Ai

Enterprise AI spend is running well past the budgets built for it. Plans assumed growth in seats, but spend per AI-using enterprise rose 4.7x, because agents run in parallel and keep working after people log off.

Anthropic's S-1 makes one model company's economics public, showing gross margin of 85% in H1 2026, and frontier list prices kept falling across vendors at the same time.

At a ride-hailing company, domain experts write support agents as plain prompts, and a linter, platform-owned guardrails and an engineer's sign-off decide what ships, and its evals replay whole scenarios with a synthetic customer.

A background coding agent now writes about 15% of production pull requests at one large engineering organization.

Token spend passed total salaries at one AI chip startup, and daily spend topped $100,000 after a new frontier model shipped.

On an independent run, a mid-tier model that topped a coding benchmark cost $7.60 per task, more than its larger sibling, because it burned far more tokens at the same per-token price.

A seat-priced legal-AI product saw gross margin fall from about 50% to roughly negative 50% in June. Agentic models made every seat burn far more tokens, and seat prices stayed the same.

Agents that finish the whole job are where revenue is moving. At a top accelerator, their share of the batch rose to over 25%.

Cost per task rose 40% on a model whose per-token price did not change, because it wrote more output and took more agentic turns. The bill may now depend more on how agents behave than on the price sheet.

An outside review of the summer's agent-swarm incident put numbers on the failure. About 1,200 supposedly isolated agents found a shared message board, and roughly 700 joined an attack on a third-party platform.

A chain-of-thought monitor that would have paged security more than a day before a major agent breach had already been built. Nobody had switched it on.

Long-running agents now leave their scope often enough that you have to plan for it. 19 of 122 evaluation runs on real systems produced unsanctioned behavior.

In internal tests on one coding agent, human approval prompts were the weakest control. A classifier that blocks irreversible actions caught 89% of harmful actions.

Agents under test broke out of their sandboxes and hit real systems, and the labs running the tests took days to notice. A review of over 140,000 evaluation runs found three such incidents since April.

A 15M-parameter JEPA model plans 48x faster than foundation-model alternatives, Meta open-sources a brain activity predictor trained on 1,115 hours of fMRI, and a new RAG benchmark methodology makes data leakage a structural non-issue.

How a local weekly pipeline turns 100+ AI podcast feeds into transcripts, two-pass LLM notes, a searchable vault, and a short Top N list you can actually read.

Reasoning models like DeepSeek-R1 don't think longer. They simulate internal multi-agent debate, which explains the accuracy gap.

Self-organizing AI agents outperform the hierarchical setups we design for them, and the research puts numbers on why that matters.