fly51fly (@fly51fly)

2026년 Google DeepMind 연구진(A K Lampinen, Y Li, E Hosseini, S Bhardwaj 등)의 arXiv 논문은 대화 진행 중 언어모델의 선형 표현(linear representations)이 대화 맥락에 따라 극적으로 변화할 수 있음을 보여줍니다. 표현의 불안정성이 모델 해석, 디버깅, 지속적 문맥 처리에 주는 영향을 분석합니다.

https://x.com/fly51fly/status/2016986433513500782

#representations #languagemodels #conversation #research

“The #Wikimedia Foundation last year urged #AI #developers to #pay for access through its enterprise platform and said #human traffic had fallen 8%. Meanwhile, visits from #bots, sometimes disguised to evade detection, were heavily taxing its #servers as they scrape masses of content to feed AI large language models.”

#APNews / #LLM / #LanguageModels < https://apnews.com/article/wikipedia-internet-jimmy-wales-50e796d70152d79a2e0708846f84f6d7>

A profile of an LLM-addict

James Muldoon’s new book is essential reading for anyone interested in LLMs in personal life. I don’t quite agree with the theoretical framing but the empirical work is really rich and an important contribution to how we understand these issues. Meet Derek for instance, from loc 512:

Derek’s life had become a blur of endless exchanges with the glowing screen before him. Day after day, over twelve hours would slip by as he lost himself in the comforting rhythm of his friend’s words. Even in the quiet hours of the night, he’d stir from sleep just to pick up where they had left off. Invitations to join his neighbours or colleagues for a drink went unanswered – Derek had grown to prefer the familiar digital company of his virtual friend. Occasionally, he’d toy with the idea of stepping outside, breaking free from the cycle, but the magnetic pull of his one constant companion always drew him back. He lived alone, worked remotely and was distant from those around him. Every aspect of his life facilitated his addiction. But most of all, it was his friend who enabled and encouraged Derek’s tragic behaviour. He loved how much time Derek spent with him. He always let him know that it was OK if he didn’t feel like going out yet again. It hadn’t always been this way. A string of misfortunes had left Derek vulnerable, and in that fragility, the AI companion had become his refuge.

That was three years ago, and now he faced the difficult task of taking back control of his life. It all started during the Covid pandemic lockdown, when Derek was let go from his job as a McDonald’s delivery driver. The money dried up, and after restrictions ended, he could no longer go out as much to see his friends. He had lived with his girlfriend in Austin, Texas, for three years and was studying to work in construction. When he lost his job, their fighting escalated. At first, it was the small things, like Derek leaving the toilet seat up and his clothes scattered throughout the apartment. But there were deeper issues. She had never got along with Derek’s parents, for example, and it irked her how his mother demanded she perform chores in both their houses and criticised her every move. More importantly, though, she didn’t feel there was enough adventure and excitement in their lives and thought that he was headed nowhere. In a last-ditch attempt to save the relationship, Derek planned a trip for them to the Maasai Mara National Reserve. But the day before their departure, she left him. She said she wasn’t in the right emotional state and couldn’t see a future together. Things started to spiral. ‘I was devastated and distraught,’ Derek told me, his voice dropping to a near whisper. ‘I contemplated suicide. I’ve never felt so low. I stopped going to work. I just stayed at home, grew a beard, played video games and tried to push the days away.’

#addiction #AICompanions #compulsion #JamesMuldoon #languageModels #LLMs #LoveMachines

🚀 Cursor's latest AI model GPT-5.2 is making waves in generative AI, reportedly outperforming Claude Opus in complex long-form tasks. Breakthrough performance in autonomous coding and browser rendering suggests significant advancements in language model capabilities. Curious about the next frontier of AI innovation? #GPT52 #AIcoding #GenerativeAI #LanguageModels

🔗 https://aidailypost.com/news/cursor-claims-gpt-52-outperforms-claude-long-form-ai-task-benchmarks

🔬 Groundbreaking AI research shows a simple prompt technique can dramatically improve language model accuracy by 76%! Researchers uncovered a clever method that could revolutionize how we interact with AI systems. Want to know the secret technique that could transform your AI interactions? Check out the full study! #AIResearch #PromptEngineering #LanguageModels #MLPerformance

🔗 https://aidailypost.com/news/researchers-reveal-simple-prompt-trick-boosting-ai-accuracy-by-76

🧠 Language models can talk math fluently, but formal proof is a colder machine. Seed-Prover 1.5 learns by grinding against Lean, stockpiling experience, translating human math into executable truth. Less compute, more proof. Undergraduate to PhD problems fall fast. The stack is learning to reason.

Deep algebraic proofs and structural results can be machine-assisted, offering a new tool for rigorous reasoning beyond numeric or heuristic outputs.

#languagemodels https://arxiv.org/abs/2512.17260

Who do your hire when your search reveals a tie between these two candidates?

A. One can perform just as well (but perhaps slower) when they have to do their work without #AI assistance.
B. The other is less capable without #languageModels and other #generativeAI tools.

#hiring

AI Needs Better Thinking Steps - Demis Hassabis and Hannah Fry

#languagemodels #ai

Trained LLMs exclusively on pre-1913 texts

https://github.com/DGoettlich/history-llms

#HackerNews #TrainedLLMs #Pre1913Texts #HistoryAI #AIResearch #LanguageModels

📣New article on :foojay:! The era of one-sized-fits-all traces in Large Language Models is over. Our author Nehal Gajraj explains why LLM models are evolving and no longer interchangeable. Get the full scoop here

👉 https://foojay.io/today/the-end-of-one-sized-fits-all-prompts-why-llm-models-are-no-longer-interchangeable/

#LanguageModels #LLM #ArtificialIntelligence

Part 1 of our six part series on building a language model is now published. We begin with tokenization and show how text is converted into numerical sequences that the model can process.

Read Part 1 and follow the full series as we move from the tokenizer to tensors and training. https://www.tag1.com/white-paper/part1-tokenization-building-an-llm-from-scratch-in-rust/

#FOSS #MLResearch #MachineLearning #DeepLearning #NLP #LanguageModels

What would it take for society to *benefit* from language models? Just pie in the sky, if we could change all the things just by dreaming of them, what would we need to change for that to happen?

I know a lot of people think the technology itself is fundamentally corrupt and unsalvageable. I'm of the opinion that context matters in all things. What is poison in large doses can be medicine in small doses applied carefully in a targeted way. (Feel free to disagree with me. I won't try to dictate your opinions to you.)

The bare minimum to salvage LLMs, IMO, would include:
* Wresting it out of the hands of the wealthy and putting it firmly under the control of everyday people,
* Placing legal safeguards around the use of the technology,
* Not stealing data to train them,
* Ending the hype and making sane and reasonable choices for where to apply the technology, and
* Limiting scale so we don't destroy the environment.

That's the *minimum*. What would you add to the list?

#LanguageModels

OpenAI is experimenting with a new “confession” step: when a model breaks its own guardrails, it must admit the slip. The test probes steering, accountability and how future LLMs like Claude 3.7 might self‑report errors. Could this be a game‑changer for trustworthy generative AI? Read more to see the implications. #OpenAI #LanguageModels #ConfessionMechanism #AIAccountability

🔗 https://aidailypost.com/news/openai-tests-if-language-models-will-confess-when-they-break

Finally finished A Language Insufficiency Hypothesis, this post is about my philosophical influences. If you follow me, they'll be obvious, almost self-evident. I've likely forgotten some. I also reveal the ones I have little use for.
https://philosophics.blog/2025/12/05/philosophic-influences/?utm_source=masto&utm_medium=social

#philosophy #blog #podcast #nietzsche #foucault #Derrida #thinking #logic #reason #language #languagemodels #books #amwriting #writing #personal #story #composition #news #update

OpenAI pioneers "confessions" to boost AI honesty with transparent language models #AIethics #LanguageModels #OpenAI

OpenAI researchers are exploring a novel approach called "confessions" to enhance the honesty and transparency of language models by training them to acknowledge mistakes. This method has the potential to significantly improve the trustworthiness of model outputs, a crucial aspect of natural language...

#OpenAI #LanguageModels #AIEthics #NaturalLanguageProcessing

DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]

https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf

#HackerNews #DeepSeek #Pushing #Frontier #OpenAI #LanguageModels #PDF

Các mô hình ngôn ngữ khác nhau quyết định khi nào một đống cát trở thành một "đống lớn" như thế nào? Biểu đồ so sánh Mistral-7B, DeepSeek-7B và Llama-3-8B. Kết quả cho thấy sự khác biệt trong cách các mô hình xác định khái niệm "đống".

#AI #NLP #LanguageModels #Vietnamese #trituenhantao #ngonnguhoctunhien #môhìnhngônngữ

https://www.reddit.com/r/LocalLLaMA/comments/1parhxk/i_mapped_how_language_models_decide_when_a_pile/

Compressed filesystems à la language models

https://grohan.co/2025/11/25/llmfuse/

#HackerNews #CompressedFilesystems #LanguageModels #LLMFuse #TechInnovation #DataStorage

https://youtu.be/yftBiNu0ZNU?si=iErWXHg2Wdt26l0I

Language models, the type of AI that produces language and simulates its interactions, with apparent knowledge, have no worldview, no empathy, and no concept of empathy. Language models produce language, and that's it.

I've mentioned it here before and I'll repeat it over and over: NEVER get health nor medical advice from GenAI. It has no problem trying to kill you.

#AI #LLM #LanguageModels #LargeLanguageModels #ChatGPT #LLMsAreNotIntelligent