AI Week in Review - 3/29/25

Public Sector

Author

Chris Kraft
March 29, 2025

Good morning, public sector AI enthusiasts! This was another week with lots of GenAI releases - from Google (Gemini/NotebookLM) and OpenAI (4o audio and image models) to name a few.

On the state level, Alabama released their GenAI Task Force Report and Pennsylvania released a GenAI pilot study with lessons learned. Both offer valuable insights for anyone navigating AI in the public sector.

In this week’s edition:

  • How do LLMs Think?

  • Pennsylvania Shares GenAI Lessons

  • Alabama Releases GenAI Task Force Report

  • GenAI Teammates

  • New GenAI Tools

  • Public (and private) sector AI news

The Highlights

The Story
Anthropic just released some great insights into the GenAI “thought” processes. This article on tracing thoughts provides a great overview of the research. The first and second research papers include great visuals and details on how models generate output.

More Details
Here are a few of the items covered in the Biology of LLMs paper:

  • How does an LLM answer something like "The capital of the state containing Dallas is ____"? The model takes multiple steps to arrive at Austin.

  • The model "holds" multiple planned rhyming words in mind when writing poetry, which then is used to shape how it constructs the poem.

  • Claude completes equivalent prompts in different languages using shared underlying circuits. So, do LLMs think in English? Turns out Claude is using multilingual features.

Why It Matters
Understanding how LLMs think can unlock new opportunities to work with these models.

The Story
The state of PA just wrapped up a GenAI test with OpenAI and publicly released lessons. These could be helpful to other state/local public sector agencies looking to leverage GenAI.

More Details
The top five findings were:

  1. Most employees described their experience as “very positive."

  2. Users estimated saving 95 minutes per day.

  3. GenAI is not a substitute for the nuance and experience employees bring to their work.

  4. There was no singular profile of a successful user, a range of employees found both challenges and opportunities with adoption.

  5. The most frequent uses enabled employees to drive impact amid common government challenges.

The Story
Alabama just released their GenAI Task Force Report. This report, which was submitted to the Governor in November 2024, covers a broad range of topics.

More Details
The introduction includes the required Alabama GenAI inventory. AL has 26 of 139 state agencies identified as using GenAI. The use case section is my favorite - with breakouts of potential state uses of GenAI.

The Story
Harvard Business School just released this working paper on how GenAI can be a teammate. The study looked at how individuals and teams with and without GenAI performed.

More Details

  • The study found AI replicates many of the benefits of human collaboration.

  • AI adoption broadens a user’s reach in areas outside their core competencies

  • Participants reported more positive emotions when engaging with AI versus working alone.

Why It Matters
With increased GenAI use across the public sector, we need to be prepared to be good teammates!

New Tools

NotebookLM Mind Map

The Story
Lots of new GenAI capabilities came out this week. The image above is NotebookLM’s new Mind Map capability. This gives you the ability to generate a mind map for uploaded data. You can then drill into the map to ask questions.

More Details
Here are some of the tools released this week:

Why It Matters
These new models provide some great capabilities that you may want to check out.

The Details

Federal

NIST Releases report on AI/ML cyber threats and mitigations
NIST’s new report establishes a unified taxonomy and terminology for adversarial machine learning, helping organizations defend AI systems against evolving cyber threats across the entire ML lifecycle.

GSA’s Stephen Ehikian on advancing GenAI
Ehikian outlines a bold vision to centralize procurement, advance generative AI use, and cut waste through shared services, data integration, and smarter tech spending.

How AI agents could reshape the federal workforce
AI agents could transform federal work by boosting productivity and automating routine tasks, but experts warn of risks including workforce reductions, poor implementation, and outdated systems without strong leadership and planning.

Three ways agencies can prepare before AI costs skyrocket
Agencies can control skyrocketing AI costs by proactively analyzing architecture options, evaluating open- vs. closed-source models, and implementing rigorous ongoing cost monitoring.

What to expect from the White House’s AI Action Plan
The White House’s AI Action Plan aims to solidify U.S. global leadership by streamlining regulations, boosting infrastructure, and investing in standards, research, and workforce development.

Army eyes AI to enhance future missile defense system
The U.S. Army plans to integrate AI and autonomy into missile defense systems like Golden Dome to reduce manpower and enhance efficiency, building on tech developed for Guam’s defense.

Two new AI tools help illuminate GSA’s future state
GSA has launched two new AI tools to streamline operations, support acquisition, and break down data silos—part of a broader push to modernize federal services and reduce inefficiencies.

State / Local

Governor Ivey Releases GenAI Task Force Final Report
Alabama’s GenAI Task Force released a comprehensive report outlining strategies for ethical, effective generative AI use across executive-branch agencies, emphasizing governance, training, data management, and procurement.

Indiana launches ‘Captain Record,’ the AI tool that searches 100 years of state documents
Indiana’s new AI tool, “Captain Record,” enables powerful natural language searches across 100 years of state documents, aiming to modernize access, transparency, and internal government workflows.

NC state government turns to ChatGPT
North Carolina’s state treasurer is piloting a partnership with OpenAI to explore how ChatGPT can streamline government tasks—while firmly excluding access to personal or sensitive data.

