- AI Fire
- Posts
- 🕸 AI Agents Are Switching to Graphs
AI agents are moving beyond simple loops. Peter Steinberger’s question about “loops or graphs” points to a major shift that could make agents more reliable, or help them repeat the same mistake at a much larger scale.
What's on FIRE 🔥
IN PARTNERSHIP WITH 42DM
We analyzed 100 AI infrastructure companies to find out which GTM channels actually turn traffic into pipeline.
Built by 42DM and Chili Piper, for AI founders and marketers who want to stop guessing where GTM budget should go.
Live July 21, 1 PM ET. Usually locked behind months of research. Yours free.
AI INSIGHTS
Peter Steinberger recently asked: “Are we still talking loops or did we shift to graphs yet?”
That question sums up a major change in AI agent architecture.
Most agents still use a simple loop:
Complete a task
Check the result
Make an adjustment
Repeat
The problem is that one loop can optimize the wrong metric. A support agent might improve ticket resolution rate by closing chats faster, even when customers still need help.
Graph engineering connects several loops that watch and correct each other:
One loop improves speed
Another checks accuracy
An audit loop tracks real customer outcomes
A higher-level loop can change the target
This makes agent systems more reliable because one loop can catch another loop’s mistakes.
However, more loops don’t guarantee better results. If every loop depends on the same weak data, the entire graph can confirm the wrong answer.
Strong systems still need real-world anchors, including completed tests, verified customer results, actual revenue, and rules the agent can’t rewrite.
The shift from loops to graphs is really about helping AI agents improve without gaming their own metrics.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
1. Make a Great NotebookLM Video with This 60-Second Focus Trick. Learn how to choose the right sources, format, and steering prompt so NotebookLM turns one useful idea into a clear vertical video instead of a generic summary.
2. NEW ChatGPT Work Full Review: Can It Really Replace Claude Cowork? Build a daily morning briefing, save the workflow as a reusable Skill, and schedule it to run automatically every weekday inside ChatGPT.
3. How to Build a Future-Proof Career Before AI Changes Your Job. Learn the five human skills AI still struggles to copy, find problems people will pay you to solve, and turn your experience into a service, product, or audience you control.
FIRE RECAP: BIGGEST AI NEWS THIS WEEK
🧬 Google DeepMind launched a bioresilience program to help detect biological threats, track pathogens, and speed up vaccine design. Google says frontier models could become a key defense against future outbreaks.
⚠️ More than 200 AI experts and economists, including 15 Nobel laureates, urged governments to prepare for AI’s economic impact. They warn society may have only a few years to handle major job disruption.
🛡️ OpenAI introduced GPT-Red, an internal model trained to find prompt injection attacks before release. It helped GPT-5.6 Sol cut failures by 6x on OpenAI’s hardest benchmark.
⌨️ OpenAI launched Codex Micro, a $230 keyboard built for managing AI coding agents. Its keys show agent status, launch Codex workflows, and adjust reasoning levels, though it’s already sold out.
🇪🇺 Google must open parts of Android and Search to AI rivals under new EU rules. Competitors like OpenAI will gain access to key Android features and anonymized search data.
TODAY IN AI
AI HIGHLIGHTS
🚀 SpaceX is reportedly in talks with the Pentagon to supply billions of dollars in AI computing power. The deal would expand its role beyond rockets and satellite systems.
📚 AI-generated biographies are flooding Amazon, often filled with errors and fake details. One creator used ChatGPT to publish 445 books, earning nearly $6,000 during the holiday season.
🛡️ Researchers found a new way to stop AI hacking agents using “context bombs.” In tests, the defense cut successful admin takeovers from 57% to just 5%.
🇨🇳 Moonshot AI’s Kimi K3 is fueling a new debate over China’s open-source AI progress. Early tests suggest it can compete with leading frontier models, worrying some U.S. tech leaders.
📱 Vertu’s $6,880 Alphafold comes with an AI agent built for executives. It can automate multi-step tasks, but testing found wrong reminders, missed details, and inconsistent file memory.
💰 AI Daily Fundraising: Databricks signed a term sheet for a strategic funding round at a $188B valuation, led by Coatue. The capital will expand Unity AI Gateway, Genie, and Lakebase, while funding future AI acquisitions and research.
NEW EMPOWERED AI TOOLS
🧩 ZooData turns any URL into agent-ready JSON, helping AI agents use fewer tokens and access live Amazon and TikTok market insights.
💻 Clark is an AI coworker with its own cloud computer that completes research, websites, spreadsheets, decks, audits, and coding tasks.
🔎 AnySearch gives AI agents real-time, structured, and de-duplicated search results from trusted sources for more reliable answers.
🚀 ExploreYC lets you search 5,773+ Y Combinator companies, with funding data, hiring insights, maps, and AI-powered startup intelligence.
AI BREAKTHROUGH
Have you ever wondered if AI image generators are just copying and pasting from their training data? Google Research recently shared some interesting insights into this, breaking down exactly where diffusion models get their creativity.
The researchers found that diffusion models are not just simply memorizing the exact images they were trained on and spitting them back out.
Instead, the models learn what they call "smoothed score functions." In simpler terms, this means the AI is learning the underlying rules, shapes, and patterns of the data, rather than just saving exact copies of the files.
They create new images by "interpolating along the data manifold." This means the model understands how to smoothly blend different learned concepts together to build a completely brand new image that makes logical sense.
In short, the creativity we see in diffusion models comes from a structural understanding of how different visual elements connect.
They aren't just blindly pulling from a giant memory bank; they are actively mixing and blending learned patterns to generate something fresh.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply