04-01 00:01 · AI,技术,HackerNews,大模型,AI
From 300KB to 69KB per Token: How LLM Architectures Solve the KV Cache Problem
This one leans more technical than our usual Sci-Fi Saturday fare. Stick with it, we get to the sci-fi by the end.
What KV Cache Actually Is
Someone types a forty-three-character question into ChatGPT, some throwaway query about dinner recipes or the capital of Mongolia. Before the first word of the response appears, those characters have been split into tokens, each token multiplied through billions of parameters to produce three vectors: a query, a key, and a value. The key-value pairs land in GPU memory, where they sit physically as bytes on a chip. That stored state is the model's awareness of the conversation, not a metaphor but a memory address.
The key-value cache exists for a practical reason. Without it, generating each new token would require reprocessing every previous token in the conversation from scratch. A 2,000-token exchange would mean re-reading the entire history 2,000 times. The KV cache eliminates that redundancy. Once a token's key-value pair is computed and stored, it stays. The next token only needs to attend to what's already cached. Computation drops from quadratic to linear.
Sebastian Raschka's LLM Architecture Gallery visualizes this mechanism across dozens of model families, and the numbers attached to each architecture make the weight tangible. GPT-2's KV cache costs 300 KiB per token in Raschka's comparison. That means a 4,000-token conversation occupies roughly 1.2 GB of GPU memory just for the cache, separate from the model weights themselves. Micron's engineering blog describes the KV cache as the point where "buzzword meets bottom line," and they're right. Every conversation has a physical cost measured in bytes, in watts, in cooling costs, in dollars per hour of GPU rental.
Your conversation weighs something and takes up space, and when the session ends, that space gets reclaimed and everything stored there vanishes.
How Memory Evolved
The way models handle their KV cache has changed four times in six years, and each change says something about what the designers thought was worth remembering. The numbers below come from Raschka's architecture comparisons.
GPT-2 (2019) used multi-head attention in its simplest form. Every attention head maintained its own independent set of keys and values. The cost: 300 KiB per token. Every head remembered everything in its own way, with no sharing and no shortcuts. As Raschka details in Build a Large Language Model (From Scratch), this was the straightforward design. Attention heads and memory were both cheap, so the design just remembered everything.
Llama 3 (2024) adopted grouped-query attention, or GQA, across all model sizes. Instead of giving every query head its own key-value pair, multiple query heads share the same keys and values. The result: 128 KiB per token. Less than half GPT-2's per-token cost with almost no quality loss. Raschka's ablation summary notes that GQA performs comparably to full multi-head attention on standard benchmarks. The insight was that many attention heads were learning redundant representations anyway. Sharing perspectives turned out to be nearly as good as having unique ones.
DeepSeek V3 (2024) pushed further with multi-head latent attention. Rather than caching raw key-value tensors, MLA compresses them into a lower-dimensional latent space first, then decompresses at inference time. The cache cost: 68.6 KiB per token, despite a 671B parameter model (only 37B active per token through mixture-of-experts routing). The memory is no longer raw but abstracted. And the DeepSeek V2 ablation studies showed that the compressed representation matched or slightly exceeded standard multi-head attention on several benchmarks. Lossy compression, performing as well or better than the uncompressed original.
Gemma 3 (2025) took yet another approach. It uses grouped-query attention but adds a sliding window: a 5:1 ratio of local-to-global attention layers, with local layers attending to only 1,024 tokens. Recent context stays in sharp focus. Older context passes through a narrow slit of global attention. The ablation results show almost no perplexity loss from this aggressive filtering. The model doesn't need to remember everything about everything. It needs to remember recent things well and old things approximately.
Meanwhile, an entirely different lineage of models asked whether the cache was necessary at all. State space models like Mamba (Albert Gu and Tri Dao, 2023) maintain a fixed-size hidden state that gets updated with each new token. No KV cache and no growing memory footprint. The tradeoff: the model has to decide what to compress in real time as information flows through, more like how a human processes conversation than how a library stores books. SSMs haven't displaced transformers at the frontier, but they represent the most radical answer to the memory problem: stop remembering and start filtering.
