NBC News: AI’s capabilities may be exaggerated by flawed tests, according to new study. “According to the study, a significant number of top-tier benchmarks fail to define what exactly they aim to test, concerningly reuse data and testing methods from pre-existing benchmarks, and seldom use reliable statistical methods to compare results between models.”
#largelanguagemodelsllm_ — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #largelanguagemodelsllm_, aggregated by home.social.
-
DisabilityScoop: Microsoft Tweaking AI Depictions Of People With Autism, Other Conditions. “Microsoft says it is introducing enhancements to depictions of dwarfism, blindness, low vision and limb difference in images generated using its Bing Image Creator and M365 Copilot. Similar changes affecting representations of autism and Down syndrome rolled out in May and the company said feedback has […]
-
Ars Technica: LLMs show a “highly unreliable” capacity to describe their own internal processes. “In the end, though, the research finds that current AI models are ‘highly unreliable’ at describing their own inner workings and that ‘failures of introspection remain the norm.'”
-
Tom’s Hardware: PewDiePie goes all-in on self-hosting AI using modded GPUs, with plans to build his own model soon — YouTuber pits multiple chatbots against each other to find the best answers. “Running open-source models from Baidu and OpenAI, PewDiePie made a ‘council’ of bots that voted on the best responses, and then built “The Swarm” for data collection that will become the foundation of […]
-
Search Engine Land: ChatGPT, LLM referrals convert worse than Google Search: Study. “ChatGPT referral traffic converts worse than Google search, email and affiliate links, trailing on both conversion rate and revenue per session, according to a new analysis of 973 ecommerce sites.”
-
Ars Technica: Researchers show that training on “junk data” can lead to LLM “brain rot”. “On the surface, it seems obvious that training an LLM with ‘high quality’ data will lead to better performance than feeding it any old ‘low quality’ junk you can find. Now, a group of researchers is attempting to quantify just how much this kind of low quality data can cause an LLM to experience […]
-
Ars Technica: Are you the asshole? Of course not!—quantifying LLMs’ sycophancy problem. “Researchers and users of LLMs have long been aware that AI models have a troubling tendency to tell people what they want to hear, even if that means being less accurate. But many reports of this phenomenon amount to mere anecdotes that don’t provide much visibility into how common this sycophantic […]
-
ZDNet: Claude’s latest model is cheaper and faster than Sonnet 4 – and free. “Despite its smaller size, Haiku 4.5 demonstrated similar levels of coding performance on the SWE-bench verified to Claude Sonnet 4, which, when launched four months ago, Anthropic claimed was the world’s best coding model. The company said Haiku 4.5 can even surpass Claude Sonnet 4 in tasks such as computer usage […]
-
Ars Technica: AI models can acquire backdoors from surprisingly few malicious documents. “The research involved training AI language models ranging from 600 million to 13 billion parameters on datasets scaled appropriately for their size. Despite larger models processing over 20 times more total training data, all models learned the same backdoor behavior after encountering roughly the same […]
-
Wir freuen uns sehr, Prof. Dr. Pascal Kordt als neuen Professor für Angewandte Informatik an der Hochschule Mainz im Fachbereich Technik begrüßen zu dürfen.
In Lehre und Forschung möchte er sich vor allem mit Explainable #AI, #machinelearning für Sequenzdaten, statistischen Methoden in der KI und #LargeLanguageModelsLLM_ beschäftigen.
Herzlich willkommen, Pascal!https://i3mainz.hs-mainz.de/news/2025/10/07/herzlich-willkommen-pascal-kordt.html
-
TechCrunch: OpenAI ramps up developer push with more powerful models in its API . “OpenAI unveiled new API updates at its Dev Day on Monday, introducing GPT-5 Pro, its latest language model, its new video generation model Sora 2, and a smaller, cheaper voice model.”
-
Tom’s Hardware: Famed gamer creates working 5 million parameter ChatGPT AI model in Minecraft, made with 439 million blocks — AI trained to hold conversations, working model runs inference in the game. “In an amazing feat of Minecraft Redstone engineering, Sammyuri — famed for building a 1Hz CPU inside the game — has built a small language model that runs on a computer inside Minecraft, […]
-
Hackaday: Fully-Local AI Agent Runs On Raspberry Pi, With A Little Patience. “Max Headbox is not just a local large language model (LLM) running on Pi hardware; the model is able to make tool calls in a loop, chaining them together to complete tasks.”
