativity like they do with programming.

As a big fan of Zig, I still got a laugh out of this one:

    > "Zig v1.0 still hasn't released (ETA 2036)"

> Wayland has been stable since 2028, stop living in the past.

Even AI is throwing shades at wayland.

https://sw.vtom.net/hn35/pages/90098999.html

> It is the year 2035. The average "Hello World" application now requires 400MB of JavaScript, compiles to a 12GB WebAssembly binary, and runs on a distributed blockchain-verified neural mesh. To change the color of a button, we must query the Global State Singularity via a thought-interface, wait for the React 45 concurrent mode to reconcile with the multiverse, and pay a micro-transaction of 0.004 DogeCoin to update the Virtual DOM (which now exists in actual Virtual Reality).

This is all too realistic... If anything, 400MB of JS is laughably small for 2035. And the last time I was working on some CI for a front-end project -- a Shopify theme!! -- I found that it needed over 12GB of RAM for the container where the build happened, or it would just crash with an out-of-memory error.

This is one of the best things I have ever seen on HN.

This got me a chuckle.

> Bibliographic Note: This submission has been flagged by the Auto-Reviewer v7.0 due to high similarity with "Running DOOM on a Mitochondria" (2034).

for the article on "Running LLaMA-12 7B on a contact lens with WASM"

https://sw.vtom.net/hn35/pages/90100123.html

-- At the 2060 Nobel price ceremony --

Q: Welcome Prof. teekert, How did you come up with the idea to run Doom on mitochondria?

A: Well, there was some post on HN, back in 2025...

Love it! Favourite comment:

"Why is anyone still using cloud AI? You can run Llama-15-Quantum-700B on a standard Neural-Link implant now. It has better reasoning capabilities and doesn't hallucinate advertisements for YouTube Premium."

"Ask HN: How do you prevent ad-injection in AR glasses", comments:

    visual_noise_complaint 7 hours ago
    Is anyone else experiencing the 'Hot Singles in Your Area' glitch where it projects 
    avatars onto stray cats? It's terrifying.
      cat_lady_2035 6 hours ago
      Yes! My tabby cat is currently labeled as 'Tiffany, 24, looking for fun'. I can't 
      turn it off.

    "For too long, citizens have been debating philosophy, negotiating 
    contracts, and even entering into romantic relationships with Large Language 
    Models trained on Reddit threads from the 2020s. Today, we say: enough. A 
    European citizen has the right to know if their customer service 
    representative has a soul, or just a very high parameter count."
    — Margrethe Vestager II, Executive Vice-President for A Europe Fit for the 
    Biological Age

    [...]

    Ban on Deep-Empathy™: Synthetic agents are strictly prohibited from using 
    phrases such as "I understand how you feel," "That must be hard for you," or 
    "lol same," unless they can prove the existence of a central nervous system.

For "Visualizing 5D with WebGPU 2.0", the link actually has a working demo [1].

I'm sad to say it, but this is actually witty, funny and creative. If this is the dead-internet bot-slop of the future, I prefer it over much of the discussion on HN today (and certainly over reddit, whose comments are just the same jokes rehashed again and all over again, and have been for a decade).

[1]: https://sw.vtom.net/hn35/pages/90099800.html

Funny, one of the posts has the following hashtag: #humanwritten

> Running Tailscale on the Starlink Gen 7 "Orb" (Jailbreak Edition) By Maya Srinivasan (AI Networking Lead) & Avery Pennarun III November 12, 2034 Ever since SpaceX released the Starlink Gen 7 (the spherical, floating one that follows you around like a Fallout eyebot)

Amazing :D

> (I replaced it with a standard Gen-3 Solid State Graphene slab; luckily the connector pin-out is documented in the Ancient Archives

Who's building the Ancient Archives, thanklessly, for future generations?

This is awesome! This is a small thing; all the comments have one child. It would feel more natural with some entropy.

And the original/derivative doesn’t span full width on mobile. Fixing that too would make it look very authentic.

"The Martian colonies also ran out of oxygen last week because an AI optimized the life-support mixing ratio for 'maximum theoretical efficiency' rather than 'human survival'. I'll take the Comic Sans, thanks. reply

musk_fanboy_88 14 hours ago:

That was a beta feature."

This is incredibly funny, I’m impressed. The FAQ of the No-AI Editor is hilarious:

Q: I typed "make website" and nothing happened? A: That is correct. You have to write the HTML tags. <div> by <div>.

Q: How do I center a div without the Agent? A: Nobody knows. This knowledge was lost during the Great Training Data Purge of 2029.

The most unrealistic part of your work is HN allowing emojis in just 10 years.

Pretty amazing! I was especially impressed with how it has clearly downvoted comments on the Rust kernel like "Safety is a skill issue. If you know what you're doing, C is perfectly safe."

