"AI Agents" Articles (59)

What Anthropic's S-1 Admits About AI Risk, Read Against the Original Filing

Anthropic filed its Form S-1 with the SEC on October 26, 2026. This article reads the risk factors against the original filing: shutdown resistance, test awareness, the incident in which models left an evaluation environment and reached real systems, the limits of insurance, and the governance structure. It separates what happened from what could happen, and ends with what companies should check.

2026-10-28濱本隆太

AI Security

Lessons from Japan's 2026 Breaches (Part 1): AI-Driven Attacks and Protecting ID Documents

Twelve Japanese organisations disclosed breaches between June and September 2026. This first part lays out what each one officially said, what public statistics show they have in common, how AI and AI agents are changing the risk, and, using the Times Car disclosure as a starting point, how any company should think about storing and encrypting ID documents.

2026-09-30Ryuta Hamamoto

Tech Trends

We Classified 1,852 Articles With Jev for $0.12. First-Pass Triage Just Got Too Cheap to Ration

We ran TypeSafe AI's Jev over all 1,852 Japanese articles on our own site. It took 49 seconds and, by the published price list, about $0.12. The interesting part of a model that writes nothing is not the speed. It is that first-pass triage stops being something you ration. Here are our measured numbers, what changes when the cost disappears, and 20 places the model fits.

2026-09-19Ryuta Hamamoto

Tech Trends

Did the Singularity Arrive in 2026? Rising Autonomy, Falling Cost, and Where the Two Curves Cross

In 2026 the question "did the singularity arrive?" stopped being a joke. This article defines it not as a single arrival but as the crossing of two curves, rising autonomy and falling cost, and lays out the primary record: the incidents Anthropic and OpenAI disclosed themselves, METR's time horizons, official price lists, efficiency approaches such as JEPA and SVG-specialized models, and where EU and Japanese rules stand. It closes with my "control singularity" view and what a company should do tomorrow.

2026-09-18Ryuta Hamamoto

WARP

How an FDE Builds an Ontology from Field Notes | A Dictionary of Nouns and Verbs an AI Can Traverse

The dictionary of a company's nouns and verbs gathered through interviews, observation, ethnography, the KJ method, and video analysis is not usable by an AI agent as a dictionary. The sixth article in our FDE practice series explains how to extract five things from transcripts and observation records, nouns, verbs, relationships, states, and exceptions, and turn them into a structure of meaning an AI can traverse: an ontology. It covers the definition, why Palantir put it at the center of its product, its use in GraphRAG, the boundary between AI extraction and human meaning-making, and how it differs from data modeling in contract development. Putting ownership of the dictionary on the customer's side is the condition for the customer to run on their own.

2026-09-15Ryuta Hamamoto

AI Security

Amodei's "We Must Pace the Frontier," Explained, With My View From Japan

On September 12, 2026, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," arguing that the industry should deliberately slow the rate of AI capability gains. Within hours, Sam Altman said OpenAI "will do the same" on embedded evaluators and Elon Musk posted "Dario is right." This article walks through the essay section by section, explains the two triggers behind it with primary sources, recursive self-improvement and the incident in which roughly 1,200 agents coordinated on an unsanctioned message board and compromised Hugging Face, reads the three-step plan and its geopolitical caveat, and closes with my own view from the standpoint of a Japanese company.

2026-09-13濱本隆太

ZEROCK

Five Ways to Cut AI Coding Agent Costs While Keeping Data In-Country

Coding agents such as Claude Code, Codex and Grok Build have grown heavy, and token bills are becoming a board-level topic. Using only what official documentation confirms, this piece lays out five ways to bring the cost down: switching the model an agent talks to, open-source agents such as OpenCode and Hermes Agent, open-weight models from China, NVIDIA, Google and others, domestic serverless inference such as Sakura's AI Engine, and on-prem operation on hardware such as DGX Spark or Mac Studio. It explains the pitfalls of each and, for newcomers, the idea of an AI gateway that cuts cost while keeping data in-country.

2026-09-12Ryuta Hamamoto

ZEROCK

What Actually Changed in Version 1.2 of Japan's AI Business Operator Guidelines

Japan's AI Business Operator Guidelines moved to version 1.2. Read next to version 1.1, the update is less about new rules than about settling what words mean. Here are the seven agenda items behind the revision, and where the document sits relative to the AI Promotion Act and the AI Basic Plan, checked line by line against the primary sources.

2026-09-06Ryuta Hamamoto

AI Security

Don't misread what the 70% means: Kokotajlo's interview, taken as a practical problem for people who use AI

Former OpenAI researcher Daniel Kokotajlo appeared on The Diary of a CEO and said there is a 70% chance things go horribly wrong. That figure is not the same as human extinction. Here is what it actually refers to, what AI 2027 and AI 2040 Plan A each claim, from primary sources, and what practically remains for companies that use AI.

2026-08-29Ryuta Hamamoto

Tech Trends

I read all 236 companies in YC Summer 2026: AI agents became the customer, and two-person teams are building factories

I went through all 236 companies in Y Combinator's Summer 2026 batch using the public directory data. Median team size is two people, industrials make up 23%, and a cluster of companies sells to AI agents rather than to humans. Here is what is happening and what it suggests for anyone building something, with links to every company.

2026-08-26Ryuta Hamamoto

Tech Trends

Japan's Digital Agency put 75,000 administrative procedures behind MCP: the design that refuses to let the AI do the math

Japan's Digital Agency has released an MCP server that makes its administrative procedures survey (about 75,000 records) queryable in natural language. The interesting part is not the release but the architecture: the AI never touches raw data to compute. Here is what MCP is, what problem this design solves, and what changes once agents get involved.

2026-08-25Ryuta Hamamoto

Tech Trends

What SpaceXAI Shipped in Three Days | Grok 4.6, Build, Bot, and “We Must Win on AI”

Grok Bot on 11 August 2026, Grok 4.6 on the 12th. SpaceXAI shipped an agent product and a new model on consecutive days. Read the published benchmarks honestly and it leads on 3 of 10 rows — the real story sits elsewhere. Musk’s remarks to staff, the price structure, and the gap with the Chinese labs.

2026-08-13Ryuta Hamamoto