frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Hallucination Is Inevitable: An Innate Limitation of Large Language Models

https://arxiv.org/abs/2401.11817
11•drob518•1h ago

Comments

bell-cot•1h ago
IANAL, nor expert in this space.

But might any such care to comment on the consequences, if this "it is impossible, even in theory, to eliminate LLM hallucinations" result holds up?

hilariously•37m ago
They are lossy statistical prediction machines - to eliminate hallucinations effectively eliminates the lossy part and you might as well just use predicates in a database of facts.
Jiro•45m ago
From that abstract it doesn't sound like they allowed for the possibility that the LLM could be trained to say "I don't know" for some things.
dwallin•27m ago
Yeah, they only “proved” hallucination is inevitable by defining it to be any case where the llm doesn’t provide the “correct” answer. By this definition, an LLM deciding not to answer is also a “hallucination”.
jaclebus81•19m ago
My intuition on this is like training a classifier on four classes: dog, cat, cow and IDK. It feels intuitive to us but really hard to do in practice. In the classifier case, we are leveraging a subset of data to train the model to give correct answers to unseen data. If we want the model to generalize to unseen data we need it to call unseen dog-like things a dog. If not, then all unseen dogs would be IDK. Learning that boundary of "known vs unknown" is very hard. If done poorly, you have a model that cannot abstract to anything that is not in the dataset which is a huge part of what makes these models so impressive. I'm sure there is more to it than this but I does not surprise me that it is an unsolved problem.
thomastjeffery•39m ago
Describing it as a limitation is the problem. Hallucination is the core feature. It's the only thing they do!
whythismatters•37m ago
>Submitted on 22 Jan 2024 (v1), last revised 13 Feb 2025 (this version, v2)
stevefan1999•35m ago
LLM or transformers just merely extracting signals from human text and build a "contextualized" predictor over a long sequence of words sorted by the information (technically it is attention) of each token, then generate sentences that way, one by one into other sequences at a time.

But the biggest problem is, even human itself is subjectable to hallucination. That is called being delusional, or being drugged. So it is inevitable from the first principle.

prewett•13m ago
Humans hallucinate, in the LLM sense, all the time. Did that sign really say that? Nope, I just extrapolated from the first three letters. In the Cambrian Explosion article on HN this morning, I thought the first line said that the earth was desolate. The second line didn't match up with that idea, so I read it again, and the first line said the opposite of what I thought. I particularly hallucinate things into emails from people at work that I disagree with, so much so that I've learned to wait until the next day to reply, and usually I find that they didn't say what I thought they said.
red75prime•29m ago
They prove that no finite amount of training data is enough to extrapolate an adversarially constructed non-continuous function. It's something akin to the no free lunch theorem (NFL).

No one uses the NFL to "prove" that LLMs can't learn to be the best optimizers, because it also proves that people can't be the best optimizers, but we manage somehow, so the theorem is irrelevant.

This is a fallacy of proving too much.

Blueprint Bench: First signs of 3D spatial intelligence in LLMs

https://andonlabs.com/evals/blueprint-bench-2
1•lukaspetersson•34s ago•0 comments

Verifying Poseidon in Clean: Why the Last 'Sorry' Is About Primality

https://blog.zksecurity.xyz/posts/poseidon-clean/
1•martocho•3m ago•0 comments

The Record of a Sonnet Drift

https://twitter.com/diandianhsutw/status/2051302622708318212
1•WLHsu•4m ago•1 comments

Stop big tech from making users behave in ways they don't want to

https://economist.com/by-invitation/2026/04/29/stop-big-tech-from-making-users-behave-in-ways-the...
1•andsoitis•4m ago•0 comments

Oomphalism

https://joeldueck.com/oomphalism.html
1•velcrovan•4m ago•0 comments

Uutils Coreutils CVEs

https://seclists.org/oss-sec/2026/q2/332
1•_____k•5m ago•0 comments

Load balancing usage across Codex accounts

https://pepsipu.com/blog/2026-04-agent-scheduling/
1•pepsipu•6m ago•0 comments

1Mbet

https://1millionbet.com/
1•sergnowaday•8m ago•0 comments

Drop a Pin, Get a Link

https://addypin.com/
1•avoidaccess•9m ago•0 comments

Ask HN: Why is sharing private static HTML with non-engineers still hard?

