
What happened
Google DeepMind introduced Gemini 4 Hydrogen and Gemini 4 Carbon, with Hydrogen being integrated into Google Search for real-time reasoning and high-volume search requests.
Why it matters
Teams building products may now need to pick models by latency and throughput needs, not just by overall benchmark scores, the piece argues.
Summaries like this, every morning.
The Gemini 4 Hydrogen and Carbon split continues a shift away from one giant model covering every task. Hydrogen is being woven into Google Search itself, where responses are expected in milliseconds, while Carbon is presented as a separate, available model. For teams that embed models in products, the practical consequence is that model selection now spans latency, throughput and context-window trade-offs, not just a headline benchmark score.
The week's research papers point in a similar direction at the system level. The memory-filtering paper (arXiv:2611.17338) finds that when an agent's resolver hard-deletes facts it judges stale, some of those facts turn out to be needed to answer complex user queries, and once removed they cannot be recovered. The suggested alternatives are softer: decaying a fact's retrieval priority, or keeping conflicting old and new facts as a version chain. LoRAWeaver (arXiv:2611.17340) addresses a related strain, the interference that arises when many task-specific LoRA adapters share one base model, by routing adapters as directional vectors in a representation space.
Separately, OpenAI opened a science-focused blog and published work toward defining the Standard Model on a lattice, saying the frontier model GPT-6 Astra helped turn an initial concept into rigorous mathematical proof steps. This connects to formal verification work such as Google DeepMind's AlphaProof Code, framing proof-and-code generation against formal systems as moving into practical use.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
BOOKERS now supports MCP, so users can connect their account to Claude or ChatGPT (web version, confirmed work…

IMPACTLAKE began offering a ChatGPT plugin for impactlake(TM) and added Model Context Protocol (MCP) support…

On a Dragonwing RB3 Gen 2, 6 of 9 InternVL3.5-2B graphs profiled via the manual AI Hub chain

The author installed the open-source agent platform OpenClaw on an EPSON Endeavor NY2100S (Celeron P4600, 4GB…

A freelance web and SEO tool developer reports running Claude Code with one COO instance, 15 sub-agent departm…

A developer using Claude Code on a personal web app discovered via --debug logs that his Stop hook passed ever…
