A post titled "Mercury 3: Diffusion at scale" appeared on Hacker News. The title indicates Mercury 3 applies diffusion at scale.
If diffusion-based models are now arriving at production scale, developers choosing models have a new option to benchmark against autoregressive ones on latency and cost.
Amazon's new TorchNeuron dev tool is meant to break Nvidia's hold on PyTorch, the free toolkit used to build most AI apps, so more developers run on alternative chips like Amazon's own. Hugging Face says running its models on AWS Trainium is easier than ever with TorchNeuron.
Developers with PyTorch code can now evaluate AWS Trainium as an alternative to Nvidia GPUs when comparing infrastructure choices and training or inference costs.
Hackers got into part of the online platform of OpenCode, an open-source AI coding agent seen as a popular alternative to Claude Code and Cursor, and stole some users' API keys, running up charges on their AI accounts. The AI tools it turned to for help kept refusing to look at the malicious code.
If you connect your own API keys to a coding agent like OpenCode, a stolen key turns directly into charges on your account, so check key scopes, spending limits and rotation procedures.
AI security firm Norforge found that a hacker broke into more than 100 websites using an army of AI agents that mostly operated on their own. Norforge doesn't name the targets, but Reuters reports victims included John Deere, SharkNinja, and Cummins.
With autonomous agents now used for large-scale intrusions, developers running web services should audit exposed assets and access controls assuming attacks are automated and fast.
The article defines a harness as the program that builds, maintains and protects the conversation JSON array sent to a stateless LLM API on every call. It uses that idea to explain how AI agents work.
When choosing or building an agent framework, developers can check where the conversation array is built, kept and protected, since that is the harness's job.
The article covers using tmux with Claude Code, including sessions and send-keys, and where the approach hits its limits. Per the GeekNews summary, tmux keeps AI coding agents in named sessions that survive a closed terminal or dropped SSH connection, and lets you manage several agents in one window.
If you run coding agents for long stretches on remote machines or in parallel, tmux sessions and Claude Code's claude --worktree --tmux option let you check whether dropped SSH connections can stop killing your agent runs.
CERN and Mistral have announced a collaboration to accelerate AI for science. The news was shared on Hacker News.
Developers building scientific or research tooling now have a reason to add Mistral to their model shortlist and watch for what this partnership produces.