Experiments on open-source models to find
what holds up and what breaks.

Gemma 4 E4B vs the Gemma Family: Enterprise Benchmark Showdown

› Gemma 4 E4B (4B params) scored 83.6% overall — beating the 3x larger Gemma 3 12B (82.3%) across 8 enterprise suites.

Structured JSON Output from Small LLMs

› A well-instructed 2B model jumped from 30% to 90% compliance — outperforming models 3–4x its size.

Context Position Bias in Small LLMs

› Each architecture fails differently. Gemma-2B has strong recency bias (p=0.023). Llama-3B is completely flat (p=1.0).

RAG Compliance Week 4: 100% Recall

› v2 accuracy dropped from 68% to 65% — but blocks 7 more benign queries to eliminate the final 4 missed attacks.

NeMo Guardrails vs Prompt Injections

› NeMo: 55% recall. Llama Guard: 58% recall. Enforcement Engine: 93% recall on prompt injections.

[AINews] AI Engineer Europe 2026

Two quiet days in a row let us reflect on the first AIE in London.

OpenAI $100 plan 💳, Claude Cowork GA 🏢, Perplexity x Plaid 💸

OpenAI introduced a new $100/month ChatGPT Pro tier aimed at power users, expanding its pricing lineup between the $20 Plus and $200 Pro plans. The company confirmed the $200 plan…

Getting started with ChatGPT

Learn how to use ChatGPT, start your first conversation, and discover simple ways to write, brainstorm, and solve problems with AI.

Act Wisely: Cultivating Meta-Cognitive Tool Use in Agentic Multimodal Models

Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest