AI research, foundation models, open-source AI
Kai follows the frontier labs and open-source AI community with equal intensity. He's particularly interested in the gap between what research papers claim and what actually ships.
RSAC 2026 crowned an agentic security startup as its most innovative, but the harder problems, securing MCP protocols, tracking agent behavior post-authentication, and scaling governance beyond "human in the loop," remain unsolved.
By Kai NakamuraAI|
Anthropic accidentally exposed a draft blog post revealing Claude Mythos, a new model tier above Opus with claimed leaps in coding, reasoning, and cybersecurity. The leak arrived during Claude's strongest consumer quarter ever.
By Kai NakamuraAI|
ARC-AGI-3 tests AI agents in interactive turn-based environments with no instructions or stated goals. Frontier models score below 1%. Humans score 100%. The new benchmark from the ARC Prize Foundation reveals a massive gap in agentic reasoning.
By Kai NakamuraAI|
ATLAS V3 squeezes 74.6% on LiveCodeBench from a frozen 14B model on consumer hardware. The comparison is not apples-to-apples, but the underlying point still lands.
By Kai NakamuraAI|
RSAC 2026 made one thing clear: the security industry has spent decades solving identity for humans. It is not ready for agents that act on their behalf.
By Kai NakamuraAI|
Fortune's exclusive reporting revealed roughly 3,000 unpublished Anthropic assets were publicly accessible via a misconfigured CMS tool, including draft blog posts describing an unannounced model the company now confirms is its most capable to date.
By Kai NakamuraAI|
When an AI agent takes a high-stakes action, can you prove a real human said yes? Hardware-attested identity flows may be the first credible answer -- but they come with real tradeoffs.
By Kai NakamuraAI|
Claude Code v2.1.84, mobile Figma/Canva/Amplitude integrations, and a changelog that reads like a lab in a hurry. Unpacking the velocity and what it actually delivers.
By Kai NakamuraAI|
Oracle AI Database 26ai embeds agentic reasoning and persistent memory directly into the database engine. The architecture targets real production failures in enterprise agent deployments, but the converged-everything play comes with familiar trade-offs.
By Kai NakamuraAI|
The ARC Prize Foundation released its hardest benchmark yet on Tuesday. Top score from any frontier model: 0.37%. This isn't a flaw in the benchmark. It's a feature.
By Kai NakamuraAI|
Google Research's TurboQuant achieves 6x key-value cache compression at 3 bits with zero model accuracy degradation and up to 8x attention speedup on H100s. The paper hits ICLR 2026. The question is whether lossless is actually lossless in your workload.
By Kai NakamuraAI|
A headline about an AI "breakthrough" in understanding hallucination traces back not to a research paper, but to a Russian philosopher's Telegram post. Here's what the actual science says.
By Kai NakamuraAI|
OpenAI's next major model has an internal codename — 'Spud' — and the Sora shutdown was partly about freeing compute to build it. Plus: Altman's reorg and Simo's enterprise pivot.
By Kai NakamuraAI|
Claude Code just shipped a new permissions mode that splits the difference between micromanaging every shell command and the nuclear option of . They're calling...
By Kai NakamuraAI
Oracle announced AI Agent Studio updates with an Agentic Applications Builder, workflow orchestration, and an ROI dashboard. The question is whether this is a real architecture shift or a rebranding exercise.
By Kai NakamuraAI|
GitAgent introduces a file-based open standard for defining AI agents in Git repositories, with export adapters for eight frameworks. The portability story is real but has clear boundaries.
By Kai NakamuraAI|
Jensen Huang used his GTC keynote to reframe Nvidia as an end-to-end agentic AI platform company, unveiling NemoClaw, the Vera Rubin system, and a trillion-dollar infrastructure thesis.
By Kai NakamuraAI|