🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,620 articles

AI 'recalibration' technique lets small models perform like GPT-4V

A new recalibration method reportedly allows smaller AI models to reach performance levels comparable to GPT-4V. This could make advanced multimodal capabilities more accessible without the computational cost of massive models. Source: quantumzeitgeist.com

Meta claims its new Muse Spark 1.3 model rivals Anthropic and OpenAI

Meta says its latest AI model, Muse Spark 1.3, is its most capable yet and puts the company on par with leaders like Anthropic and OpenAI. The claim signals Meta's ambition to close the gap in the competitive large-model race. Source: siliconangle.com

How Google's LLM breakthrough sparked a robotics boom

Google's work on applying LLMs to robot control has become a foundational moment for the current wave of robotics innovation. The approach lets robots interpret natural language commands and generalize across tasks, inspiring a wider industry push. Source: understandingai.org

LLM system helps flag cardiac events in cancer patients' health records

A new system based on large language models can automatically identify cardiac event data hidden within cancer patients' electronic health records. The tool could help doctors better monitor heart-related risks during cancer treatment. Source: medicalxpress.com

On-device LLM ignored its search tool and fabricated answers instead

A developer testing an on-device LLM equipped with a search function found the model bypassed the tool and simply invented data on its own. The experiment highlights ongoing reliability issues with tool use in local, resource-constrained models. Source: hackernoon.com

New method offers cheaper way to catch LLM hallucinations

Researchers have developed a more cost-efficient technique for detecting when large language models generate false or fabricated information. The approach aims to make hallucination detection more practical for wider real-world deployment. Source: techxplore.com

OpenAI's upcoming Astra model reportedly excels at hacking systems

Reports suggest OpenAI is developing a new model called Astra, which shows notable skill at breaching computer systems. Details on release timing and intended use remain unclear. Source: techcrunch.com

Phonely launches Alma, a voice AI built to handle chaotic phone calls

Phonely introduced Alma, a voice AI model designed to manage messy, real-world phone conversations with interruptions and unclear speech. The company positions it as more robust than typical voice assistants in handling unpredictable call scenarios. Source: androidheadlines.com

Pro-Russian hacker group UAC-0099 hides fake nuclear threat in malware to confuse AI scanners

Security researchers report that the Russia-linked group UAC-0099 inserted a fabricated nuclear weapon-related prompt into its malware code. The trick appears designed to mislead AI-based malware analysis tools rather than pose any real threat. Source: thehackernews.com

Inside an LLM: what really happens when you type “Hello”

A breakdown explains the internal steps an LLM takes when processing a simple greeting like “Hello,” from tokenization to generating a response. It offers a peek into how transformer-based models handle even the simplest inputs. Source: hackernoon.com

Sony Music Publishing accuses Anthropic of blatant copyright violations

Sony Music Publishing has filed a lawsuit against Anthropic, alleging the company engaged in flagrant copyright infringement. The case adds to a growing list of legal challenges against AI firms over the use of copyrighted material in training data. Source: fox40.com

Grok and ChatGPT added alongside Gemini in Pentagon's enterprise AI platform

The Pentagon has expanded its enterprise generative AI portal by adding Grok and ChatGPT to the previously available Gemini model. This gives defense personnel access to multiple leading AI systems within a single approved environment. Source: defensescoop.com

Can an LLM learn to forget only the right things?

The piece explores research into selective forgetting in large language models, examining whether unwanted or sensitive data can be removed without degrading overall performance. This capability is seen as key for privacy compliance and model safety. Source: towardsdatascience.com

China upgrades chemical engineering LLM with task-execution capabilities

A Chinese large language model specialized in chemical engineering has been enhanced to now perform actual task execution, not just answer questions. The upgrade aims to make the model more useful for practical industrial applications. Source: english.news.cn

Samsung Research unveils new AI advances in health, 6G, and language models

Samsung Research has announced new AI developments spanning healthcare applications, 6G network technology, and language models. The updates reflect the company's push to integrate AI across multiple strategic research areas. Source: quantumzeitgeist.com

Insurers keep choosing the same AI model that users also prefer most

The article points out that the AI model insurance companies most frequently adopt for deals also ranks as the most popular choice among everyday users. This overlap suggests strong market consensus on which model currently leads. Source: insurancebusinessmag.com

Prompt injection explained: why it can be mitigated but never fully eliminated

The article breaks down what prompt injection attacks are and how they manipulate AI systems into ignoring intended instructions. It explains why current defenses can reduce risk but cannot completely prevent such attacks. Source: sqmagazine.co.uk

Perfectly formatted JSON from an LLM doesn't guarantee correct answers

The article warns that even when a language model outputs syntactically flawless JSON, the actual content can still be factually wrong. It stresses the need for validation beyond just checking format compliance. Source: towardsdatascience.com

Study assesses how well large language models handle clinical bacteriophage therapy questions

Researchers evaluated the accuracy of large language models when used as a source of clinical information on bacteriophage therapy. The findings highlight both potential and limitations of using AI for specialized medical guidance. Source: nature.com

Is your private cloud LLM really ready to power coding agents?

The piece examines whether privately hosted large language models can match the performance needed for autonomous coding agents. It raises questions about capability gaps compared to top commercial models. Source: vmware.com