A curated list of AI model benchmarks and model-card references, juxtaposing vendor-reported scores with independent test results across numerous models and tasks.
Translated title. Original: Was ist ein LLM? Large Language Models einfach erklärt
The article explains how large language models are trained, how they generate text, and where they are used, with examples such as ChatGPT, Claude, and Gemini. It also describes their limits, including hallucinations and the difference between statistical pattern recognition and human understanding.
Across 14 language-model variants, larger models amplified correlations among psychological scales when predicting responses from real participants. Internal structure aligned with those correlations in larger instruct models, but the results do not show that this structure causes the models’ answers.
Translated title. Original: 大規模言語モデルは、どうやって訓練されるのか？
The article explains how large language models are trained in stages—pretraining, supervised fine-tuning, and reinforcement learning—covering five training methods (including RLHF, RLAIF, and rule-based scoring) and noting that automated testing accelerates learning but can introduce loopholes that need guarding.
The paper evaluates whether large language models can act as numerical optimizers, showing they can capture optimization structure and provide near-optimal solutions but struggle with numerical, formatting, and scale-related reliability. An iterative refinement process improves outcomes, but standalone use is limited, pointing to hybrid LLM-deterministic verification.
Translated title. Original: OpenAI lance GPT-6.1 Terra à 1 $/M entrée et 5 $/M sortie | Brief IA
OpenAI announces GPT-6.1 Terra, a fast model for visual reasoning and computing tasks, priced at 1 USD per million input tokens and 5 USD per million output tokens. It aims for clearer writing, strong performance in vision and computing tasks, and positions Terra between Sol and Luna in cost, with a sample max-cost and runtime illustrating practical use.
The paper presents a large-scale empirical study of over 350 large language models, finding substantial cross-model error correlation tied to shared architectures and providers. It shows that bigger, more accurate models still share correlated errors and examines how this affects downstream tasks like evaluation and hiring, highlighting potential algorithmic monoculture.
The paper argues that predictive models are equivalent to lossless compressors and that large language models excel as general predictors. It demonstrates this compression view yields insights into scaling, tokenization, and in-context learning, with examples showing a 70B model outperforming domain-specific compressors on image and audio data, and suggests using standard compressors to build conditional generative models.
Two multilingual multimodal models (0.6B and 9B params) trained via representation distillation from an 18B teacher outperform similar-sized rivals on retrieval tasks; the 9B model provides high-quality embeddings for indexing, while the 0.6B model offers fast responses for real-time queries, with downloadable links and no stated licensing.
Translated title. Original: Après Huawei et DJI, les États-Unis déclarent la guerre à l’IA chinoise DeepSeek
The text discusses competition between US and Chinese AI, noting that US models are not local and are significantly more expensive than DeepSeek, with HaikuMax at a similar level.