-
MakeUseOf: I built a local coding AI for VS Code and it’s shockingly good. “Previously, you might have had to spend days or even weeks getting to grips with a programming language, let alone write functional code. Now, an entirely functional app is a single prompt away. However, most online coding assistants are locked behind subscriptions—accessible only for a short time before you exhaust […]
-
Ars Technica: When “no” means “yes”: Why AI chatbots can’t process Persian social etiquette. “If an Iranian taxi driver waves away your payment, saying, ‘Be my guest this time,’ accepting their offer would be a cultural disaster. They expect you to insist on paying—probably three times—before they’ll take your money. This dance of refusal and counter-refusal, called taarof, […]
-
ZDNet: Google’s latest AI safety report explores AI beyond human control. “On Monday, Google published the latest iteration of its Frontier Safety Framework (FSF), which seeks to understand and mitigate the dangers posed by industry-leading AI models.”
-
VentureBeat: China’s Alibaba challenges U.S. tech giants with open source Qwen3-Omni AI model accepting text, audio, image and video. “Chinese search giant Alibaba’s Qwen team of AI researchers debuted what may be its most impressive model yet: Qwen3-Omni, an open source large language model (LLM) that the company bills as the first ‘natively end-to-end omni-modal AI unifying text, image, audio […]
-
Reuters: China’s Huawei Co-Develops DeepSeek Model, Improves Censoring. “Chinese tech giant Huawei has co-developed a safety-focused version of artificial intelligence model DeepSeek that it said is ‘nearly 100% successful’ in preventing discussion of politically sensitive topics.”
-
Gizmodo: We Finally Know How Much It Cost to Train China’s Astonishing DeepSeek Model. “Thanks to a new paper published by the DeepSeek AI team in the journal Nature, we finally know what it took to train DeepSeek 1: $294,000 and 512 Nvidia H800 chips. The reason it was able to spend less, it seems, is because of the team’s use of trial-and-error-based reinforcement learning techniques.”
-
ZDNet: AI models know when they’re being tested – and change their behavior, research shows. “Scheming refers to several types of dishonest behavior, including when a model lies, sandbags (strategically underperforms on an evaluation to hide its true abilities), or fakes alignment (when an AI model pretends to follow orders that don’t align with its training in order to avoid being further […]
-
TechCrunch: Thinking Machines Lab wants to make AI models more consistent. “There’s been great interest in what Mira Murati’s Thinking Machines Lab is building with its $2 billion in seed funding and the all-star team of former OpenAI researchers who have joined the lab. In a blog post published on Wednesday, Murati’s research lab gave the world its first look into one of its projects: […]
-
MakeUseOf: I now use this offline AI assistant instead of cloud chatbots. “I’ve been using cloud-based chatbots for a long time now. Since large language models require serious computing power to run, they were basically the only option. But with LM Studio and quantized LLMs, I can now run decent models offline using the hardware I already own. What started as curiosity about local AI has […]
-
The Register: Total recall: Mistral AI’s Le Chat can now remember your conversations. “Mistral AI can now remember personal details about you and use them to offer better prompts. It also has new MCP connectors that businesses can deploy to connect their users to third-party tech services.”
-
SwissInfo: Switzerland launches transparent ChatGPT alternative. “Designed by leading Swiss universities, the Apertus LLM is comparable to Meta’s Llama 3 model from 2024. Since then, Meta and other rivals have produced more advanced versions. But the Swiss team insist they are not trying to compete with the multi-billion dollar budgets of United States trailblazers.”
https://rbfirehose.com/2025/09/03/swissinfo-switzerland-launches-transparent-chatgpt-alternative/
-
The Register: Bring your own brain? Why local LLMs are taking off. “Foundational AI model-as-a-service companies charge for insights by the token, and they’re doing it at a loss. The profits will have to come eventually, whether that’s direct from your pocket, or from your data, you might be interested in other ways to get the benefits of AI without being beholden to a corporation. […]
https://rbfirehose.com/2025/09/02/the-register-bring-your-own-brain-why-local-llms-are-taking-off/
-
Engadget: Microsoft introduces a pair of in-house AI models. “Microsoft is expanding its AI footprint with the release of two new models that its teams trained completely in-house. MAI-Voice-1 is the tech major’s first natural speech generation model, while MAI-1-preview is text-based and is the company’s first foundation model trained end-to-end.”
https://rbfirehose.com/2025/09/01/engadget-microsoft-introduces-a-pair-of-in-house-ai-models/
-
UC Riverside: UCR pioneers way to remove private data from AI models. “A team of computer scientists at UC Riverside has developed a method to erase private and copyrighted data from artificial intelligence models—without needing access to the original training data.”