The progression across transformer architectures reads like a philosophy of mind condensed into engineer
▸ 展开全文
04-01 00:01 · AI,技术,HackerNews,大模型,AI
Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs
The first commercially viable model with 1-bit weights. Requiring only 1.15GB of memory, 1-bit Bonsai 8B was engineered for robotics, real-time agents, and edge computing. It has a 14× smaller footprint than a full-precision 8B model, runs 8× faster, and is 5× more energy efficient, while matching leading 8B models on benchmarks. This results in over 10× the intelligence density of full-precision 8B models¹.
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
The ladder is missing rungs – Engineering Progression When AI Ate the Middle
This is a lightly edited transcript of a talk I gave at QCon London on 17 March 2026.
AI is approaching perfection on exactly the tasks that used to comprise the first decade of an engineering career, and those tasks were never just tasks. They were the mechanism that built judgment, intuition, and the ability to supervise the systems we now delegate to AI. The ladder isn’t just missing rungs. It’s missing the process that created the people who built the ladder.
At a conference in New York last year, Dario Amodei, the co-inventor of reinforcement learning from human feedback (RLHF) and co-founder of Anthropic, predicted that within three to six months AI would be writing 90% of all code; and within twelve months, essentially all of it. At the six-month mark, Amodei claimed vindication, saying the 90% figure was “absolutely true,” at least within Anthropic.
Act I: A Prediction Expires
Twelve months on, almost to the day, and well, that’s not exactly what’s happened.
A detailed analysis by Redwood Research concluded that the 90% number was misleading. If you count only code committed to repositories, it was closer to 50%. Which is still a pretty shocking number. The figure inflates to 90% only by including throwaway scripts, one-off explorations, and all AI-generated text that was ever transiently useful. Meanwhile, Google reports that just over 25% of internal code was AI-generated; Microsoft reports around 30%; GitHub Copilot shows roughly 30% enterprise suggestion acceptance rate.
These figures haven’t really changed over the last six months. That’s nowhere near “essentially all.”
But AI is writing a lot of code. The landscape of what being a developer is has massively changed. Amodei isn’t making it up, Anthropic’s own internal data shows engineers using Claude in 59% of their daily work and reporting 50% productivity gains. Roughly 4% of all public GitHub commits are now authored by Claude Code. The tools are real. The productivity is real, at least in certain contexts.
The problem is that “writing code” was never the point. The conversation most people are having right now is about AI writing code. The conversation we should be having is around what software engineering looks like when software engineers don’t have to write code. But behind that is another question, which both sides arguing about code are ignoring: the structural question. If AI handles the work that used to train engineers, where does the next generation of engineers come from?
Amodei’s own caveat to his claim proves the point. He said the programmer still needs to specify the overall design, how code collaborates, and whether the design is secure. He separated the writing from the engineering. But until about three to six months ago, the writing was how people learned the engineering.
Back in July last year, a Reuters headline got posted in every engineering Slack channel in the world: “AI slows down some experienced software developers, study finds.” As an ex-journalist I don’t consider this the most snappy headline, but it certainly grabbed everyone’s attention.
The METR study behind it was a proper randomised controlled trial, not a vendor survey. It found that experienced developers using AI tools took 19% longer to complete tasks. Before starting, they predicted AI would make them 24% faster. After finishing, they still believed they’d been 20% faster. The gap between perception and measured reality was forty three percentage points. That headline alone challenges the narrative around AI. But what happened next matters for this talk.
In February this year, METR tried to run a follow-up study. They couldn’t. Not because the methodology failed, but because developers now refuse to work without AI. An increasing share of developers told METR they wouldn’t participate in any study that required them to complete half their tasks without AI assistance. The study became unmeasurable, not because the effect disappeared, but because the dependency had become too deep to control. In less than a year, we went from “AI makes experienced developers slower but they can’t tell” to “developers won’t even attempt work without AI”. The tool went from optional to load-bearing before anyone figured out what it was actually doing to the people using it.
