Hacker News Reader

wangiion Using LLM for confidence scoring is a bad idea

I'm surprised nobody mention ELO yet. We used it since 2023.

Java spent decades making pointer-rich identity semantics cheaper, before increasingly recognizing that many things programmers call “objects” are actually values.

We may spend a decade making persistent token-level KV state cheaper, before recognizing that much of what we call “context” should actually be transient, compressed, reconstructed, or represented in an entirely different state space.

KV cache risks turning an optimization of today's representation of history into tomorrow's semantics of memory.

wangiion Context Language Models

is kvcache a prematural optimization?

wangiion LensVLM: Compressing long context as images, expanding only relevant pages

yep, deepseek

wangiion Our position on open-weights models

Is Dario Amodei another version of Trump? https://www.youtube.com/watch?v=RDrfE9I8_hs

wangiion Late Bronze Age Collapse

yep, that's my first thought came to my mind as well. it's a great talk fueled with the passion of engineer, especially relevant in the era of LLM.

wangiion GLM 5.2 and the coming AI margin collapse

good observation! however, I'd argue it's the distribution channel and installation friction did the job in the cases you mentioned.

given how easy it's to replace LLM API in claude code, and how easy it is to write a claude code clone with itself (Fable is pretty good!), the collapse is coming.

wangiion Artificial intelligence is not conscious – Ted Chiang

Ted Chiang is a great Sci-Fi author, but his remark `GPT-3 and similar models as a "blurry JPEG of the Web"` was misleading at the best.

I share the view with others the term `consciousness` is not well-define yet, therefore his assessment is pointless. Maybe the real question to ask, if consciousness is merely an illusion at the macro level, when the observers not looking at the tech/implementation level deep enough, just like the term "intelligence" itself, might be more precisely captured by drifting amongst high dimensional manifold pitches.

wangiion I don't think AI will make your processes go faster

The idea of software developers' job is to translate specification into codes is dying. Now what I'm experimenting is to make my developers into competing project manager and developers are replaced by AI. I think the key advantages of developers, junior or senior, are the process + data oriented style of thinking, therefore way ahead of people from many other domain.

I told my developers: the day of making good money by sitting in front of computer and typing is long gone. Go to the clients' scenes and build the whole thing from scratch, with some assistant from sales and domain experts. Now you are the PM.

wangiion Xilem – An experimental Rust native UI framework

what's the difference between Xilem and GPUI? GPUI is used in zed and pretty cool!

wangiion TPUs vs. GPUs and why Google is positioned to win AI race in the long term

It's all about the right mind set at the very top level. At the beginning of the PC era, nobody would bet IBM to lose. Same in the dawn of internet, all money was on MS. so it happened to Nokia and Ericsson.

Google is a giant without a direction. The ads money is so good that it just doesn't have the gut to leave it on the table.

wangiion Spatial intelligence is AI’s next frontier

she's done pretty important work but since then obsessed with the vague term `spatial intelligence`. what does it mean? there isn't a clear definition in the piece. it seems very intuitive & fundamental but tbh not *rigorous*, nor insightful.

I bet it's a dead end.

wangiion Raised by Wolves Is Original Sci-Fi at Its Most Polarizing (2020)

I have only one problem w/ the S1: the creators just not have the guts to kill Mother or Father. Don't get me wrong, I like them but it could make the whole series more distant and cold.

on S2: a total disaster.

wangiion As AI Gets Smarter, It Acts More Evil

I expected to read some nash equilibrium like insight to demonstrate the inevitability of machines turning evil, but got a few headlines.

wangiion Show HN: CXXStateTree – A modern C++ library for hierarchical state machines

how is it better than https://github.com/boost-ext/sml ?

there are about 1 million c++ state machines, and sml happens to be the best, or one of them. how does yours differentiate?

wangiion Urtext: The Python plaintext library for people who've tried everything else

could anyone give one sentence pitch why it's better than markdown files?

wangiion The Friendship Recession: The Lost Art of Connecting

Maybe it's b/c vast majority not doing well financially?!

wangiion LLMs can see and hear without any training

Exactly! There is definitely something wrong with FAIR.

wangiion Google is making AI in Gmail and Docs free, but raising the price of Workspace

it feels like google in panic mode, the only thing it can think of is to put a chatbot everywhere, just b/c it can. I don't see a value proposition at all.

wangiion OpenAI O3 breakthrough high score on ARC-AGI-PUB

Interesting formulation! it captures the intuition of the "smartness" when solving a problem. However, what about asking good questions or proposing conjectures?

wangiion Clay – UI Layout Library

Looks impressive! What are the connection/comparison with imgui though?

wangiion Unit is a general purpose visual programming system

The flow based visual programming has always being challenging field, and the problem remains: what are the values added comparing with the text based coding?

I really want to love it, please give me a strong reason.

wangiion Consistency LLM: converting LLMs to parallel decoders accelerates inference 3.5x

are you practicing any meditation? it's regarded as "awaken" state in some practice! if you have any method, please share with me! thanks!

wangiion Consistency LLM: converting LLMs to parallel decoders accelerates inference 3.5x

According to the original Jacobi decoding paper, it's set in the machine translation tasks, with encoder + decoder, in which parallel algo applied only to the decoder part.

wangiion Consistency LLM: converting LLMs to parallel decoders accelerates inference 3.5x

I feel it's a pretty dangerous optimization before we REALLY understand what's going on inside of the LLM. e.g. guys believe in the geometric interpretation will have something to say, and it would probably hurt if you are using "filler" tokens.

Besides, the assumption (not a universal fact) that "forming complete sentences in mind before articulating word by word" seems overly simplifies activities happens in our mind: do we really have a complete planning before start talking/typing? as a Buddhist I lean towards it's an illusion. further more, what about simultaneous thoughts? are we linear thinker in the sentence level?

anyway, pretty neat math!

wangiion An infinite canvas for code exploration

not able to sign in with github:

Unable to process request due to missing initial state. This may happen if browser sessionStorage is inaccessible or accidentally cleared. Some specific scenarios are - 1) Using IDP-Initiated SAML SSO. 2) Using signInWithRedirect in a storage-partitioned browser environment.

wangiion Better and Faster Large Language Models via Multi-Token Prediction

the problem is then the total number of computation drops dramatically therefore leads to much less "thinking" power. i think the idea originated from an understanding that when we write/speak, we have an overall idea. my current hypothesis is it's probably an illusion.

you may want to search for "filler" papers to read.

wangiion The language you speak changes your perception of time

I believe it's a co-evolving relationship, and it's extremely difficult to track the first/root cause.

wangiion The Triumph and Terror of Wang Huning

People make same kind of mistakes all the time, instead of judging the contents, we look down to things that are "mostly repackaging" what already exist. Brand new things are extremely rare, mostly just remix of things we've seen everyday.