AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
Ahead of AI
Reporting and perspectives from Ahead of AI. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Ahead of AI published a technical walkthrough of positional embeddings in LLMs, tracing how GPT-2's learned absolute vectors and the original transformer's fixed sinusoidal encodings give way to RoPE, which injects relative position inside attention in each block.

Sebastian Raschka released eight coding videos, totaling 11 hours, that build a reasoning model from scratch in Python and PyTorch on a small pre-trained Qwen3 model, covering evaluation, inference-time scaling, RLVR/GRPO training and distillation.

A new technical write-up places Jev, a much-discussed classifier model, in the lineage of text classification from bag-of-words through RNNs, CNNs and transformers, arguing its edge is generality at lower cost and latency than large LLMs.

OpenAI released GPT-6 Astra, which reportedly scored 99.9% on ARC-AGI-3 versus 7.8% for GPT-5.6 Sol, and The Information reports it uses "recurrent depth" or looped transformer blocks. The article examines that design and the rumor that Astra hides its chain of thought.

Anthropic has begun watermarking text generated by its Claude models, making it possible for the company to identify its own output, and Sebastian Raschka has published a 48-minute video lecture and slide deck explaining the mechanism, its detection, and how it can fail or be removed.

A tutorial project describes building a simple AI text detector by fine-tuning a DistilBERT classifier that returns a 0-100 score, then using it as a verifier to train a small language model to produce text that avoids detection.

Ahead of AI's new article explains how reasoning models get multiple reasoning-effort settings, using OpenAI's recently released GPT-5.6 family as the motivating example. GPT-5.6 ships in three sizes, each with roughly five or six effort levels ranging from Light to Ultra.

Sebastian Raschka published a curated list of LLM research papers he bookmarked from January through May 2026, organized into ten categories. He notes it is not a complete list of the year's publications and that he read only a subset of the papers in detail.

A survey of open-weight LLM architecture changes from April to May shows a common push to cut long-context costs, including cross-layer KV cache sharing in Google's Gemma 4 E2B and E4B, which saves about 2.7 GB and 6 GB respectively at 128K context in bfloat16.