Blog
Notes and analysis on AI development.
Robinhood opens MCP trading, but agent losses stay with users
Robinhood opened Trading MCP and Banking MCP for AI agents. The real developer story is the permission, approval, and liability model around financial tool calls.
OpenAI launches Rosalind Biodefense as a trusted-access test for GPT-Rosalind
OpenAI has launched Rosalind Biodefense, pairing GPT-Rosalind, Codex life-science tooling, and trusted access for public-health defense work.
Arm open-sources Metis, a security agent aimed at SAST false positives
Arm has open-sourced Metis, an agentic AI security framework for code review, SARIF triage, evidence chains, and lower false-positive cost.
LFM2.5 8B Brings Tool Calling to Local Agents
Liquid AI released LFM2.5-8B-A1B, a 1.5B active MoE model with 128K context for local tool calling, structured outputs, and agent workflows.
AI Passed the CAPTCHA, but 0.88 AUC Caught the Process
Roundtable and an arXiv paper argue that AI agents can match CAPTCHA answers while still revealing themselves through click order and behavior.
Claroty Claire Launches With Approval Boundaries for Factory and Hospital AI Security
Claroty introduced Claire, a CPS-native AI security agent. The launch shows why AI action in factories and hospitals has to be tied to asset data, approvals, and audit trails.
Reactor raises $59M for real-time AI world APIs
Reactor launched as infrastructure for real-time world models, moving AI video competition from rendered clips toward latency, sessions, and API operations.
Kog claims 3,000 tokens/s, and coding agents hit a latency wall
Kog KIE tech preview claims 3,000 tokens/s on 8x MI300X. The useful question is what 2B-model, batch-1 latency means for coding agents.
Antigravity Ran 93 Agents and Put a Price Tag on OS Demos
Google Antigravity teamwork-preview used 93 subagents, 15,314 model calls, and 2.6B+ tokens to build an OS demo. The useful signal is the cost model.
Five LLMs split on 67% of fact-checks, and AI search absorbs the cost
Lenz Research tested 1,000 real fact-check claims across five frontier LLMs and found that 67% did not receive the same verdict.
AI coding teams ship daily, but DevOps is paying the bill
Harness 2026 survey data links heavy AI coding use with faster deployment, more delivery pressure, and downstream security, rollback, and burnout signals.
Fujitsu brings Claude to 100,000 staff and adds Codex to SI delivery
Fujitsu announced OpenAI and Anthropic collaborations on the same day, pointing to a multi-model SI strategy built around Claude, Codex, FDE, and enterprise controls.