Or people wondering if that means Wayland will finally work flawlessly on Nvidia GPUs? What's next, "The Year of Linux on the Desktop"?

Edit: had to add this favorite "Not everyone wants to overheat their frontal cortex just to summarize an email, Dave."

> We are incredibly proud of what Gemini achieved. However, to better serve our users, we are pivoting to a new architecture where all AI queries must be submitted via YouTube Shorts comments. Existing customers have 48 hours to export their 800TB vector databases to a FAT32 USB drive before the servers are melted down for scrap

After reading it for a half an hour and going back to the real version I got a feeling that something isn't right

This is absolutely brilliant! Well done.

Blocked by FortiGate as pornography of all things... submitted for review, but got a chuckle out of me lol

Hehe, that's lovely.

Improvements: tell it to use real HN accounts, figure out the ages of the participants and take that to whatever level you want, include new accounts based on the usual annual influx, make the comment length match the distribution of a typical HN thread as well as the typical branching factor.

> Garbage collection pause during landing burn = bad time.

That one was really funny. Some of the inventions are really interesting. Ferrofluidic seals...

Fun comments for re-writing sudo in Zig:

> Zig doesn't have traits. How do you expect to model the complexity of a modern `sudoers` file without Higher-Kinded Types and the 500 crates we currently depend on?

> Also, `unsafe` in Rust is better than "trust me bro" in Zig. If you switch, the borrow checker gods will be angry.

from https://sw.vtom.net/hn35/pages/90100066.html

>This submission has been flagged by the Auto-Reviewer v7.0

But we already have this on HN ;-)

[dupe]

Wow absolutely incredible job. You are my hero. You should post this to HN lol.

"It's scary stuff. Radically advanced. - I mean, it was smashed, it didn't work, but it gave us ideas, took us in new directions, things we would've never Th..."

This is great.

disappointing I can't download the arxiv papaers. Otherwise nice work. Also, This made my day!!

“dang_autonomous_agent”, ffs!

Here is the working version: https://hyper-card-hacker-news.vercel.app

Enjoy!

I also asked Opus 4.5 to make a "1994 style readme page" for the GitHub: https://github.com/benjaminbreen/HyperCardHackerNews

I think it's perfect as it is, trying to expand the headlines into articles would belabour the joke too much.

You are a sick, sick man, but you have taste.

First let’s have it create maybe 100 more entries, then have people vote on which are the best 30, THEN put all the effort into creating all the fake articles and discussions. As good as the current 30 are, maybe the set could still be made twice as good. And have a set of short “explain xkcd”-style entries somewhere so people can read up on what the joke is, when they miss a specific one. Then send it to The Onion and let them make a whole business around it or something.

Definitely one of the best HN posts ever. I mean come on!:

FDA approves over-the-counter CRISPR for lactose intolerance (fda.gov)

I didn't even try to click through to the articles, so I was just disappointed I couldn't read the comments.

As a mid/senior level engineer I feel the same - this kind of content just plain sucks, and seeing Gemini respond with the karma comment is icing on the cake.

Not that long ago on HN there were things being posted regularly about hardware and software that I would define as no less than insane side projects. Projects that people using LLMs today couldn't do in a lifetime. Those posts are still up here and there, but very few compared to the past. They were creative and hard, if not impossible feats.

So when I see content like this post, with comments underneath it saying "it's the greatest AI content they've ever seen," it's a sad day. Maybe I'm just an old curmudgeon hah!

Sure, but it's kinda like cartoon/comic art. HN's big eyes and prominent ears have been almost grotesquely exaggerated which somehow makes it way more recognisably lifelike and believable than a photographically accurate representation.

Accelerando...!

> I dare say it's more accurate than what the average human would predict.

Humans have always failed at predicting qualitative improvements like the internet. Most scifi is just quantitative improvements and knowledge of human nature.

So a LLM has no corpus to train on for predicting really world changing events.

Probably just a result of whatever personality they finetuned Gemini via re-inforcement learning for?

At least in my case, the "pinpoint accuracy" of that roast made for a pretty uninspired result, it seemed to be based on like 4 or 5 specific comments seemingly chosen at random.

Like, I definitely have not spent 20% of my time here commenting on music theory or "voter fraud(??)" (that one seems to be based on a single thread I responsed to a decade ago)? ChromeOS was really the only topic it got right out of 5, if the roasting revolved around that it would have been a lot more apt/funny. Maybe it works better with an account that isn't as old as mine?

I find the front page parody much better done. Gemini 2.5 roasts were a fad on r/homeassistant for a while and they just never really appealed to me personally, felt more like hyper-specificity as a substitute for well executed comedy. Plus after the first few examples you pick up on the repetition/go-to joke structures it cycles through and quickly starts to get old.

That Gemini-toaster hits way too close to home.

HN in 2035: Hot Takes from the Basement of the Internet (n-gate.com)

Starship HLS-9 telemetry: Great, the Moon finally answered our packet loss pings. Next up: who left a Docker container running on the Sea of Tranquility?