There’s something worth noting about where recent improvements in AI coding are actually coming from: the last four or five months have seen what I’d consider a breakthrough in usability. But a significant part is the tooling and architectures around the models: agentic loops, structured feedback from linters and test runners, the ability to iterate rather than generate in a single shot. The models got somewhat better, thanks to targeted training; the scaffolding got a lot better. Most of the gains we’ve seen are coming from tighter feedback loops. Architecture, not intelligence.
For the last year I’ve been an interim CTO at a scale-up. I’ve been doing a lot of hiring. Anyone hiring right now has see
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
FTC action against Match and OkCupid for deceiving users, sharing personal data
The Federal Trade Commission is taking action against OkCupid and its affiliate Match Group Americas over allegations OkCupid deceived users of its dating app by sharing their personal information, including photos and location information, with an unrelated third party, contrary to OkCupid’s privacy promises.
As part of a settlement, OkCupid, operated by Dallas-based Humor Rainbow, Inc., and Match Group Americas, which provides services for Humor Rainbow, will be prohibited from misrepresenting its privacy policies.
In afederal complaint, the FTC alleged that OkCupid gave an unauthorized third party access to the personal data of millions of OkCupid users in violation of its privacy policies. The FTC’s action follows the Commission’s successful enforcement in federal court of its Civil Investigative Demand, which required OkCupid to turn over information requested by the agency.
“The FTC enforces the privacy promises that companies make,” said Christopher Mufarrige, Director of the FTC’s Bureau of Consumer Protection. “We will investigate, and where appropriate, take action against companies that promise to safeguard your data but fail to follow through—even if that means we have to enforce our Civil Investigative Demands in court.”
As alleged in the complaint, OkCupid told consumers that it doesn’t share “your personal information with others except as indicated in this Privacy Policy or when we inform you and give you an opportunity to opt out of having your personal information shared.” Its privacy policy at the time claimed it may share personal information with service providers, business partners, other entities within its family of businesses or when it informed consumers about such data sharing and gave consumers the chance to opt out.
Despite these promises, the FTC alleged that OkCupid shared users’ personal data with a third party—even though it was not a service provider, business partner, or family affiliate—and did not inform consumers or give them the chance to opt out of such sharing.
Even though it did not have any business relationship with OkCupid, the third-party data recipient asked the company to share large datasets of OkCupid user photos and related data with it because OkCupid’s founders were financial investors in the third party. OkCupid provided the third party with access to nearly three million OkCupid user photos as well as location and other information without placing any formal or contractual restrictions on how the information could be used, the FTC alleged.
The FTC also alleged that, since September 2014, Match and OkCupid took extensive steps to conceal—including through trying to obstruct the FTC’s investigation—and deny that OkCupid shared users’ personal information with the data recipient. For example, when a news story revealed that the third party had obtained large OkCupid datasets, OkCupid claimed to the media and OkCupid users that it was not involved with the third party.
Underthe proposed settlement, OkCupid and Match are permanently prohibited from misrepresenting or assisting others in misrepresenting:
The Commission vote authorizing the staff to file the complaint and stipulated final order was 2-0.
The FTC filed the complaint and final order in the U.S. District Court for the Northern District of Texas, Dallas Division.
NOTE:The Commission files a complaint when it has “reason to believe” that the named defendants are violating or are about to violate the law and it appears to the Commission that a proceeding is in the public interest. Stipulated final orders have the force of law when approved and signed by the District Court judge.
The lead staff attorneys on this matter are Sarah Choi and Alejandro Rosenberg.
The Federal Trade Commission works to promote competition andprotect and educate consumers. The FTC will never demand money, make threats, tell you to transfer money, or promise you a prize. Learn more about consumer topics atconsumer.ftc.gov, or report fraud, scams, and bad business practices atReportFraud.ftc.gov. Follow theFTC on social media, readconsumer alertsand thebusiness blog, andsign up to get the latest FTC news and alerts.
