- AI Fire
- Posts
- 🍏 iSearch's Reborn With Googl(e)ye
MIT may have just cracked one of AI’s biggest limits, the “long‑context blindness”, and even call itself to handle 10‑million‑token inputs without forgetting a thing…
What are on FIRE 🔥
IN PARTNERSHIP WITH LUMOS
You don’t need another “AI overlay.” You need a system that understands context, acts autonomously, and learns over time.
While most companies rely on outdated approvals and ticket workflows, a few are handing it off to identity agents that don’t just assist… they execute.
🗓️ On Thursday, October 30th, Lumos is running a LIVE webinar on how to run access policies, approvals, and reviews using an autonomous AI framework!
If you're building systems that scale, secure, and self-improve, this is it - all handled by an AI that understands your org like a trusted teammate.
🎟️ Save your seat and let trusted agents take it from here.
AI INSIGHTS
If you’ve ever wished your AI chatbot could “look things up” or “take notes” like a real researcher, you need to read this. MIT dropped Recursive Language Models (RLMs), like a simple trick where LMs can handle huge input & avoid “context rot.”
Instead of cramming 300K tokens into a single prompt and hoping the model doesn’t forget what happened on line 1… RLMs can:
Peek at chunks of the context,
Grep for keywords or patterns (like “User: 67144”),
Split the data into smaller parts,
And recursively call itself to answer sub-queries.
It’s like watching a dev debug a huge dataset using code blocks (the dev is the model):
RLM(GPT-5-mini) outperforms GPT-5 by 114%↑ on accuracy
RLM(GPT-5) was the only method staying accurate as the doc count hit 1000
You give the RLM a giant block of context and a question, it’ll know to to handle all. It can also write programmatic long-form outputs like Git diffs.
Why it matters: It sounds like an agent but it’s not, it decides how to think (agents follow the rules). Now someone pls tell OpenAI to plug it into their agent system 🤡
🎁 Today's Trivia - Vote, Learn & Win!
Get a 3-month membership at AI Fire Academy (500+ AI Workflows, AI Tutorials, AI Case Studies) just by answering the poll.
How do RLMs overcome “context rot”? |
PRESENTED BY HUBSPOT
How 15 Small Brands Achieved Remarkable Marketing Results
Stop believing you need a big budget to make an impact. Our latest collection highlights 15 small brands that transformed limited resources into significant market disruption through innovative thinking.
Case studies revealing ingenious approaches to common marketing challenges
Practical tactics that delivered 900%+ ROI with minimal investment
Strategic frameworks for amplifying your brand without amplifying your budget
These actionable insights can be implemented immediately, regardless of your team or budget size. See how small brands are making big waves in today's market.
TODAY IN AI
AI HIGHLIGHTS
📚 Google NotebookLM now supports arXiv papers, just like talking to your very own professor. This will definitely our fav feature. You can view the video demo here.
🧰 Y'all realize that Anthropic is back in the spotlight again after a quiet stretch!? You can now give Claude particular skills to automate your tasks. See it in action here.
🧠 ChatGPT can now auto manage your saved memories, no more “memory full” alert. You can also choose which to re-prioritize in settings, see how-to guide here.
🧸 A stress-relief AI pet aimed at Gen Z just flopped, exactly as psychologists predicted. Users said it felt more awkward than soothing. See what happened.
💥 It’s no longer AI vs AI like ChatGPT or Claude. Now we got AIs vs Govs. The White House clashed with Anthropic over AI regulation, while D.C. calls it fear‑mongering.
💰 AI Daily Fundraising: n8n raised $180M, reaching a $2.5B valuation. Backed by Accel, NVIDIA, and Sequoia, it builds AI that blends autonomy with control.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
AI QUICK HITS
🎬 Sora 2 brings storyboards to users & expands video length
🌎 OpenAI’s low-cost Go tier is expanded to now 89 countries.
🚨 Chinese entities misused ChatGPT to spread communist ideas
📷 Runway introduced a collection of better video editing tools
🖥️ Microsoft is turning every Windows 11 into an AI powerhouse
AI CHART
Meet DeepMMSearch-R1, a multimodal LLM that searches the web, crops images, rewrites its queries, and self-corrects mid-conversation. Here’s how it works:
It decides when to search and what to search for, then issues real web queries
It can crop images automatically to zoom in on the right part before searching
It reflects on its own answers: if results don’t help, it rewrites the query and searches again
In benchmark tests, DeepMMSearch-R1 outperformed every open-source baseline:
+21 points higher than RAG workflows, which over-retrieve and add noise
+9 points over prompt-based search agents, which rely on brittle prompts
Nearly matched GPT-o3, despite being built on a smaller Qwen2.5-VL-7B backbone
The model also learned to use fewer, more relevant searches after RL training. And unlike retrieval-augmented models, it doesn’t need a massive index, it just learns to use the web intelligently.
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