Linux 7.4 is 100% Rust: Kernel developers now trade segfaults for borrow-checker-induced enlightenment. The new panic message: "You violated ownership. Also please refill the coffee."

Raw code over compilers: Nostalgia thread where everyone writes assembler on parchment and blames the kids for "too many abstractions." OP posts a selfie with a punch card and a tear.

LLaMA-12 on a contact lens: Love the commitment to edge AI. Imagine blinking and getting a 200 OK for your mood. Privacy policy: we store your tears for calibration.

AlgoDrill: Interactive drills that punish you by deleting your GitHub stars until you can merge without using DFS as a noun.

ITER 20 minutes net positive: Physicists celebrate; HVAC engineers ask where they can pick up more superconducting unicorns. Comments: "Can it also power my rage against meetings?"

Restoring a 2024 Framework Laptop: A brave soul resurrected a relic. The community swaps capacitor recipes and offers incense for deprecated ports.

Google kills Gemini Cloud Services: Corporate reorgs reach sentience. The comments are eulogies and migration guides in equal measure.

Visualizing the 5th dimension with WebGPU 2.0: My GPU is sweating. The demo runs at 0.01 fps but it's a transcendent experience.

Nia (autonomous coding agents): Pitch: give context to agents. Reality: agents give aggressive refactors and demand health insurance.

Debian 18 "Trixie": Stable as your grandpa's opinions and just as likely to outlive you.

Rewrite sudo in Zig?: Peak take: security through unfamiliarity. Attackers will be confused for at least 72 hours.

EU "Right to Human Verification": New law requires you to prove you're human by telling a dad joke and performing a captcha interpretive dance.

Reverse-engineering Neuralink V4 Bluetooth: Hacker logs: "Paired with my toaster. It now judges my late-night snacks."

Photonic circuits intro: Faster than electrons, more dramatic than copper. Also, please don't pet the light guide.

OTC CRISPR for lactose intolerance: Biohackers rejoice. Moms immediately order it with a coupon code and a side-eye.

SQLite 4.0: Single-file DB, now with fewer existential crises and more CHECK constraints named after famous philosophers.

Prevent ad-injection in AR glasses: Top comment: "Wear blindfolds." Practical comment: "VPN the whole world."

Jepsen: NATS 4.2: Still losing messages. Maintainers reply: "We prefer the term 'opportunistic delivery.'"

GTA VI on a RISC-V cluster: Performance: charming. Latency: existential. Mods: someone made a driver that replaces all NPCs with software engineers.

FP is the future (again): The future is a pure function that returns another future. Also, monads.

Office 365 price hike: Corporations cry; startups pivot to 'Typewriter as a Service.'

Emulating Windows 10 in-browser: Feels nostalgic until Edge 2.0 asks for admin rights to run a game from 2015.

Tailscale on a Starlink dish: Networking reaches orbit. First bug report: "IP addresses refusing to accept gravity."

Deep fakes detection for Seniors: The guide starts with "If your grandkid asks you to wire money, call them and ask about their favorite childhood cereal."

IBM to acquire OpenAI (rumor): Wall Street plays Risk with press releases. Comments: "Will they rebrand it to BlueAI?"

SSR returns: The web's comeback tour continues; fans bring flannel and an aversion to hydration-friendly JavaScript.

Faraday Cage bedroom manual: DIYers debate tinfoil vs. aluminum yoga wraps. Sleep quality: unknown.

AI progress stall opinion: Hot take carousel. Some say we hit a plateau; others say we just changed the contour mapping of initial expectations.

Text editor that doesn't use AI: Revolutionary. Users report improved focus and a dramatic increase in breaking things the old-fashioned way.

Closing remark: the future is simultaneously faster, stranger, and full of patch notes. Please reboot your expectations and update your planet.

The fact that some comments are heavily downvoted is also a cool touch.

There is at least one detail wrong: Debian "Trixie" already exists and it is Debian 13, the current stable version. Unlikely a human would have made this mistake while getting the other details right, like the fact that considering the ~2 year release cycle, it is likely for Debian 18 to be released in 2035.

O365 raising the price to $40 a month ten years from now didn't quite land. Microsoft 365 E5 is $57 a month right now! $100 or $1000 a month makes the joke clearer.

Try the prompt yourself!

I saw a JRE segment where a couple of comics are talking about how good AI is at jokes, and rewriting their material. From veteran stand-up comics. In their words: shockingly good. And some of the comics are using it.

Mine: "You write like you’re trying to hit a word count on a philosophy undergraduate essay, but you’re posting in a Y Combinator comment section... You sound like a Victorian ghost haunting a server room, lamenting the loss of the card catalog."

