Skip to content

Frontier AI Review

Subject desk / 4 articles

Training

Pre-training and post-training advances that change the cost, speed and behavior of AI models.

Pre-training · Post-training

Mistral’s Forge illustration shows agents learning through actions and rewards in an environment. It illustrates reinforcement learning, not Large 4’s undisclosed training setup. Image: Mistral AI

Order

4 articles

Analysis

Mistral’s Forge illustration shows agents learning through actions and rewards in an environment. It illustrates reinforcement learning, not Large 4’s undisclosed training setup.

Mistral’s training recipe makes the curriculum a moving target

Reinforcement learning must keep challenging a model as it improves. The engineering problem is producing useful attempts, trustworthy rewards and a curriculum that does not stand still.

Post-trainingMistral

4 min read · 11 Oct 2026Read analysis ↗

Analysis

NVIDIA’s Blackwell processor imagery. Hardware context for the precision trade-offs in model training.

NVIDIA’s four-bit training makes precision a budget decision

Nemotron 3 Ultra shows why cheaper pretraining depends on knowing which calculations cannot afford to be cheap.

Pre-trainingNVIDIA

4 min read · 11 Oct 2026Read analysis ↗

Analysis

Official FlashAttention-4 research artwork from Together AI.

FlashAttention-4 targets the waiting hidden inside model training

Faster chips cannot remove every bottleneck. The new attention kernels show why training software determines how much research a GPU budget buys.

Pre-trainingTogether AI

4 min read · 11 Oct 2026Read analysis ↗

Analysis

Official artwork accompanying the Thinking Machines and UIUC research on SQL reinforcement learning.

Thinking Machines’ SQL study puts expertise inside the training signal

A specialist-model result suggests that knowing what a correct answer means can matter more than adding another layer of prompts.

Post-trainingThinking Machines

4 min read · 11 Oct 2026Read analysis ↗