Big Tech's Decarbonization Pledges Waver as AI Data Centers Expand
Among the developments rounded up in "The Batch," deeplearning.ai's newsletter, the most striking is the warning that the rapid expansion of AI infrastructure is threatening greenhouse gas reduction targets. Major players such as Alphabet, Amazon, Meta, and Microsoft have pledged to curb emissions, but as they push ahead with massive data center buildouts, those commitments are said to be in jeopardy. Much of the near-term power is expected to come from fossil fuels, and reports suggest a short-term rise in emissions is unavoidable.
The same issue also covers news on the model front. GPT-5.5 reportedly outperforms existing models while also being flagged for hallucinations (factual errors), and Kimi K2.6 is said to be leading the top tier of open LLMs. The issue also features research comparing strategic thinking in LLMs versus humans, once again calling into question the gap between rising benchmark scores and the actual quality of reasoning (detailed figures for each model are not included in the data we have, so we keep this to a headline-level summary).
Google Pulls AI Studio App That Gathered 800,000 Pre-Registrations
According to reports on Hacker News and elsewhere, Google canceled the release of an app related to "AI Studio" after it drew more than 800,000 pre-registrations. Details are limited, and neither an official explanation from Google nor the requirements for the app in question could be confirmed at this time. The last-minute withdrawal of a project that attracted such a large audience is likely to raise questions about the provider's decision-making criteria.
A "Poison Font" That Disrupts AI Processing Debuts
Creative Bloq reports that type designers have released a free font that "poisons" AI. While the information available is limited, the font appears to be designed so that the letterforms themselves interfere with AI's ability to correctly process and learn from text and images. It can also be read as part of the pushback against AI training on creators' work without permission, and it is an intriguing defense operating at the "character" level, distinct from existing techniques for scrambling AI (specific mechanisms and validation results are not included in the data currently available).
"IssueTrojanBench" Sets a Trap for Coding Agents
"IssueTrojanBench," published on arXiv, proposes a benchmark that evaluates AI coding agents against "malicious issues." It appears aimed at quantifying the risk of agents being tricked, through cunning requests disguised as developer tickets, into unknowingly writing code with planted vulnerabilities. It is research that highlights a concern that grows with agent autonomy: the instructions fed to an agent can themselves become an attack vector.
The FISA Reform Gap That Lets Surveillance Go AI
The Center Square argues that with reform of the Foreign Intelligence Surveillance Act (FISA) stalled, the federal government continues to be positioned to use AI to surveil "entire communities." The legal framework has failed to keep pace with technological evolution, and AI-driven mass data processing could dramatically expand the scope of surveillance — a state of affairs that lays bare the lag in regulatory debate.
Passes Every Exam, but Never Has an Idea
An essay titled "AI Has Passed Every Exam. It Has Never Had an Idea" is drawing attention. It questions the distance between benchmark achievements and genuine creativity or original vision — a critique that lands harder the higher the performance scores climb.