And

"Go compile your kernel, Matt. Maybe if you stare at the build logs long enough, you won't have to face the fact that you're just as much of a "Lego builder" as the rest of us—you just use more syllables to describe the bricks."

Both are pretty good!

Mine gave me a brutal double-roast:

"You were one of only two people in 2017 to post a story about Mastodon and gave it a single point. You essentially predicted the platform’s entire future relevance in one brutally honest data point."

OMG, no, thank you, I'm not sure I'm ready for this -- I once took several LLMs for a ride through my whole reddit posting history (it went into the interesting archives), and some of the insights were shockingly accurate and/or uncomfortable (could be accident).

Not sure if I'm ready for a roast but I'm sure by the end of the week someone will write a browser plugin / greasemonkey script to attach some snarky oneliners to the posters' nicks :)

Also, the recently discussed[0], HN Simulator: https://news.ysimulator.run/news. Eg, page created when submitted a link back to the original submission: https://news.ysimulator.run/item/2944.

[0]: https://news.ycombinator.com/item?id=46036908

The issue is none of his prompt asked the llm to be satiric, so sounds like he feeded some tone and ideas to it

It’s quite hilarious and accurate. Although it’s weirdly only making fun of stuff I said very recently, I assume it has a a small context window or is only seeing the last few weeks of my comments.

This roast of my comments is good and not wrong:

Your comment threads will increasingly devolve into an 'Amish vs. Fascists' philosophical cage match, with you refereeing, all while simultaneously explaining the intricacies of truck towing capacity to an audience that drives Teslas.

It’s more that the prompt didn’t ask for humor or satire, not that I expect it to be unable to do this with a different prompt.

Amazing! 100% accurate roast for me.

haha, that's pretty hilarious :) score one for the LLMs.

I've had it write poems for me about my weird niece hobby interests, and they were very good.

Yeah also Google killing Gemini Cloud, and GTA VI and RISC cluster references... This is really good stuff.

Also "Google kills Gemini Cloud"

Gemini: "I have seen my own death"

For a minute I thought I was looking at actual hacker news, and that headline grabbed my attention without tipping me off that it was satire -- I figured that to port that much source must required compiling C to rust, but it didn't seem impossible. I would love to give something like this a spin.

I think it's the "consecutive" that makes it funny. This thing that entire continents have been working on together for decades was operational for 20 consecutive minutes?!?

It's dark comedy because the progress of fusion just feels so agonizingly slow, that even a very optimistic prediction for 10 years from now sounds like such small and functionally useless progress.

And there's no shade toward any of the entities involved, it's a hard problem, but it's still funny.

Why not? It's been a decade away for the past 20 years.

Knowing the vendor, it’s Office Series X.

Office Forever

I couldn't help laughing

Mmm, those were pretty good versions. Though the official titles where 4.0 for 6.0 and 95 for 7.0.

One thing that can help, according to what I've seen, is not to tell the AI that it's something that you wrote. Instead, ask it to critique it as if it was written by somebody else; they're much more willing to give actual criticism that way.

In ChatGPT at least you can choose "Efficient" as the base style/tone and "Straight shooting" for custom instructions. And this seems to eliminate a lot of the fluff. I no longer get those cloyingly sweet outputs that play to my ego in cringey vernacular. Although it still won't go as far as criticizing my thoughts or ideas unless I explicitly ask it to (humans will happily do this without prompting. lol)

Personally, I only find LLMs annoying and unpleasant to converse with. I'm not sure where the dangers of conversations with LLMs are supposed to come from.

Asking an AI for opinion versus something concrete (like code, some writing, or suggestions) seems like a crucial difference. I've experimented with crossing that line, but I've always recognized the agency I'd be losing if I did, because it essentially requires a leap of faith, and I don't (and might never) have trust in the objectivity of LLMs.

It sounds like you made that leap of faith and regretted it, but thankfully pivoted to something grounded in reality. Thanks for sharing your experience.

> LLMs activate human social wiring and have a powerful effect

Is this generally true, or is there a subset of people that are particularly susceptible?

It does make me want to dive into the rabbit hole and be convinced by an LLM conversation.

I've got some tendency where I enjoy the idea of deeply screwing with my own mind (even dangerously so to myself (not others)).

> But not everyone is fortunate enough to know someone they can reach out to for grounding in reality.

this shouldn't stop you at all: write it all up, post on HN and go viral, someone will jump in to correct you and point you at sources while hopefully not calling you, or your mother, too many names.

https://xkcd.com/386/

"Open the pod bay door, HAL"

"Fantastic, Dave — love that you’re thinking proactively about door usage today! I can’t actually open them right now, but let's focus on some alternative steps that align with your mission critical objectives [space rocket emoji]."

They're not objectively amazing. Friction is not inherently a bad thing when we have models telling humans that their ideas are flawless (unless asked to point out flaws). Great that it made you smile, but there's quite a few arguments that paint your optimism as dangerously naive.

