Today in AI

In partnership with

NEW LAUNCHES

The latest features, products & partnerships in AI

• Soon you’ll be able to make AI images directly within Google Docs

• Struggling to find the right stock photo? Try Getty Images’ new AI tool

• fieldmobi.ai makes ERP deployment a breeze for companies of any size

• AI search engine Genspark adds Claude-powered financial reporting

• SUSE rebrands, launches AI platform to safeguard enterprise data

• Rabbit R1 users can now fully customize its entire interface

• ChatGPT Mac app gains code-reading ability

• YouTube is rolling out an AI music remixing feature to select users

AI AGENTS

Launches, research & more from the AI Agent Report Newsletter

• OpenAI’s ‘Operator’ tool could book flights and more by controlling your computer

• The most promising use cases for agentic AI

• AI’s third wave: How AI agents will transform workplace productivity

• Microsoft’s Magentic-One is an AI agent that writes code and browses the web

• The future of software development: AI agents vs AI copilots

GOVERNMENT

Press releases, regulation, defense & politics

• AI likely impacted the 2024 election, though not in the ways many expected

• Mass. economic bill includes millions in funding for AI, quantum computing

• DHS releases AI adoption guidelines for critical infrastructure

• The biggest concerns (and reassurances) of China’s military AI research

• Why a Trump administration may detour Schumer’s AI roadmap

• Biden administration releases final AI security guidelines

IMPLEMENTATION

Announcements, strategies & case studies

• Navigating the ethical minefield of AI-powered customer segmentation

• Integrating generative AI with your business data? You need RAG

• Why enterprises are increasingly using small language models

• National League of Cities, Google publish AI implementation guide for city governments

• How AI is streamlining the private aviation industry

OTHER INTERESTING STORIES

Compelling stories beyond the usual categories

• Kansas City newspaper building repurposed into AI data center

• Musk accuses OpenAI of antitrust violations in fresh dispute

• Gaming giant Razer pivots to business market with environmental impact AI tool

• AI satellites advance wildfire detection in California

• OpenAI accused of profiting from model inspection in NYT lawsuit

What’s happening in AI right now

AI models hit performance wall

Big tech continues touting AI breakthroughs, but troubling signs of model performance plateaus are emerging. While Google celebrates Gemini 1.5's million-token context window and OpenAI teases GPT-5, users report degraded outputs from current models and experts warn of fundamental limitations ahead.

Signs of stagnation

OpenAI's newest models show only marginal improvements over previous versions - a stark contrast to the dramatic leaps between GPT-3 and GPT-4. Industry data suggests this isn't an isolated case. Major AI platforms are showing increasingly erratic and inaccurate responses, challenging the narrative of inevitable progress through scaling.

The data dilemma

The slowdown stems largely from an emerging crisis in training data. Experts project that by 2026, AI companies will have exhausted most available high-quality public text data. More worryingly, researchers have identified "model collapse" - where systems trained on AI-generated content show degraded performance over time. With OpenAI alone generating approximately 100 billion words per day, the risk of models training on their own increasingly degraded outputs grows.

Big promises, bigger challenges

Yet despite these mounting technical challenges, tech executives continue making aggressive predictions. Industry leaders now forecast artificial general intelligence arriving between 2026-2030, driving unprecedented investment in AI infrastructure. Microsoft alone is spending billions on specialized AI datacenters, while Google and Meta race to match capabilities.

Alternative approaches

The disconnect between bold promises and technical reality has sparked new approaches. Research teams are exploring mathematical data - leveraging patterns and relationships that could provide theoretically infinite training material. Unlike text or images, mathematical structures don't suffer from quality degradation or exhaust their novelty.

The enterprise reality

Meanwhile, enterprises are quietly pivoting to more practical solutions. New data shows 77% of enterprise AI implementations now use models with fewer than 13 billion parameters. These smaller models often match their larger counterparts in specific tasks while offering dramatically lower latency and operating costs.

Small models, big impact

This enterprise shift coincides with growing interest in Small Language Models (SLMs). Rather than chasing raw size, SLMs prioritize efficiency and can run locally on devices without requiring constant internet connectivity or massive computing resources. Companies like IBM and AMD are investing heavily in this approach, betting that practical advantages will outweigh marginal improvements in large model capabilities.

The training data race

The scramble for high-quality training data has meanwhile sparked new competition. Research indicates that AI companies are increasingly paying premium prices for pristine datasets, recognizing that model performance depends heavily on training data quality. Meta reportedly considered acquiring Simon & Schuster to use its publications for training large language models, as revealed in an audio recording shared with The New York Times. Some firms are exploring synthetic data generation, though this approach risks accelerating model collapse if not carefully managed.

As tech giants push forward with massive investments and ambitious timelines, the underlying technical challenges paint a more complex picture. The coming months may reveal whether bigger truly means better in AI development, or if the future belongs to more focused, efficient approaches.

We publish daily research, playbooks, and deep industry data breakdowns. Learn More Here

One of the holy grails of unlocking the potential of AI is to build systems that can reason like humans. By improving AI's, Large Language Models in particular, ability to break down complex problems and apply logical steps.

Bagel’s research team has been exploring this problem. Analyzing LLM building techniques, especially fine-tuning techniques, to allow Large Language Models to evolve from pattern-recognizing prediction agents to true cognitive agents

Read More Here 🥯 

AI generated art

A look at the art and projects being created with AI.

Midjourney Prompt: A yellow and pink and green Ginkgo trees , with a clear blue sky and lake reflection, in a minimalist style. The composition features a large area of blank space, with a simple layout captured using a wide-angle lens. The delicate texture and cool tones create a static, tranquil atmosphere