AI Roundup 113: Liberation Day

April 11, 2025.

Apr 11, 2025

Liberation Day

In the chaotic aftermath of President Trump's new tariff policies, Semianalysis has put together an excellent deep dive into what they mean for AI infrastructure.

The big picture:

The tariffs are currently set to 145% on Chinese imports and 10-50% on other countries, though semiconductors and Mexico/Canada trade have key exemptions.
A critical USMCA loophole allows companies to avoid GPU tariffs by routing through Mexico - which may explain why companies like Foxconn are rapidly building factories there.
Ultimately, the most significant risk isn't the direct cost impact but how economic uncertainty could force AI labs to delay or abandon training efforts to preserve cash during a potential global recession.

Elsewhere in AI geopolitics:

After Jensen Huang paid $1M for a Mar-a-Lago dinner, the White House rescinded planned Nvidia H20 export restrictions to China.
China has established an AI Industry Investment Fund worth $8.2B to support early-stage AI projects.
The EU announced plans to build AI gigafactories with 100K advanced AI chips as part of its AI Continent Action Plan to compete with the US and China.
OpenAI and Google have rejected the UK's proposal to allow AI training on copyrighted work without permission unless rights holders opt-out.
And the White House has ordered federal agencies to appoint chief AI officers and expand AI use, reversing previous Biden-era safeguards.

Cloud Next

Google's Cloud Next conference was this week. The two biggest reveals, model-wise, were Gemini 2.5 Flash, a fast reasoning model, and Sec-Gemini v1, an experimental AI model for cybersecurity.

All the other announcements I could find:

Firebase Studio, an AI-powered IDE for creating custom apps in-browser.
Agent2Agent, an open protocol for AI agent collaboration across platforms (plus support for Anthropic's Model Context Protocol in Gemini models).
Workspace Flows for AI workflow automation alongside new features for Docs, Sheets, and other tools.
Ironwood, its seventh-generation TPU optimized for inference processing.
Unified Security, a platform integrating multiple security operations features.
AI agents to Gemini Code Assist for advanced coding tasks like app creation from specifications.
An upgrade to Deep Research for Gemini Advanced subscribers to use Gemini 2.5 Pro Experimental.
Gemini Live camera features to Pixel 9 and Galaxy S25 devices in 45 languages.
AI Mode in Search will be available to more US Labs users with new multimodal capabilities.
And Reddit integrates Gemini into its conversational AI search tool Reddit Answers.

Elsewhere in the FAANG free-for-all:

Meta launches Llama 4 and rolls it out across its messaging platforms, while also telegraphing a Llama 4 reasoning model announcement for later this month.
xAI debuts Grok 3 API access with reasoning capabilities and tiered pricing for its standard and Mini versions.
Amazon introduces Nova Sonic, claiming it's the most cost-efficient AI voice model available through Bedrock.
Microsoft's AI chief advocates for an "off-frontier" strategy of developing models several months behind the cutting edge.
And GitHub implements premium requests for Copilot users switching between AI models.

Llama 4 is a technical triumph and a strategic stumble

Charlie Guo

Apr 9

Read full story

People are worried about AI safety

OpenAI and Google have been accelerating their AI model releases while appearing to cut corners on safety testing and transparency, in yet another example of the tension between competitive pressures and responsible AI development.

Between the lines:

OpenAI has cut safety testing from months to days for its latest models, while Google has launched new Gemini versions without publishing standard safety documentation.
As models become more capable, identifying potential safety concerns becomes more difficult - one source has said that some of GPT-4's dangerous capabilities were discovered after two months of testing.
And while AI labs say they've made testing more efficient through automation, critics argue this rush to market fundamentally compromises public safety.

Elsewhere in OpenAI:

OpenAI is rolling out a memory feature in ChatGPT that references past conversations for Pro and Plus subscribers, excluding UK and EEA users.
The company launched its Pioneers Program to develop specialized AI benchmarks for sectors like legal and finance.
OpenAI has countersued Elon Musk for harassment and seeks to prevent further attacks regarding the company's structure.
Reportedly, OpenAI discussed acquiring io Products, the AI hardware company from Sam Altman and Jony Ive, for over $500 million.
And OpenAI is testing image watermarks for ChatGPT's free tier, though implementation remains uncertain.

Elsewhere in AI anxiety:

Emissions from AI chip production grew 357% in 2024 according to Greenpeace, outpacing electricity consumption due to heavy reliance on fossil fuels.
Spammers used OpenAI's API to generate unique messages that bypassed filters and flooded over 80,000 sites with SEO spam in just four months.
US officials report DOGE is using AI surveillance to monitor at least one federal agency for anti-Trump discussions.
And leaked documents reveal Outlier and Scale AI employ freelancers to write stress-test prompts about sensitive topics like suicide and terrorism while banning CSAM content.