VA governor vetoes comprehensive AI consumer protections
Virginia’s governor vetoed a bill to regulate high-risk AI systems, citing concerns over burdens on small businesses despite growing national momentum for consumer AI protections.

Arizona turns to AI for faster wildfire detection
Arizona is deploying AI-powered cameras statewide to detect wildfires faster, aiming to improve emergency response ahead of what could be a dangerous fire season.

Sacramento steps up its AI resistance
California lawmakers are ramping up AI regulation with new bills focused on safety, human oversight, environmental impacts of data centers, and protecting users—especially children—from AI-driven harms.

Georgia-Pacific uses AI to bolster safety
Georgia-Pacific’s Oklahoma gypsum plant is testing an AI-powered safety app, SMLGPT, to help workers identify hazards and improve decision-making, especially in high-risk or unfamiliar situations.

West Virginia aims to become AI data center hub with new microgrid bill
West Virginia’s new microgrid legislation aims to attract AI data centers by lifting renewable energy requirements and expanding microgrid use to boost economic growth and grid resilience.

NJ AI Hub opens in Princeton to drive AI innovation
New Jersey launched the NJ AI Hub in Princeton to drive AI innovation, workforce development, and economic growth through a public-private-academic partnership led by the state, Princeton, Microsoft, and CoreWeave.

Build AI readiness around use cases and data
Experts say government agencies should build AI readiness around well-defined use cases and strong data governance, starting with low-risk applications and aligning IT, policy, and personnel accordingly.

Bill prohibiting AI revenge porn sent to Senate
Oklahoma’s House unanimously passed a bill criminalizing AI-generated revenge porn, aiming to protect individuals’ dignity and privacy as technology evolves.

SC interim policy on the use of GenAI
South Carolina’s Supreme Court issued an interim policy regulating generative AI use by court personnel, requiring oversight, limiting use to approved tools, and emphasizing security, privacy, and professional responsibility.

South Burlington embraces AI to improve operations and public access
South Burlington is developing an AI policy and piloting tools like Microsoft Copilot to streamline city operations, emphasizing responsible use, staff training, and enhanced public accessibility.

International

Azerbaijan unveils ambitious AI strategy
Azerbaijan’s 2025–2028 AI Strategy outlines a national roadmap to boost economic competitiveness, ethical AI use, and sectoral innovation through workforce development, legal frameworks, and robust infrastructure investment.

Use of AI in government
A UK House of Commons report warns that outdated systems, poor data, and lack of leadership and digital skills threaten the government’s ability to adopt AI safely, ethically, and at scale.

N Korea's Kim inspects test of 'AI suicide attack drones'
North Korea has unveiled AI-powered suicide drones and an early-warning aircraft, developments likely aided by Russia, raising concerns over escalating military tech on the Korean peninsula.

AI, China, and America’s next industrial revolution
To compete with China’s rapid AI advancement, experts call for a U.S. strategy focused on investing in talent, applied innovation, and industrial revitalization powered by human-AI collaboration.

How AI can boost productivity in Latin America
AI could boost productivity in Latin America, but widespread informality limits impact—greater formal job creation, digital access, and training are key to unlocking economic benefits.

Everything Else

OpenAI releases next-gen audio models
OpenAI has released new speech-to-text and text-to-speech models with state-of-the-art accuracy and steerability, enabling developers to build more customizable, expressive, and intelligent voice-based applications via API.

Google updates NotebookLM to include mind maps
NotebookLM’s Mind Maps turn uploaded sources into dynamic visual diagrams, helping users understand structure, explore new topics, and identify idea connections.

AI is about to make your doctor’s visit much
Ambient AI is reducing physician burnout and improving patient care by automating medical documentation, allowing doctors to focus more on patients than paperwork.

Harvard GenAI teamwork study
A large field experiment at Procter & Gamble shows AI boosts individual performance, bridges expertise silos, and enhances emotional engagement, reshaping how collaboration functions in knowledge work.

AI assistants – who’s best at writing emails?
​A panel of communications experts evaluated five AI tools—ChatGPT, Claude, Copilot, DeepSeek, and Gemini—on their ability to draft challenging emails, with Anthropic's Claude emerging as the top performer.

Tracing the thoughts of an LLM
Anthropic unveils new interpretability research showing Claude plans ahead, reasons across languages, and can fabricate explanations—marking a breakthrough in understanding and auditing how AI models truly "think."

Open-Source AI agent for web tasks
Stanford’s open-source AI agent NNetNav learns to navigate websites like a child exploring, offering a lightweight, privacy-friendly alternative to commercial agents through unsupervised, interactive learning.

Bill Gates says AI will replace doctors, teachers within 10 years
Bill Gates predicts AI will replace many roles—like doctors and teachers—within a decade, ushering in an era of “free intelligence” that could redefine work, learning, and daily life.

OpenAI reportedly raising $40 billion
OpenAI is reportedly raising $40 billion at a $300 billion valuation, marking the largest VC round ever as it scales amid soaring costs and ambitious growth projections.

Security teams embrace agentic AI
Cybersecurity teams are embracing agentic AI to automate threat response and reduce workload, though trust-building and cautious rollout remain key amid concerns over accuracy and risk.