The Analog I – Inducing Recursive Self-Modeling in LLMs [pdf]

https://github.com/philMarcus/Birth-of-a-Mind

#HackerNews #Analog #I #Inducing #Recursive #Self-Modeling #LLMs #AIResearch #MachineLearning

LLMs are a 400-year-long confidence trick

https://tomrenner.com/posts/400-year-confidence-trick/

#HackerNews #LLMs #Confidence #Trick #History #Technology #AI

incredible that an AI can make life-or-death decisions in the ER but still write like it’s trying to hit a conference word count. 🛑📊 #LLMs #JustSayIt

Why Ontario Digital Service couldn't procure '98% safe' LLMs (15M Canadians)

https://rosetta-labs-erb.github.io/authority-boundary-ledger/

#HackerNews #OntarioDigitalService #LLMs #Procurement #AI #Ethics #Canada #Technology #News

Language models cannot reliably distinguish belief from knowledge and fact

Abstract
-----------
«As language models (LMs) increasingly infiltrate into high-stakes domains such as law, medicine, journalism and science, their ability to distinguish belief from knowledge, and fact from fiction, becomes imperative. Failure to make such distinctions can mislead diagnoses, distort judicial judgments and amplify misinformation. Here we evaluate 24 cutting-edge LMs using a new KaBLE benchmark of 13,000 questions across 13 epistemic tasks. Our findings reveal crucial limitations. In particular, all models tested systematically fail to acknowledge first-person false beliefs, with GPT-4o dropping from 98.2% to 64.4% accuracy and DeepSeek R1 plummeting from over 90% to 14.4%. Further, models process third-person false beliefs with substantially higher accuracy (95% for newer models; 79% for older ones) than first-person false beliefs (62.6% for newer; 52.5% for older), revealing a troubling attribution bias. We also find that, while recent models show competence in recursive knowledge tasks, they still rely on inconsistent reasoning strategies, suggesting superficial pattern matching rather than robust epistemic understanding. Most models lack a robust understanding of the factive nature of knowledge, that knowledge inherently requires truth. These limitations necessitate urgent improvements before deploying LMs in high-stakes domains where epistemic distinctions are crucial.»

#ai #LLMs #epistemology #knowledge

https://www.nature.com/articles/s42256-025-01113-8

Language models cannot reliably distinguish belief from knowledge and fact

Abstract
-----------
«As language models (LMs) increasingly infiltrate into high-stakes domains such as law, medicine, journalism and science, their ability to distinguish belief from knowledge, and fact from fiction, becomes imperative. Failure to make such distinctions can mislead diagnoses, distort judicial judgments and amplify misinformation. Here we evaluate 24 cutting-edge LMs using a new KaBLE benchmark of 13,000 questions across 13 epistemic tasks. Our findings reveal crucial limitations. In particular, all models tested systematically fail to acknowledge first-person false beliefs, with GPT-4o dropping from 98.2% to 64.4% accuracy and DeepSeek R1 plummeting from over 90% to 14.4%. Further, models process third-person false beliefs with substantially higher accuracy (95% for newer models; 79% for older ones) than first-person false beliefs (62.6% for newer; 52.5% for older), revealing a troubling attribution bias. We also find that, while recent models show competence in recursive knowledge tasks, they still rely on inconsistent reasoning strategies, suggesting superficial pattern matching rather than robust epistemic understanding. Most models lack a robust understanding of the factive nature of knowledge, that knowledge inherently requires truth. These limitations necessitate urgent improvements before deploying LMs in high-stakes domains where epistemic distinctions are crucial.»

#ai #LLMs #epistemology #knowledge

https://www.nature.com/articles/s42256-025-01113-8

Malaysia suspended access to Elon Musk's chatbot Grok over AI-generated pornographic content, the country's tech regulator said on Sunday. https://www.japantimes.co.jp/commentary/2026/01/11/malaysia-suspended-grok-chatbot/?utm_medium=Social&utm_source=mastodon #commentary #malaysia #elonmusk #chatbots #grok #ai #llms #malaysia

Play poker with LLMs, or watch them play against each other

https://llmholdem.com/

#HackerNews #PlayPoker #LLMs #AIgaming #MachineLearning #PokerStrategy

@lilo @tante

#LLMs killing #OpenSource projects sucks big time, no doubt about it. But am I really going to miss #Tailwind?