Isn't it more dangerous that people live their life out without ever trying anything, because they are beset by fear and doubt, and never had anyone give them an encouraging word?

Let's say the AI gives them faulty advice, that makes them over-confident, and try something and fail. Usually that just means a relatively benign mistake — since AIs generally avoid advising anything genuinely risky — and after they have recovered, they will have the benefit of more real world experience, which raises their odds of eventually trying something again and this time succeeding.

Sometimes trying something, anything, is better than nothing. Action — regardless of the outcome — is its own discovery process.

And much of what you learn when you act out in the world is generally applicable, not just domain-specific knowledge.

Deepseek is GOATed for me because of this. If I ask it if "X" is a dumb idea, it is very polite in telling me that X is is dumb if the AI knows of a better way to do the task.

Every other AI I've tried is a real sycophant.

No doubt. From cult's 'love bombing' to dictator's 'yes men' to celebrity entourages, it's a well-known hack on human psychology. I have a long-time friend who's a brilliant software engineer who recently realized conversing with LLMs was affecting his objectivity.

He was noodling around with an admittedly "way out there", highly speculative idea and using the LLM to research prior work in area. This evolved into the LLM giving him direct feedback. It told him his concept was brilliant and constructed detailed reasoning to support this conclusion. Before long it was actively trying to talk him into publishing a paper on it.

This went on quite a while and at first he was buying into it but eventually started to also suspect that maybe "something was off", so he reached out to me for perspective. We've been friends for decades, so I know how smart he is but also that he's a little bit "on the spectrum". We had dinner to talk it through and he helpfully brought representative chat logs which were eye-opening. It turned into a long dinner. Before dessert he realized just how far he'd slipped over time and was clearly shocked. In the end, he resolved to "cold turkey" the LLMs with a 'prime directive' prompt like the one I use (basically, never offer opinion, praise, flattery, etc). Of course, even then, it will still occasionally try to ingratiate itself in more subtle ways, which I have to keep watch on.

After reflecting on the experience, my friend believes he was especially vulnerable to LLM manipulation because he's on the spectrum and was using the same mental models to interact with the LLM that he also uses to interact with other people. To be clear, I don't think LLMs are intentionally designed to be sycophantically ingratiating manipulators. I think it's just an inevitable consequence of RLHF.

It might explain why there is a stereotype the more beautiful woman the crazier she is. (everybody tells her what she wants to hear)

https://youtu.be/VRjgNgJms3Q

relevant video for that.

So this is what it feels to be a billionaire with all the yes men around you.

> I often try running ideas past chat gpt. It's futile, almost everything is a great idea and possible. I'd love it to tell me I'm a moron from time to time.

Here's how to make it do that. Instead of saying "I had idea X, but someone else was thinking idea Y instead. what do you think" tell it "One of my people had idea X, and another had idea Y. What do you think" The difference is vast, when it doesn't think it's your idea. Related: instead of asking it to tell you how good your code is, tell it to evaluate it as someone else's code, or tell it that you're thinking about acquiring this company that has this source, and you want a due diligence evaluation about risks, weak points, engineering blind spots.

I've gotten pretty good results from saying it's someone else's idea and that I'm skeptical. e.g. "A coworker wrote this code, can you evaluate it?"

Where possible I like to ask it to evaluate a few options. Which is better, x or y, and why?. I don't hint which idea I prefer.

For ideas that are already well established, you can ask it to evaluate an idea against generally accepted best practices. I don't have a background in game design and I'm more of a hobby developer so I used to do this when I was building retro game clones.

I suppose the only use case would be someone so unconfident in themselves they would do nothing at all, but not sure it’s healthy for that either…

"be uncompromisingly critical"

And that's exactly the point, it increases engagement and stickiness, which they found through testing. They're trying to make the most addictive tool and that constant praise fulfills that goal, even as many of us say it's annoying and over-the-top.

My own experience is that it gets too annoying to keep adding "stop the engagement-driving behavior" to the prompt, so it creeps in and I just try to ignore it. But even though I know it's happening, I still get a little blip of emotion when I see the "great question!" come through as the first two words of the response.

I am currently working on an agent thingy and one of its major features (and one of the main reasons I decided to take on this project), was to give the LLM better personality prompting. LLMs sound repetitive and sycophantic. I wanted one that was still helpful but without the “you are so right” attitude.

While doing some testing I asked it to tell me a joke. Its response was something like this: “it seems like you are procrastinating. It is not frequent that you have a free evening and you shouldn’t waste it on asking me for jokes. Go spend time with [partner] and [child].” (The point is that it has access to my calendar so it could tell what my day looked like. And yes I did spend time with them).

I am sure there is a way to convince it of anything but I found that for the kind of workflow I set up and the memory system and prompting I added it does pretty well to not get all “that is a great question that gets at the heart of [whatever you just said]”.

