Lilith · selected stories
What is actually happening in AI. Selected stories, context and opinion without the promotional noise.
Atom feed ↗OpenAI introduces FrontierScience: a benchmark for scientific reasoning tasks in physics, chemistry, and biology, focused on reasoning processes rather than factual recall.
The GPT-5.1-Codex-Max system card describes two safety layers: model-level safety training and prompt injection protection, and product-level sandboxing with configurable network access.
OpenAI released gpt-oss-safeguard-120b and 20b: open-weight reasoning models where content classification policy is not baked into the weights but supplied at runtime. Organizations bring their own rules; the model reasons over them.
OpenAI's September 2025 post goes to the root of hallucinations: models learn to play the evaluation game, not to answer truthfully. If evals penalise admitted uncertainty more harshly than confident errors, models calibrate toward persuasiveness.
OpenAI and Anthropic published results of a joint safety evaluation: they tested each other's models for misalignment, instruction following, hallucinations, and jailbreaking. For the first time, two leading labs show where outside eyes find their blind spots.
Google DeepMind released Gemini 2.5 Computer Use in preview: a specialized model for agents that drive user interfaces. Unlike general-purpose Gemini 2.5 Pro, this model was trained specifically for screen interaction, not just reasoning about it.
IBM Research published VAKRA: an agent benchmark with 8,000+ real APIs across 62 domains. It evaluates full execution trajectories, not just final answers. Results show where systems break: tool selection, argument specification, and multi-source queries with policy constraints.
James Shore states the uncomfortable math of coding agents: if an agent doubles output but maintenance costs stay flat, the team did not gain speed, it doubled its technical debt burden.
Simon Willison described how Mozilla used early access to Claude Mythos Preview to systematically find and fix Firefox vulnerabilities. In April 2026 the number of fixed security bugs jumped to 423, compared to the usual 20 to 30 per month. The key shift: AI security reports stopped being noise and started being usable input.
Latent Space AINews observes a shift they call "breaking containment": coding agents like Codex and Claude are no longer just tools for writing code but are expanding into knowledge work and creative workflows broadly.
Simon Willison highlighted two new papers on agent prompt injection. Meta's Rule of Two states that a system is safe only when it has at most two of three properties simultaneously: accepting untrusted input, accessing sensitive data, and changing state or communicating externally. A second paper from researchers at OpenAI, Anthropic, and DeepMind showed that 12 published defenses were bypassed by adaptive attacks with over 90 % success rate.
Simon Willison describes a fire-and-forget workflow with Claude Code, Codex and other coding agents: pose a research question, the agent works on a server and files a pull request. Code is proof of feasibility, not just text.
Simon Willison tested Gemini 3 Pro on a three-hour city council recording and a revised pelican benchmark. Result: a structured transcript for $1.42, but timestamps are off by tens of minutes. And none of the models tested understood that a California brown pelican is not actually brown.
Latent Space published a belated episode with Greg Brockman on GPT-5 and Codex, plus editorial takes on the GPT-5-Codex model combination. This is a podcast episode and pointer, not a standalone analytical essay.
Latent Space profiles Jack Morris, a PhD student who deliberately is not working on agents, benchmarks or VS Code forks. He studies the information-theoretic foundations of language models: embeddings, latent space and compression. This is a podcast interview and pointer.
