EN · International edition · 04/11/2026
Inferama daily · 04/11/2026
The most important AI developments since the previous newsletter, selected for editorial relevance.
News coverage period: 03/11/2026 08:30 → 04/11/2026 08:30 (Europe/Madrid).
Baidu releases Unlimited OCR 2 for processing long documents in a single request
Baidu released Unlimited OCR 2 on October 27, 2026. The model processes pages in sequence and uses an attention technique designed to reduce the cost of handling long documents. Its performance and “unlimited” horizon claims come from the provider and have not been independently verified in the available evidence.
Read article →
Parameters in Machine Learning: What a Model Learns—and What Its Count Doesn’t Reveal
A parameter is an internal value that a model learns or adjusts and uses to transform inputs into outputs. Understanding the term helps you interpret models and parameter counts without confusing them with data, predictions, or direct measures of quality.
Read article →
Foundation Model: What It Is—and What the Label Does Not Imply
A foundation model is a model trained on broad data that can serve as a starting point for different downstream tasks. The label describes a general role; it does not guarantee capabilities, quality, or a particular architecture.
Read article →
DeepSeek V3.2 Evaluations: What Its Benchmarks Measure and How Far They Can Be Compared
The available documentation confirms the DeepSeek-V3.2 paper, distinguishes V3.2 from V3.2-Speciale, and includes external analyses of API benchmarks and V3.2-Exp. It is not enough, however, to reconstruct a results table or verify the protocols test by test.
Read article →
A shared robots.txt signal for AI training: what changed and what remains pending
OpenAI, Google, Microsoft, and Anthropic have announced their support for a voluntary mechanism that allows websites to express preferences regarding the use of their content for training AI models through a single rule in robots.txt. While it simplifies configuration, the standard remains in draft stage at the IETF and does not guarantee universal compliance.
Read article →
When a Tool Fails: How to Stop an Agent from Repeating an Action
A lost response does not prove that an action failed. This guide offers a procedure for checking the state, deciding whether to continue or stop, and testing safe recovery before deployment.
Read article →
OpenAI to unify ChatGPT and Codex usage limits in December 2026
Starting in December, Plus, Pro, and Business plans will share a single token allowance across conversation, agent tasks, and coding. Instant mode is exempt from consumption, but the exact date and specific per-plan limits have not yet been detailed.
Read article →
GLM-5.3-Flash under scrutiny: what its benchmarks support and what a real task costs
Z.AI attributes improvements over GLM-5.2 to GLM-5.3-Flash and reports a DeepSWE score of 63.4 versus 46.2. Available documentation confirms some model details, but it is not enough to interpret those figures or establish a universal price.
Read article →
GPT-4.1 Under the Microscope: Capabilities, Access, Pricing, and the Limits of the Evidence
OpenAI presents GPT-4.1 as a model with improvements in coding, instruction following, and long-context processing. The available documentation confirms some details, but does not establish every result, current pricing, or the present status of safety documentation.
Read article →