AI news, with the context that matters.

Technology & Research

Can Genie 4 replace a game engine? How Gemini watches over a generated world, and where it falls short

Google DeepMind's Genie 4 is a world model that generates video and sound together in response to key presses and text instructions. We read the official description of how Gemini 3.8 Flash watches the scene and steers it, then compare Genie 4 with game engines, which keep world state as numbers, and sort out what game developers can try with it.

Technology & Research
Read article →

What can you build with OpenAI Decisions API? Turning everyday language into application actions

Follow a player’s request from conversation to an NPC’s action, then explore responsive voice interfaces, work routing, and document and image review. When is a separate decision API worth using alongside a conventional LLM?

Technology & Research
Read article →

What GoTo’s regional-language AI reveals about language costs and reasoning strengths

Equivalent meaning can take different numbers of tokens across languages and tokenizers. Starting with GoTo’s Gemma 3 derivative, we separate billing, inference computation, reasoning text, and internal representations, and examine when English translation helps and when the original matters.

Technology & Research
Read article →

Why Anthropic moved from Git to Scion

Supporting more than 30,000 daily merges takes more than faster coding. Scion changes how code changes are recorded, history is queried and files are fetched, reducing the work done by development infrastructure.

Technology & Research
Read article →

Anthropic’s distillation defenses and the limits on reusing model reasoning

Anthropic combines restrictions on reasoning reuse, classifiers, and account controls. We examine what the reported 90–99% reduction measures, the full-trace versus summary experiment, and the implications for legitimate agent developers.

Technology & Research
Read article →

ColGrep-Agent delegates code search to a local 2B model

How a local 2B model searches for a larger coding agent, what the token savings trade off against time and accuracy, and how this relates to local LLM configurations.

Technology & Research
Read article →

Can Liquid AI d1 run on the same PC as your game?

d1-3B and the experimental 600M return decisions in one pass. What the published 8 ms really measures, and how to budget memory and integrate decisions without stalling a game.

Technology & Research
Read article →

Can AI check AI-written code? What GitHub ReviewBench measures

GitHub’s ReviewBench measures AI code review through coverage and finding quality. We explain its 219 public PRs, what 96.6% agreement means, and how to interpret the results for real development.

Technology & Research
Read article →

Strata runs a large MoE model by sharing computation between the CPU and GPU

Strata v0.1.39 speeds up local generation and adds concurrent requests and Codex CLI support. We explain the memory layout behind a large MoE on a 12 GB GPU, RAM and SSD requirements, long-prompt and concurrency trade-offs, and the limits of its Responses API.

Technology & Research
Read article →

Strands Decider 2B is a small AI model for making choices locally

A large LLM plans the work while a small model handles recurring choices. We explain how Strands Decider 2B works, how it can route tasks to local specialists, what confidence means, and the conditions behind its hardware support and performance figures.

Technology & Research
Read article →

GLM-5.3 changes attack economics and cyber defense priorities

When AI takes on exploit development and preparation, the same number of attackers may be able to try more targets. GLM-5.3 evaluations, observed AI misuse, and five Japanese incidents help explain the changing economics of attacks and the defensive work that needs urgency.

Technology & Research
Read article →

TensorFold speeds up local AI generation by drafting and verifying tokens

TensorFold runs models on Macs and supported NVIDIA GPUs. We explain draft verification, quantization, memory, and model support. Project measurements show faster single-request generation but lower throughput than vLLM at eight concurrent requests, so the workload changes the verdict.

Technology & Research
Read article →

When AI remembers you, who can read its memory?

What do device-derived keys and isolated execution protect in Google's new server-side memory design? We examine the encryption, two unresolved audit findings, and what is still unknown about availability, using primary sources.

Technology & Research
Read article →

Anthropic says Claude discovered a previously unknown enzyme system

Researchers followed up on patterns identified by AI with further analysis and experiments. We examine what is new, what has been verified, and why ten attempts to repeat the search failed to rediscover it.

Technology & Research
Read article →