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
Mathematical methods and human thought in the age of AI
Mathematics > History and Overview
Title:Mathematical methods and human thought in the age of AI
View PDF HTML (experimental)Abstract:Artificial intelligence (AI) is the name popularly given to a broad spectrum of computer tools designed to perform increasingly complex cognitive tasks, including many that used to solely be the province of humans. As these tools become exponentially sophisticated and pervasive, the justifications for their rapid development and integration into society are frequently called into question, particularly as they consume finite resources and pose existential risks to the livelihoods of those skilled individuals they appear to replace.
In this paper, we consider the rapidly evolving impact of AI to the traditional questions of philosophy
with an emphasis on its application in mathematics and on the broader real-world outcomes of its more general use. We assert that artificial intelligence is a natural evolution of human tools developed throughout history to facilitate the creation, organization, and dissemination of ideas, and argue that it is paramount that the development and application of AI remain fundamentally human-centered. With an eye toward innovating solutions to meet human needs, enhancing the human quality of life and expanding the capacity for human thought and understanding, we propose a pathway to integrating AI into our most challenging and intellectually rigorous fields to the benefit of all humankind.
Bibliographic and Citation Tools
Code, Data and Media Associated with this Article
Demos
Recommenders and Search Tools
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
I am definitely missing the pre-AI writing era
Yesterday, I wrote my first technical draft on what I was working on with the goal to share it publicly on here (well using an account dedicated to technical post), and did not realize how wanting to sound perfect actually steal the ''my voice" in the paper. Although 80 % of the content was my own writing, the fact that it was run in a LLM enginee for grammar and vocabulary cross-check, made it failed the "probable written by AI " metric; and it was rejected. I am not complaining, well, it was kinda my fault after all cause the instructions did specify the no-use of LLM, when sending first post. And the rejection made me think: why did I even need the AI to validate what I worked on?
The scary part, is I love writing, I have been writing blogs and all sort of things since I can't even remember. Have some think pieces on popular sites, and other personal blog that I used to share. And you know, before 2023, my writing was so good that I rarely had to have a second read or correction to make sure the flow stays consistent, since english is my 4th language. I always sought external review or check with tools like : Grammarly or Quilbot, for mistakes or grammars misphrasing. But now, this "superskill" has faded away, like seriously. I can't even write a 1000 words think piece, without needing or wanting to know what AI thinks or how it could have phrase it better, even for my emails. I can't write poems as before, without sounding generic, and this is mainly because I have now trained my brain to rely on these automated tools that It can not be creative anymore or think of itself. My writing has deterioted, I attempted to write a slam-poem, recently for a competition I wanted to partake in, and upon the completion when I read it, my reaction was : Who is this? Who wrote this? This is bad !
Before, I would sit and write a piece and I kid you not, the first draft was always perfect that I rarely had to write, and my best poem that got me to some "art"ies festival, was just written while I was sitting outside my hostel, with a piece of paper and a pen, and a thought that was bothering me. But now, I am slowly realizing, it has been changing in negative way.
Sorry that this sounds like a rant, well, it is one ! But like, my point was to share it out loud, and kinda know if I am the only person experiencing this or if there are others? And for people who successfully taken back their creative writing skills, how did you do it?
This post, is written without any tools assistance I just wrote what my brain is instructing to type (might not reread it before posting). Obviously, you would see some mistakes, some paraphrasing issue, some article misuse; I personally do not see them but, hey, I am not expert in literature. So bear with me, also, I think that is the beauty of writing, the raw , unedited emotions of the person behind every words either for entertainment or educational purposes, is what makes it special.
So, thank you ''LessWrong" moderators, for rejecting my first draft, it is kinda a wake up call that made me realize, maybe I just need to be focusing on my own voice, and there is nothing wrong into sounding "more or less wrong". Yeah, I know, people advocate into embracing AI as it would be/is in every aspect of our life, but maybe it is good that most of us do not let this technology rephrase our own thought. Cause those words, that the AI told you: sound wrong or aggressive or less formal, are the ones that carry the ''emotions", you ought to share.
