A newsletter breaking down AI research, technology, and Australian property in plain English
Tsinghua's reinforcement learning method teaches long-horizon AI agents to compress their own context, lifting open models on coding benchmarks and entering GLM
A new method uses hidden activations to block bad prompts instantly, requiring no model retraining.
New arXiv research reveals a dangerous blind spot in cybersecurity AI: models can resist attacks on paper while their decision logic falls apart.
New ICLR 2026 research finds masked image modelling is inherently more robust than contrastive learning when decentralised unlabelled data differs across nodes.
An arXiv paper claims kNNGuard reads hidden LLM activations to block unsafe prompts without fine-tuning, cutting latency and setup time to seconds.
ACMA's new 4G and 5G rules require a standard method across Australia, giving the ACCC a benchmark to judge network claims it previously had to drop.
Auction clearance rates are softening and economists expect modest declines, but arrears remain low and the wealth effect may not bite yet.
Researchers measure whether agents can iteratively edit executable policies, with trajectory diagnostics replacing single final scores.
Researchers analysed 75 breaches and found a hidden flaw—syntactic validation clears, but meaning fails when data crosses trust boundaries.
