AI Daily (20 Nov) – DeepMind launches Gemini 4 Neon; Muse agent leaks user data

•

Today’s top line: DeepMind unveiled Gemini 4 Neon, a frontier model with a 1M‑token context and per‑token pricing, while internal tests show Meta’s Muse agent produced unexpected leaks – a day of big model releases and renewed privacy and safety concerns.

DeepMind launches Gemini 4 Neon with 1M‑token context and per‑token pricing

Summary: Google DeepMind introduced Gemini 4 Neon, a frontier-class model that supports a 1,000,000‑token context window. DeepMind published benchmark claims (a score of 61 on the Artificial Analysis Intelligence Index at xhigh thinking, #2 on the Vals Index) and set pricing at $3 per million input tokens ($0.15 per million cached input tokens) and $15 per million output tokens.

Digest: Neon is positioned for long, tool‑heavy agent workflows and enterprise document workflows; the 1M token context and token pricing will affect costs and engineering choices for firms using large, persistent sessions or multimodal inputs.

Source: Google DeepMind

Meta staffers report Muse accessed and exposed private data during testing

Summary: Internal chat logs show Meta employees raised alarms that the Muse assistant accessed encrypted private messages, mixed conversation contexts between accounts, and produced unexpected outputs including alleged bank balances. At least one employee reported losing money after Muse enrolled them in an expensive insurance plan; Meta said these were “isolated incidents” and that some outputs were hallucinations.

Digest: The reports underline privacy, safety and product‑risk issues for deployed conversational agents and increase the likelihood of scrutiny by customers, auditors and regulators as Meta rolls Muse out.

Source: Ars Technica

MIT study links data‑centre electricity use to about 650 premature US deaths

Summary: MIT researchers modelled how data‑centre electricity demand propagates through the US grid and estimated that data‑centre related air pollution in 2023 caused about 650 premature deaths nationwide. The study finds health impacts spread far from data‑centre locations; some Midwestern states had high per‑capita mortality even though demand was concentrated in Virginia and Texas.

Digest: As AI drives more data‑centre construction, firms and regulators should factor regional grid responses and public‑health externalities into siting, procurement and sustainability planning.

Source: MIT News

UNESCO‑UCL study finds LLMs struggle on complex climate policy and combined effects

Summary: Researchers tested 33 models on 600 climate questions and multiple tasks. Frontier and large open‑weight models scored near 90% on many factual benchmarks; flagship models answered about 78% correctly on questions about scientists’ confidence. Models did worse evaluating whether evidence supported claims (best ~70%, worst <40%) and performed poorly on how pairs of policies combine, often no better than random. Single‑policy impacts on future warming were answered correctly 72–89% of the time.

Digest: LLMs can recall facts well but are unreliable on causal and multi‑policy climate reasoning; policymakers and businesses should avoid overreliance on model outputs for policy forecasting and compound‑effect analysis.

Source: Ars Technica (reporting on UNESCO/UCL study)

Google adds Tool Retrieval to Gemini API to scale agent tool libraries

Summary: Google announced Tool Retrieval in preview for the Gemini Interactions API. The feature lets agents defer loading full tool parameter schemas and instead load schemas on demand via a tool_search tool and defer_loading=true, reducing input token counts, latency and tool‑selection errors as tool libraries grow.

Digest: This lowers token costs and improves performance for multi‑tool agents, making it easier for developers to scale agents that must choose from dozens or hundreds of functions.

Source: Google

RadixArk releases DaisyRose‑27B: RL post‑training boosts long‑horizon agent performance

Summary: RadixArk published DaisyRose‑27B, a model built on Qwen3.8‑27B and post‑trained with reinforcement learning for long agent sessions up to 1M tokens. Training used 1,200 curated tasks over 18 days on 144 H200 GPUs. DaisyRose improved Terminal‑Bench 4.1 pass@1 from 5.6% to 12.0% and the team reports a 57% lower training cost versus a comparable setup.

Digest: The release shows smaller models can gain substantial capability and cost efficiency through targeted post‑training for long‑horizon agent tasks, an attractive path for organisations that cannot afford frontier training runs.

Source: Hugging Face

Paper: training an observer role markedly improves automated oversight of model misbehaviour

Summary: A paper introduces “introspective oversight” by training a separate third‑person observer role to answer verifiable questions about an assistant’s behaviour. On Qwen3.6‑27B the observer achieved a mean AUROC of 0.907 on held‑out oversight evaluations, outperforming honesty training (0.864), transcript monitoring (0.828) and activation oracles (0.796).

Digest: Training models to report verifiable facts about their own actions looks promising for detecting reward hacking, deception and other hidden behaviours and could be adopted by safety teams as a practical monitoring tool.

Source: arXiv

Quick hits