Have a productive week, dear reader !
P.S: If you see this! I am celebrating. It means, I passed the text. Yay!
I can definitely relate to what you wrote here. I've been writing in a particular NSFW niche since 2014 across several usernames, and have generally received positive feedback on my stories. I always wrote fast and clean - my superpower was knocking out thousands of words in a single evening like a machine, and I mostly didn't need to edit it.
I was an early and enthusiastic adopter of AI as a writing companion in 2022. I loved NovelAI, and had tons of fun using various LLMs to write the "boring stuff" and writing the spicy stuff they refused to write myself. The response to my AI-assisted efforts (never disclosed) were about as positive as my 100% human stuff. But... a switch flipped in 2025, and now I'm just not having fun all around. I used to love just writing. I used to love AI-assisted writing. But lately, I can hardly get the motivation for any of it. My superpower is gone, and I still get people commenting on my unfinished human-made and ai-assisted stories, wondering when my next post is going to be.
I'm still cautiously pro-current AI, and still a little worried about the future of AI from an existential risk standpoint (probably like many here on Less Wrong), but I know for me it has kind of been a long-term poison pill in a way. I haven't cultivated that joy again, but I have been trying to de-stress and get back into AI-writing at least,
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
Recover Apple Keychain
I had the misfortune of locking myself out of my work MacBook.
You’d be justified in asking: Don’t you use it every day?
Yes, I do—and probably more than I should admit. But a few things led to this situation. For one, I’d gotten too used to relying on Touch ID. As a result, if my muscle memory failed me, I was stuck. On top of that, I kept entering my office Active Directory (AD) password instead of my Mac password—again, muscle memory at work. By the time I realized what I was doing wrong, I was locked out.
So, like every corporate employee, I reached out to IT. I’ll spare you the rant about how useful—or not—that decision was.
Using macOS recovery options, I managed to reset my laptop’s password—without fully understanding the ramifications. When I logged back in with the new password, boom: all my keychain data and some local user data were gone, and syncing had stopped.
It turns out that this data was encrypted using my old password, and the reset utility doesn’t decrypt and re-encrypt it with the new one. In hindsight, that makes sense from a security standpoint. Still, I had assumed there might be some kind of master key that would handle this automatically during a password reset.
After the reset, macOS gave me the option to resync with my old data. Unfortunately, I had already exhausted the allowed attempts to enter my old password. Every time I tried to resync, I was met with the dreaded “delete iCloud data” prompt.
As a long-time Linux (Arch Linux) tinkerer, I figured there had to be a way to either:
- Reset the failed login attempt counter, or
- Recover the data directly
But since this was Apple, I knew it wouldn’t be straightforward.
After some digging, I found that keychain data is stored at:
~/Library/Keychains
The main file is:
login.keychain-db
During the password reset, macOS had moved the original keychain to:
login_renamed_1.keychain-db
and created a fresh login.keychain-db
.
Restoring my old data turned out to be surprisingly simple: replace the new keychain file with the old one.
Steps to restore keychain data
Navigate to:
~/Library/Keychains
Remove the newly created keychain (make sure it doesn’t contain anything important):
rm login.keychain-db
Rename the old keychain back to the default:
mv login_renamed_1.keychain-db login.keychain-db
Voilà—data restored.
When I opened Keychain Access, it prompted me for my old password. After entering it, everything synced correctly and updated to use the new password going forward.
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
Show HN: Coasts – Containerized Hosts for Agents
Coasts (Containerized Hosts) is a CLI tool with a local observability UI for running multiple isolated instances of a full development environment on a single machine. It works out of the box with your current setup: no changes to your existing application code, just a small Coastfile
at your repo root. If you already use Docker Compose, Coasts can boot from your existing docker-compose.yml
; if you do not use Docker or Compose, Coasts works just as well.
Build once and run N instances with whatever volume and networking topology your project needs. Check out one coast at a time to bind canonical ports to your host, and use dynamic ports to peek into the progress of any worktree.