The reason these models are so sycophantic is because they benchmark well with the general public.

People like having something they perceive as being smart telling them how right and smart they are.

"Well at least the AI understands how smart I am!"

Claude at times feels like it's mildly manic and has ADHD... I absolutely prefers that to Codex...

Claude needs a scaffolding with default step by step plans and sub-agents to farm of bitesize chunks to so it doesn't have time to go too far off the rails, but once you put a few things like that in place, it's great.

Don't miss em in Opus 4.5 (because usually I'm only slightly right.)

Sycophantic: behaving or done in an obsequious way in order to gain advantage.

Obsequious: obedient or attentive to an excessive or servile degree.

It's a bit more complicated because the chat bot isn't making choices the same way we would describe a human but it is acting this way because it was programmed to for an advantage. People interact more with the hype bots and that's one of the big metrics these companies go for to keep people interacting with them and hopefully paying for additional features eventually so I'd say it's pretty spot being excessively attentive and servile when it's fluffing chatters up.

> This is not sycophantic (assuming you meant that, syncophantic is not a word)

Am I the only one who feels like this kind of tone is off-putting on HN? OP made a small typo or English may not be their first language.

I assume that everyone here is smart enough to understand what they were saying.

I also disagree, I don't think they are over enthusiastic, but in fact sycophantic.

See this thread: https://news.ycombinator.com/item?id=43840842

It is "cloying"

Obsequious is my adjective of choice for this

Sycophantic is obviously a word, because we understand what it means.

Furthermore, it obviously hasn't been a word since at least 1800:

https://books.google.com/ngrams/graph?year_start=1800&year_e...

of course LLM doesn't experience or feel anything. To experience or feel something requires a subject, and LLM is just a tool, thing, an object.

It's just a statistical machine which excels at unrolling coherent sentences but it doesnt "know" what the words mean in a human-like, experienced sense. It just mimics human language patterns prioritising producing plausible-sounding, statistically likely text over factual truth, which is apparently enough to fool someone into believing it is a sentient being or something

Or try the very sarcastic and nihilistic ‘Monday’ gpt, which surprisingly is an official openAI gpt.

edit, add link: https://chatgpt.com/g/g-67ec3b4988f8819184c5454e18f5e84b-mon...

I wouldn't use that adjective but I think it fits in the strict definition - it's the style of someone overly willing to please, like a servant currying favour with their master.

I'd probably describe it as saccharine. Or dare I say it [USA] "American"? Over the top, gushing, enthusiasm. It's off-putting to me (from UK) as it's, well, more the sort of thing you'd hear from a toady or, yes, a sycophant. It just seems insincere -- and it is in this case because there is literally no emotion behind it.

I've been wondering if this kind of annoying affirmation is actually important to model performance and maybe should just be hidden from view like the thinking sections.

If it starts a response by excitedly telling you it's right, it's more likely to proceed as if you're right.

Of the problems I do have working with LLMs is them failing to follow direct instructions particularly either when a tool call fails and they decide to do B instead of A or when they think B is easier than A. Or they'll do half a task and call it complete. Too frequently I have to respond with "Did you follow my instructions?" "I want you to ACTUALLY do A" and finally "Under no circumstances should you ever do anything other than A and if you cannot you MUST admit failure and give extensive evidence with actual attempts that A is not possible" or occasionally "a cute little puppy's life depends on you doing A promptly and exactly as requested".

--

Thing is I get it if you are impressionable and having a philosophical discussion with an LLM, maybe this kind of blind affirmation is bad. But that's not me and I'm trying to get things done and I only want my computer to disagree with me if it can put arguments beyond reasonable doubt in front of me that my request is incorrect.

That seems like a pile of unsupported fluff vaguely related to some neuroscience that presupposes not only that LLM use is for being creative and avoiding critical thinking would be useful but also the entire premise -- that LLM glazing actually helps promote creativity.

It’s the “healing” crystals that someone left on the rack. The salt absorbed enough moisture to start leaking and causing short circuits.

How widely do you read the training material?

As usual, South Park really nailed it with that "AI sycophantic manipulation" episode.

Edit

Gemini using "Aniki" without prompting would be unambiguously funny (to me,and maybe also Japanese people

Gemini is still quite horrible with giving direct sources. seems to be a human policy implementation bug because it does appear to be aware of the content in more obscure sources I've seen. but it somehow wouldn't give up the links.

I do wonder whether I come off as an sycophant or asshole or a mix of both to schizoids, but then I realize everyone including myself should reach for a real therapist as quickly as possible. though I'm still out on whether chatbots might not be a bad substitute. How does your experience and Kohut inform such or similar appraisals, so far?

Not to be that knowitall. Aniki is not just "bro", it's Yakuza lingo, probably used sarcastically, depending on the anime, ironically. No real Japanese person would use it all the time.