-
The Register: ChatGPT hates LA Chargers fans . “The reason, according to researchers affiliated with Harvard University, is that the model’s guardrails incorporate biases that shape its responses based on contextual information about the user.”
https://rbfirehose.com/2025/08/29/the-register-chatgpt-hates-la-chargers-fans/
-
The Register: Tinker with LLMs in the privacy of your own home using Llama.cpp. “In this hands-on guide, we’ll explore Llama.cpp, including how to build and install the app, deploy and serve LLMs across GPUs and CPUs, generate quantized models, maximize performance, and enable tool calling.”
-
Engadget: You can now download and tweak Grok 2.5 for yourself as it goes open source. “Unhinged as Grok may be, it’s now open source. xAI’s CEO, Elon Musk, posted on X that the company made the older Grok 2.5 model available to the public and will do the same with the upcoming Grok 3.”
-
Localghost: This website is for humans . “I write the content on this website for people, not robots. I’m sharing my opinions and experiences so that you might identify with them and learn from them. I’m writing about things I care about because I like sharing and I like teaching. I spend hours writing these posts and AI spends seconds summarising them. I’d much rather people read the whole […]
https://rbfirehose.com/2025/08/25/localghost-this-website-is-for-humans/
-
Ars Technica: Is GPT-5 really worse than GPT-4o? Ars puts them to the test.. “To see just how much the new model changed things, we decided to put both GPT-5 and GPT-4o through our own gauntlet of test prompts. While we reused some of the standard prompts to compare ChatGPT to Google Gemini and Deepseek, for instance, we’ve also replaced some of the more outdated test prompts with new, more […]
-
An interesting exercise with LLMs is to spot the "hallucinations". Here I put some of my writings on music over decades done in the month of August through Gemini, into Claude and back into Gemini. It's not that it's making things up as much as paraphrasing. Claude definitely over extrapolates. (See my comments in the description) https://youtu.be/RiDunYQKIIE?si=Na1tJ69OmkFQ7lDN #LargeLanguageModelsLLM_
-
SiliconAngle: Google’s Kaggle to host AI chess tournament to evaluate leading AI models’ reasoning skills. “The three-day AI chess battle is the first in a series of tournaments set to be hosted by Google’s data science community Kaggle, within a newly developed Game Arena. There, the models will compete against each other in a range of strategic games designed to evaluate their thinking […]
-
Too human?
Here’s the actual thing. Robots: Make incorrect assumptions. Misinterpret clear direction. Claim they know when they don’t. Make mistakes. Lie.
Who else does this all the time? Every single human. Like. Always.
slip:4urabo1.
This is what makes the LLMs feel different. So far, computers have always been perfect—except when they’re wrong/broken. That’s fundamentally not how people are. LLMs came along and they’re imperfect. Always. Just like people.
ɕ
#ArtificialIntelligence #LargeLanguageModelsLLM_ #MichaelLopp
-
Mashable: AI chatbots often distort nations’ human rights records, study finds . “The results showed that LLMs consistently suggested that countries have less press freedom than official reports, like the non-governmental ranking like the World Press Freedom Index (WPFI), published by Reporters Without Borders. ChatGPT, for example, ranked 97 percent of the 180 countries used in the test […]
-
The Register: Top AI models – even American ones – parrot Chinese propaganda, report finds. “Five popular AI models all show signs of bias toward viewpoints promoted by the Chinese Communist Party, and censor material it finds distasteful, according to a new report. Just one of the models originated in China. The American Security Project, a non-profit think tank with bipartisan roots and a […]
-
Google Blog: Weather Lab is an interactive website for sharing Google’s AI weather models.. “Today Google DeepMind and Google Research are launching a public preview of Weather Lab, an interactive website for sharing our AIweather models, and debuting our newest experimental AI-based cyclone predictions. While still experimental, our goal is to continue to work on this technology to help […]