Coasts is agnostic to AI providers and agent harnesses. The only host requirement is Git worktrees, so you can switch tools without changing how you work and without any harness-specific environment setup.
Coasts is also offline-first with no hosted service dependency, so there is no vendor lock-in risk: even if we disappeared, your local workflow would keep running.
Install the latest public release:
eval "$(curl -fsSL https://coasts.dev/install)"
Visit coasts.dev for the website, docs, and installation instructions.
For the full user-facing documentation, see the Coasts docs.
Want a concrete example to explore? Check out the coasts-demo
repository for a small demo project you can use to try Coasts end to end.
To contribute, read the contributing guide for PR guidelines.
Note: Coasts is currently macOS-first. Linux development works, but canonical ports below
1024
require host setup beforecoast checkout
can bind them.For local HTTPS stacks that use Caddy, Coast now reuses one local Caddy root CA per Coast installation. Trust that root once and recreated workspaces under the same
COAST_HOME
keep using it.coast
andcoast-dev
stay separate because they use different Coast homes.
- Rust (stable toolchain)
- Docker
- Node.js
- socat (
brew install socat
on macOS,sudo apt install socat
on Ubuntu) - Git
Run the setup script once to build the web UI, compile the workspace, and symlink coast-dev
/ coastd-dev
into ~/.local/bin
:
./dev_setup.sh
On first run it adds ~/.local/bin
to your PATH — restart your shell or source ~/.zshrc
to pick it up.
Dev mode uses ~/.coast-dev/
and port 31416, so it never conflicts with a global coast install on port 31415.
That separation also applies to local HTTPS trust. coast-dev
reuses a stable Caddy root under ~/.coast-dev/caddy/pki/...
, while the regular install uses ~/.coast/caddy/pki/...
.
You can inspect or export the active install's root certificate with:
coast cert info
coast cert export --to ~/Downloads/coast-root.crt
Coast does not install the certificate into your OS or browser trust store automatically. Export it, then import/trust it wherever your environment needs it.
You'll want three terminals:
Terminal 1 — dev daemon:
coast-dev daemon start # start in background
# or: coastd-dev --foreground # start in foreground for log output
After Rust changes are rebuilt by make watch
, restart the daemon to pick them up:
coast-dev daemon restart
Terminal 2 — Rust rebuild on save:
make watch
This runs cargo watch
and recompiles the workspace whenever Rust source files change. After a rebuild completes, restart the daemon in Terminal 1.
Terminal 3 — web UI with hot reload:
cd coast-guard
npm install
npm run dev:coast-dev
This starts the Vite dev server on http://localhost:5173
with hot module replacement, proxying /api
requests to the dev daemon at localhost:31416
.
Use
npm run dev
(without:coast-dev
) if you're developing the UI against a production daemon on port 31415.
The Makefile is the primary entry point for development tasks:
The web UI depends on TypeScript types generated from Rust structs via ts-rs
. After changing any Rust types that are used by the UI, regenerate the bindings:
cd coast-guard
npm run generate:types
This runs cargo test -p coast-core export_bindings
and rebuilds the barrel file in src/types/generated/
.
The docs viewer in the UI reads from a generated manifest. After changing any markdown files in docs/
, regenerate it:
cd coast-guard
npm run generate:docs
Translation and search index generation are centralized Python scripts invoked via the Makefile:
make docs-status # show which docs need translation
make translate LOCALE=es # translate docs for one locale
make translate-all # translate all supported locales
make doc-search LOCALE=en # generate search index for one locale
make doc-search-all # generate search indexes for all locales
Both scripts read OPENAI_API_KEY
from the environment or from .env
in the project root. See .env.example
.
make test
Runs cargo test --workspace
across all crates.