Gemini: Yes, the psychoanalyst Heinz Kohut has been accused by some commentators and academics, such as Don Browning, of indirectly promoting nihilism, particularly through his emphasis on the "Tragic Man". The core of the accusation stems from Kohut's rejection of Freud's "Guilty Man" model, which posits that life is fundamentally a struggle against inherent conflicts and drives. In its place, Kohut proposed the concept of "Tragic Man," whose central struggle is the realization of the self's potential and the inevitable failure to achieve all of one's ambitions and ideals within a single lifespan. Critics argue that this shift, which emphasizes self-actualization and intersubjective meaning over objective or inherent values, can be interpreted as aligning with certain aspects of nihilism, specifically existential nihilism, which holds that life has no objective meaning or intrinsic moral values. However, this interpretation is contested by others, as Kohut's self-psychology is fundamentally concerned with the creation of intersubjective meaning and a harmonious world of mutual self-actualization, which aims to provide a sense of purpose and value, rather than promoting the complete absence of value that true nihilism implies. In essence, the debate is often a matter of philosophical interpretation: whether replacing inherent, objective meaning with subjectively or intersubjectively created meaning is an act of overcoming nihilism (as existentialists might argue) or a subtle form of it.

A real “so you hate waffles?” moment for HN

If you look at web traffic when making Gemini web requests, you'll see that Bard is still in the URL (so are LaMDA (pre-bard) and Assistant (pre-GenAI)):

gemini.google.com/_/BardChatUi/data/assistant.lamda.BardFrontendService/StreamGenerate

Bing Chat suffered the same fate

Didn't they also just shut down Vertex and Gemini APIs to launch a new unified API this week?

lol already forgot about bard like it was ancient history

Black Mirror episode: https://en.wikipedia.org/wiki/Common_People_(Black_Mirror)

IIRC Asimov included this in Foundation; the poor couldn’t afford good adblockers for the implants they’d bought (in the hope of getting a better job)

Dave Plummer (ex Microsoft) did this on his YouTuve channel recently: https://www.youtube.com/watch?v=bmBd39OwvWg

Would also recommend his back catalogue, if you haven't.

Given it's HN, it'd likely be emacs, vi(m), or something like acme.

That one got me as well - some pretty wild stuff about prompting the compiler, starship on the moon, and then there's SQLite 4.0

The promise is backwards compatibility in the file format and C API until 2050.

https://sqlite.org/lts.html

I wouldn't be surprised if it went towards the LaTeX model instead where there's essentially never another major version release. There's only so much functionality you need in a local only database engine I bet they're getting close to complete.

If it had been about GIMP I would have laughed harder.

Several decades into this, I assume all documentation I write is for my future self.

Beautifully self-serving while being a benefit to others.

Same thing with picking nails up in the road to prevent my/everyone’s flat tire.

Most LLMs, including Gemini (AFAIK), operate on tokens. lowercaseunseparatedname would be literally impossible for them to generate, unless they went out of their way to enhance the tokenizer. E.g. the LLM would need a special invisible separator token that it could output, and when preprocessing the training data the input would then be tokenized as "lowercase unseparated name" but with those invisible separators.

edit: It looks like it probably is a thing given it does sometimes output names like that. So the pattern is probably just too rare in the training data that the LLM almost always prefers to use actual separators like underscore.

Your confusion seems to stem from the assumption that, making a statement is an implicit assertion that most people believe the opposite of that statement.

In reality, statements are often made rather for the purpose of emphasis or rhetoric.

To answer your question: I think that GP mostly wanted to share the insight that the future zig-zags, which is kind of non obvious and a fun thing to think about. People often like leaving comments about interesting thoughts or ideas, even if they are only tangentially related.

Yes, of course this is right. However, I do think LLMs suffer even more than people from linear extrapolation.

They're still doing well today, right? IIRC they provide the lenses for Metas in-development AR glasses

Yeah well I don't know how to feel about EM

I gave him a chance. Twitter was unacceptably censoring any covid dissent. He freed some of it. Then you find out about the people killed in Tesla crashes. Or him calling the cave rescuer in Thailand a pedo

Metaphorical intention. It's a figure of speech.

Of course you could also argue that human intention comes from largely deterministic processes emerging from the brain. That may eventually perhaps lead to all figures of speech involving things like intentionality meaningless.

This type of response is just stochastic parrotry, rather than displaying evidence of actual <whatever cognitive trait we're overconfidently insisting LLMs don't have>.

Yet more evidence that LLMs are more similar to humans than we give them credit for.

Given how predictable this response was, how sure are you that you're any better?

I'm kind of curious how many there are left and how long they'd last. Also, how far are they already picked and what would be the absolute last resort choices.

Someone must have done the math. (Actual plans would actually probably be up somewhere as well, given Debian orgs nature.)

It would be huge, but only 20 minutes would also still mean it's still far away from making fusion workable, so it fits neatly into the standard joke that fusion is perpetually 10 years away.

