Explore what's new in AI

Anthropic tests Bugcrawl with 10 parallel repo-scanning agents

If this ships, Claude Code gets a much more parallel bug-hunting path for teams and enterprise, with obvious spend and safety implications.

Greg Brockman says GPT-5.5 is a beginning, not an endpoint

The quote frames GPT-5.5 as an intermediate step, which is useful context for teams trying to judge whether to wait or ship now.

CodexBar 0.23 adds Mistral support and GPT-5.5 pricing widgets

A small release, but it closes a bunch of rough edges that matter when you live inside agentic editor workflows all day.

Vtrivedy10's Alpha-Eval treats evals as training data for agents

The core idea is practical: use failed evals and production traces to generate better training signals, instead of treating evaluation as a dead end.

ArithTuple tensors in CuTeDSL get a practical walkthrough

A useful kernel-programming explainer if you want to think about tensor layouts and code generation without drowning in compiler jargon.

Linear and Granola turn meeting notes into issues through MCP

The integration closes the loop from meeting to task, which is exactly the kind of agent workflow that saves teams from manual copy-paste.

MIT's recursive LLM diagram keeps context small with surgical tool calls

The pattern is simple: inspect files surgically, recurse on subproblems, and avoid stuffing everything into one giant prompt window.

maharshii's custom torch.compile backend hits 1.5x on fused ops

A small backend shim before Inductor delivered a real speedup, which is a good reminder that compiler plumbing still has easy wins left.

Google tests website generation and catalog uploads in Pomelli

Pomelli is moving from marketing prompts toward site and catalog generation, which would make it far more useful for SMB workflows.

trycua opens Cua Driver for macOS app control

The new driver lets any agent drive native Mac apps, not just browsers, which expands computer-use workflows beyond web-only automation.

wacrawl archives and searches local WhatsApp Desktop data on macOS

It snapshots WhatsApp SQLite databases into its own archive and adds FTS search, which is handy for compliance, e-discovery, or personal lookup.

Google open-sources DESIGN.md for reusable design rules

The format makes design intent portable across tools and agents, so teams can keep branding consistent without re-explaining every mockup.

Ethan Mollick argues AI reviewers can reconstruct papers from methods and data

He says frontier agents can rebuild complex papers from methods and data alone, which could change what a competent review workflow looks like.

Avichawla's prompt formula adds ARQ and verbalized sampling

A useful cheat sheet for getting better outputs from the same model before you jump to a new harness or a bigger model.

thdxr finds tool-call pruning changes cache hits and spend

The result is less straightforward than it sounds, pruning can improve cache behavior on some workloads, but the dollar savings can still be small.

Scaling test-time compute for agentic coding uses summaries and voting

The paper shows big gains by compressing rollouts into structured summaries, then reusing them for recursive voting and refinement.

PARE benchmarks proactive agents with stateful app simulations

The benchmark models apps as finite state machines, so agents have to infer goals and act at the right time instead of just firing tool calls.

Microsoft's DELEGATE-52 shows long document workflows corrupt files

The benchmark says even frontier models silently degrade documents over long delegated edits, and tool use does not reliably fix it.

Sakana's TRINITY routes tasks across thinker, worker, and verifier roles

A tiny coordinator with under 20K parameters beat larger setups by dynamically assigning models to different roles at test time.

BytePlus adds GLM-5.1 to its coding plan

The plan bundles multiple frontier models behind one pricing layer, which makes it easier to benchmark and switch by task.

Anthropic brings Claude Code web and mobile closer to desktop

The new sidebar, drag-and-drop layout, and routines view make the web and mobile clients feel less like second-class companions.

Google prepares Gemini Usage Limits and a new Images tab

The UI leak suggests Google is tightening quota visibility while expanding image workflows, a sign Gemini is becoming a broader agent surface.

Hugging Face lets GPT-5.5 join ml-intern with full infra access

Giving the model access to buckets, jobs, and repos turns it into a real research operator, not just a chat endpoint.

OpenAI says the next AI step is a context-aware model

Altman frames the next jump as context, goals, people, and device history. That is basically the product case for richer memory and agent state.

Microsoft rolls GPT-5.5 into Copilot, M365, and Foundry

The model is moving from a standalone launch into enterprise surfaces, which is where adoption and workflow lock-in really happen.