Integration tests live in integrated-examples/
and exercise full end-to-end coast workflows. Most tests can run inside a Docker-in-Docker container via the DinD runner, which only requires Docker:
make run-dind-integration TEST=test_assign # run one test
make run-dind-integration TEST=al
▸ 展开全文
03-31 00:01 · AI,技术,HackerNews,AI
Bird brains (2023)
I was doomscrolling Reddit at 1am (as you do) and someone had posted a video from the New Zealand Transport Agency. Road workers near a tunnel by Milford Sound kept finding their traffic cones in weird places. Dragged into the road, rearranged, sometimes actively rerouting traffic. Nobody could figure out what was going on, so they checked the cameras.
Kea. Native to New Zealand, these big parrots are usually seen on the route to Milford Sound harassing tourists. A flock of them is officially called a "circus" or a "curiosity" -- whoever named them clearly met one. The footage showed them just... casually shoving cones around a construction site. But here's the insane bit -- workers said the kea wouldlisten for cars coming through the tunnelBEFORE moving the cones, timing it so the cars would have to stop. Why? Because stopped cars mean humans getting out. Humans getting out means food.
These birds are smarter than some adults I know. Move cone → car stops → human gets out → human feeds me. They independently invented toll booths.
The transport agency's solution was equally funny. They switched to heavier cones the birds couldn't move, and then -- I'm not making this up -- they built "kea gyms" by the roadside. Puzzle stations and contraptions to keep them entertained. A government agency literally built a playground for parrots because they were too smart for traffic management. Honestly, I'm fine with my tax dollars going to this.
Obviously now I had to know -- is this the smartest bird in the world? And hold on, how do you actually measure how smart a bird is? So I whipped out ChatGPT and Google Scholar and here's what I learned.
Turns out there's no single test -- researchers have come up with a bunch of different experiments over the years, each designed to measure a different type of intelligence. Some of these I'd fail too tbh.
First up, the mirror test. You stick a coloured mark on a bird somewhere it can only see in a mirror. If it looks at the mirror and then tries to remove the mark from its own body, it recognises that the reflection isitself. That's self-awareness. Most animals completely fail this -- dogs fail it, cats fail it. Eurasian magpiespass it. One of the very few non-mammals to do so. Your local magpie has a stronger sense of self than your golden retriever. Pretty humbling for the dog.
Then there's a cool one called Aesop's Fable -- my favourite. It's literally named after the fable where a thirsty crow drops stones into a pitcher to raise the water level. Researchers put food floating in a narrow tube of water that the bird can't reach. The question is whether it'll figure out to drop objects in to raise the water level and get the food. Rooks, New Caledonian crows, and Eurasian jaysall pass. Some of them even figure out that heavy objects sink (useful) while light objects float (useless). A fable from 600 BC and it turns out Aesop was just reporting the news.
Next, the delayed gratification test. The marshmallow test, but for birds. Offer an OK snack now, or a much better snack if they wait. Ravens pick the better future rewardover 70% of the time. They'll even choose atoolthey'll need later over an immediate food reward. That's more self-control than I have around a bowl of chips.
There's also vocal mimicry and communication, which goes way beyond "Polly wants a cracker." Dr. Irene Pepperberg studied an African grey parrot named Alex for 30 years. Alex could identify objects, colours, shapes, and numbers. He understood abstract concepts like "same" and "different." His vocabulary exceeded 100 words. When he died in 2007, his last words to Pepperberg were reportedly "You be good. I love you. See you tomorrow." I don't care how you define intelligence -- that one's hard to brush off.
And finally, spatial memory. Clark's nutcrackers cache up to 33,000 seeds across thousands of locations each autumn -- and remember where most of them are months later. I lose my keys in a two-bedroom apartment.
Here's the wild part. A 2016 study inPNASfound that parrots and songbirds pack roughlytwiceas many neurons into their forebrains as primate brains of the same mass. The neurons are just much smaller and more densely packed. A crow's brain weighs about 10 grams. A chimpanzee's weighs about 400 grams. And yet corvids demonstrate cognitive abilities that rival great apes: tool use, planning, and social reasoning.
Forebrain neurons vs. brain weight
A macaw's brain weighs 20 grams and has roughly the same number of forebrain neurons as a macaque monkey's brain at 70 grams. Ounce for ounce, bird brains are some of the most computationally dense organs in the animal kingdom. Calling someone a "bird brain" is honestly more of a compliment.