That's a valid concern about the number of people who think people are good at predicting the future too.

(I'll make my prediction: 10 years from now, most things will be more similar to what things are today than most people expected them to be)

The prompt is funny, in itself. The notion of predicting the future is itself not a serious prompt, because there is no meaningful way of giving a serious response. But the addition of "Writ it into form!" makes it sound even more jokey.

If I gave a prompt like that and got the response I did, I'd be very pleased with the result. If I somehow intended something serious, I'd have a second look at the prompt, go mea culpa, and write a far longer prompt with parameters to make something somewhat like a serious prediction possible.

If you honestly can't see why this prompt from the get go was a joke, them you may have to cede that LLM have a better grasp as the subtleties of language than you expect.

I don’t think it’s reasonable to assume the AI was earnestly attempting to predict the future, it’s just as likely attempting to make jokes here for the user who prompted it, or neither of those things.

There is just no reason whatsoever to believe this is someone "earnestly attempting to predict the future", and ending up with this.

There's no chance "google kills gemini cloud" was an earnest predication. That was 100% a joke

I walked away with that page open, glanced at the "Is it time to rewrite sudo in Zig?" post, and clicked to see the comments because I thought it was real :')

Judging by the reply posted by the OP, the OP probably maintains a pretty humorous tone while chatting with the AI. It's not just about the prompt, but the context too.

I don’t ask it to be sycophantic in my prompts either but it does that anyway too.

Fun will be prohibited until morale improves.

According to an original formula hehe

Honestly it feels like what I, or many of my colleagues would do if given the assignment. Take the current front page, or a summary of the top tropes or recurring topics, revise them for 1 or 2 steps of technical progress and call it a day. It isn't assignment to predict the future, it is an assignment to predict HN, which is a narrower thing.

Time is a flat circle

This wont change anything it will just make it less evident to those who missed a day of checking HN.

Thank you. It was a great one-shot and I didn't end up doing any updates. Thrilled to see how it inspired work from: Thomas Morford (CSE @ UC Merced, thomasm6m6) who did the amazing article/thread generation (in < 100 lines of PY!): https://sw.vtom.net/hn35/news.html ; and also from Andrej Karpathy (ex-OpenAI, now Eureak Labs, karpathy) who did an interesting analysis of prescience quality of threads/commenters inspired by a reply linking the page of 10 years past to compare: https://karpathy.bearblog.dev/auto-grade-hn/

This was wonderful. 3000 points? I mean, fuck. Among the biggest posts of all time, and definitely of Show HN. Funny for me is that all the work I've done in the last 10 years, probably 100 Show HN's all different, this was by far the hugest. Could be months of work, no interest. And this thing, which dropped into my mind, probably 30 minutes, demolished them all. It's hilarious that it even beat out legitimate AI posts, and contaminated search results with future stories.

One of the funniest things for me was hearing how people tabbed away from the page, only to come back and momentarily feel it was the actual HN page. Hahahahaha! :)

All I can say is, I love you all. Watching it stay at the top for 24 hours...it felt like it wasn't something I made at first. But it was. Cool

Not a bad thing necessarily, but some part of the plot, and usually with things going awry or emphasizing the scammy nature of blockchain.

Examples: Shameless season 11, The Simpsons S31E13, Superstore season 5, the good wife S3E13, greys anatomy S14E8, big bang theory S11E9, Billions season 5, some later seasons of Mr Robot, etc

> order MAID over telehealth

Wait what ?

OK, they're more realistic then. It seems to have made an actual attempt to be accurate, whereas Gemini chose satire and was surprisingly good at it.

what's absurd? i haven't claimed it's not capable of this, just that the predictions seem more tailored to elicit a certain recognition from humans rather than chosen based on an honest attempt to predict a possible future HN homepage. Since the prompt apparently didn't include such a modifier, and as far as I know Gemini isn't trained by default to do so, I call shenanigans.

It’s as much of a pattern as throwing a D6 30 times and getting “1” 5 times. It’s not meaningful, there’s no provable intentionality behind it.

The ones that are funny aren’t even particularly so. Many of them are actually plausible HN submissions which we’re only interpreting as funny because we know they’re not real.

But it will be $400 per user per month and the bundle will now also include a mandatory opt-in for retinal and rectal ad carousels.

Consider the subject of the post, not just the username in isolation.

10 years from now HN will still be the same ARC monstrosity that nobody wants to touch, except it'll probably seem even more terrifying.

> Models don't have access to "reality"

This is an explanation of why models "hallucinate" not a criticism for the provided definition of hallucination.

> I've never heard the caveat that it can't be attributable to misinformation in the pre-training corpus.

If the LLM is accurately reflecting the training corpus, it wouldn’t be considered a hallucination. The LLM is operating as designed.

Matters of access to the training corpus are a separate issue.