AI videos, nothing else.

Matthew Berman tests DeepSeek V4.1 Flash

Matthew Berman tests DeepSeek V4.1 Flash on visual interpretation, browser-based applications and locally generated software, then discusses speed, benchmarks and cost.

Fuad Ali on rebuilding an AI agent harness

Fuad Ali describes lessons from rebuilding an in-product agent, including planning, memory, regression tests and human judgment.

Running AI Locally at Gannon University - 0xSero

At Gannon University, 0xSero explains the basics of model files and memory, and argues that running AI locally gives users more control over access and capability.

AI found a new result. Nobody can explain it.

Mo Bitar considers a reported AI-generated physics result and argues that human understanding remains important in scientific research.

Testing Genie 4 Interactive World Models - Tim Simmons

Tim Simmons explores Genie 4’s interactive generated worlds, sound and prompt-based changes while showing limitations in scale, object permanence and game mechanics.

Stop Giving Your Agent Tools. Give It a Sandbox - Mikyo King

Mikyo King explains how Arize Phoenix lets agents search for tools and compose their use in sandboxed code, and discusses the resulting efficiency gains and security trade-offs.

Anthropic Just Filed to Go Public

Jack Roberts and Nick Saraev discuss GPT-6.1 Terra’s pricing and computer use, then consider claims about Anthropic’s public filing and investment across the AI stack.

Deepseek 4.1 Flash on 2 DGX Sparks changed my life

Alex Finn connects two DGX Sparks to run DeepSeek V4.1 Flash locally, then tests conversational and game-building tasks and discusses continuous agent workloads.

What a bank payment error teaches us about AI agents

John Feminella uses a major bank payment error to explain five controls for giving AI agents useful autonomy while managing operational risk.

Four principles for avoiding AI agent productivity theater

Nate B Jones argues that useful AI agent work depends on accounting for human follow-up, choosing experiments deliberately, setting stop conditions and protecting attention.

AI mathematical breakthroughs and computational limits

Wes Roth examines reported AI mathematics results and explains how the exponential time hypothesis relates to questions about computational complexity.

Four questions shaping AI this week

Nathaniel Whittemore examines questions about open-weight models, cybersecurity, data center finance and funding for frontier AI companies.

Alex Stamos on AI Cybersecurity and Defence Economics

Alex Stamos explains why security teams need affordable AI-assisted patching and better coordination as attackers gain capabilities.

OpenBot: Local Agents, Model Switching and Security

AICodeKing compares OpenBot’s local agent workflows with managed cloud agents and highlights its security and licensing limits.

Tim Tyler on Accountability for AI Agents

Tim Tyler proposes treating agents as a governed digital workforce with identities, accountable sponsors and bounded authority.

Kimi K3.1 tested on creative coding and video

Pat Simmons compares Kimi K3.1 with other models on three creative builds and finds mixed results in his informal tests.

What heavy AI token use costs and why model prices change

Theo Browne questions a high estimate of AI token spending and compares the effects of caching, model prices and changing workflows.

Building feedback loops to improve AI agents

Fuad Ali demonstrates an agent improvement loop that combines production traces, failure cases, evaluations, proposed fixes and human review.

AI, Cybersecurity and Stolen API Keys

Riley Brown and Ansh Nanda discuss vulnerabilities, stolen API keys and how cheaper AI tools could affect cyber defence.

How to Test Your Agents on Real Tasks With Harbor - Xander Song, Arize AI

Xander Song explains how realistic tasks, isolated environments and outcome verifiers can help teams test agent changes, then demonstrates a Phoenix experiment that reduces unnecessary tool calls.