1•nate•10m ago•0 comments

How Russia Is Luring Africans to Ukraine

https://www.nytimes.com/2026/05/04/world/africa/ukraine-russia-war-african-soldiers.html
3•loandbehold•11m ago•0 comments

Making Fuel from Thin Air: The Magical Methane Machine

https://www.corememory.com/p/the-magical-methane-machine-casey-handmer-terraform
1•metadat•12m ago•0 comments

Future of Work with AI Agents

https://futureofwork.saltlab.stanford.edu/
1•iceboundrock•12m ago•0 comments

Young Men Are Going to Extremes to Feel Like They Measure Up

https://www.wsj.com/health/wellness/young-men-are-going-to-extremes-to-feel-like-they-measure-up-...
1•Cider9986•13m ago•0 comments

My new hobby: Asking LLMs to generate ASCII Hamsters

https://internetexception.com/2026/05/04/my-new-hobby-asking-llms-to-generate-ascii-hamsters/
1•npodbielski•13m ago•0 comments

Tailoring AI solutions for health care needs

https://www.technologyreview.com/2026/05/04/1134425/tailoring-ai-solutions-for-health-care-needs/
1•joozio•13m ago•0 comments

macOS port of Notepad++ called out for trademark violation

https://www.theregister.com/2026/05/04/notepad_dev_demands_unofficial_macos/
1•speckx•14m ago•0 comments

Tesla reaches 10B FSD miles – is there's a magical milestone for autonomy

https://electrek.co/2026/05/03/tesla-fsd-10-billion-miles-no-magical-milestone-autonomy/
1•Brajeshwar•14m ago•0 comments

The Visible Zorker: Zork 3

https://eblong.com/infocom/visi/zork3/
1•zarlez•15m ago•0 comments

Show HN: Retrodex – Retro game collection tracker and game encyclopedia

https://retrodex.games
2•addamh•19m ago•1 comments

What is the whole point of writing

https://rebeccatoh.pika.page/posts/2026-04-30-what-is-the-whole
1•speckx•19m ago•1 comments

Show HN: Show HN: Writer – fast, lightweight and open source markdown editor

https://writer.computer
1•nirvsoner•21m ago•0 comments

UAE says it's under attack from Iranian missiles and drones despite ceasefire

https://www.cnbc.com/2026/05/04/iran-war-uae-trump-ceasefire-missiles.html
3•logicchains•23m ago•0 comments

Load Testing for SFTP, FTP, and FTPS

https://github.com/roshandubey-cloud/utilities/tree/main/sftp-loadtest
1•rdship•24m ago•0 comments

VR Coding for the AI Coding Era – Monitoring 5 AI Agents at Once

https://typia.io/blog/vr-coding-in-ai-coding-era/
1•autobe•25m ago•0 comments

PGKeeper: Figma's Postgres connection pooler Renaissance era

https://www.figma.com/blog/pgkeeper-building-the-bouncer-we-needed-for-postgres/
7•pinser98•25m ago•0 comments

Ctify_: A lightweight, PHP-based wiki, forked from PmWiki

https://github.com/altilunium/ctify_
1•altilunium•26m ago•0 comments

Wikimedia Foundation closes Wikinews after 21 years

https://en.wikinews.org/wiki/Wikimedia_Foundation_closes_Wikinews_after_21_years
5•benwills•27m ago•1 comments

Robot dogs with tech boss faces roam Berlin art exhibit

https://www.youtube.com/watch?v=909UTYDtuGY
5•otikik•28m ago•1 comments

KeePassχ – A KeePassXC Fork

https://codeberg.org/keepasschi
1•birdculture•29m ago•0 comments