I do this for the love of writing — each subscriber costs me money. But please subscribe anyway and spread the word.
There's no definitive answer because different species dominate different areas. But after going through all of this, if you made me ran
▸ 展开全文
03-30 00:01 · AI,技术,HackerNews,AI
Miasma: A tool to trap AI web scrapers in an endless poison pit
AI companies continually scrape the internet at an enormous scale, swallowing up all of its contents to use as training data for their next models. If you have a public website, they are already stealing your work.
Miasma is here to help you fight back! Spin up the server and point any malicious traffic towards it. Miasma will send poisoned training data from the poison fountain alongside multiple self-referential links. It's an endless buffet of slop for the slop machines.
Miasma is very fast and has a minimal memory footprint - you should not have to waste compute resources fending off the internet's leeches.
Caution
There is inherent risk in deploying this software. Please fully read configuration and disclaimer before use.
Sample Miasma Response
Install with cargo (recommended):
cargo install miasma
Or, download a pre-built binary from releases.
Start Miasma with default configuration:
miasma
View all available configuration options:
miasma --help
Let's walk through an example of setting up a server to trap scrapers with Miasma. We'll pick /naughty-bots
as our server's path to direct scraper traffic. We'll be using Nginx as our server's reverse proxy, but the same result can be achieved with many different setups.
When we're done, scrapers will be trapped like so:
Embedding Hidden Links
Within our site, we'll include a few hidden links leading to /naughty-bots
.
<a href="/naughty-bots" style="display: none;" aria-hidden="true" tabindex="-1">
Amazing high quality data here!
</a>
The style="display: none;"
, aria-hidden="true"
, and tabindex="-1"
attributes ensure links are totally invisible to human visitors and will be ignored by screen readers and keyboard navigation. They will only be visible to scrapers.
Since our hidden links point to /naughty-bots
, we'll configure this path to proxy Miasma. Let's assume we're running Miasma on port 9855
.
We'll also set up aggressive rate limiting based on the scraper's user agent to help ensure we don't accidentally DDoS ourselves.
http {
# Reserve 8MB memory for tracking user agents
limit_req_zone $http_user_agent zone=miasma:8m rate=1r/s;
server {
location ~ ^/naughty-bots($|/.*)$ {
# Rate limit via the 'miasma' zone with no 429 delay
limit_req_status 429;
limit_req zone=miasma burst=5 nodelay;
# Proxy requests to Miasma
proxy_pass http://localhost:9855;
}
}
}
This will match all variations of the /naughty-bots
path -> /naughty-bots
, /naughty-bots/
, /naughty-bots/12345
, etc.
Lastly, we'll start Miasma and specify /naughty-bots
as the link prefix. This instructs Miasma to start links with /naughty-bots/
, which ensures scrapers are properly routed through our Nginx proxy back to Miasma.
We'll also limit the number of max in-flight connections to 50. At 50 connections, we can expect 50-60 MB peak memory usage. Note that any requests exceeding this limit will immediately receive a 429 response rather than being added to a queue.
miasma --link-prefix '/naughty-bots' -p 9855 -c 50
Let's deploy and watch as misbehaving bots greedily eat from our endless slop machine!
Be sure to protect well-behaved bots and search engines from Miasma via your robots.txt
!
User-agent: *
Disallow: /naughty-bots
Miasma can be configured via its CLI options:
Contributions are welcome! Please open an issue for bugs reports or feature requests. Primarily AI-generated contributions will be automatically rejected.
Miasma is not affiliated with the poison fountain. We have no control over its responses and cannot guarantee the safety of its contents. You should never direct users towards your Miasma location.
Miasma is not responsible for any retaliation from operators of affected scrapers. It is your responsibility to comply with applicable laws and hosting provider policies. See LICENSE (GPL-v3) for full warranty & limitation of liability details.
Cover art by @delphoxlover334
▸ 展开全文