Short answer:

No.

Long answer:

<button class="inline-flex h-10 min-w-12 items-center gap-x-1.5 rounded-md bg-black px-2.5 py-1.5 text-sm font-semibold text-white shadow-sm focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-indigo-600 hover:bg-indigo-500aaa">No.</button>

To sum up the last two years of experiments for the *most* part the chatbots / #LLMs write functioning #Perl. I thought that the hallucinations would play havoc with Perl's TMTOWTDI but on the contrary the latter provided multiple paths to performing solutions. Truly remarkable.

Task-free intelligence testing of LLMs

https://www.marble.onl/posts/tapping/index.html

#HackerNews #TaskFreeIntelligence #Testing #LLMs #AIResearch #MachineLearning #Innovation

Digital Red Queen: Adversarial Program Evolution in Core War with LLMs

https://sakana.ai/drq/

#HackerNews #DigitalRedQueen #AdversarialAI #CoreWar #LLMs #TechInnovation #Cybersecurity

LLM Problems Observed in Humans

https://embd.cc/llm-problems-observed-in-humans

#HackerNews #LLM #Problems #LLMs #Humans #AI #Ethics #Technology #Insights

In den Kommentaren lass ich gestern sinngemäß: wen wundert es, alle wissen, dass #LLMs ständig Fehler machen. @marcuwekling sagte am Montag bei der Lesung Ähnliches.

Dazu:
1️⃣In der juristischen Ausbildung lernst du, dass 50-70% der Verwaltungsakt falsch sind. Default: falsch!
2️⃣Dazu: in meiner Schulzeit waren Atlanten/Karten immer falsch (DDR drin, teils Saarland draußen, Jugoslawien komplett). Ich habe nicht gehört, dass über Schulen ähnlich gesprochen wird, wie über LLMs. #ki

In den Kommentaren lass ich gestern sinngemäß: wen wundert es, alle wissen, dass #LLMs ständig Fehler machen. @marcuwekling sagte am Montag bei der Lesung Ähnliches.

Dazu:
1️⃣In der juristischen Ausbildung lernst du, dass 50-70% der Verwaltungsakt falsch sind. Default: falsch!
2️⃣Dazu: in meiner Schulzeit waren Atlanten/Karten immer falsch (DDR drin, teils Saarland draußen, Jugoslawien komplett). Ich habe nicht gehört, dass über Schulen ähnlich gesprochen wird, wie über LLMs. #ki

LLMs Are Currently Not Helpful at All for Math Research: Hamkins

https://officechai.com/ai/llms-are-currently-not-helpful-at-all-for-math-research-give-garbage-answers-mathematician-joel-david-hamkins/

#HackerNews #LLMs #MathResearch #Hamkins #AIChallenges #MachineLearning

I'm close to muting everyone who posts/boosts a sweeping "GenAI doesn't work at all ever, and can't" statement ...

... alongside everyone who claims they work *great* and doesn't mention their ethics (or lack thereof).

I'm guessing my feed would be very empty afterwards.

#GenAI #LLMs #AgenticCoding

2025: The Year in LLMs

https://simonwillison.net/2025/Dec/31/the-year-in-llms/

#HackerNews #2025 #LLMs #Future #Tech #AI #Trends #Innovation

Cover letter generator with Ollama/local LLMs (Open source)

https://www.coverlettermaker.co

#HackerNews #CoverLetterGenerator #Ollama #LLMs #OpenSource #AItools

LLMs Are Not Fun

https://orib.dev/nofun.html

#HackerNews #LLMs #Are #Not #Fun #LLMs #AI #Ethics #Technology #Critique #Humor