- AI Fire
- Posts
- 📊 Why 99% Aren’t Seeing AI’s Power?
In partnership with
Ever wonder why people disagree so much about AI? Karpathy just breaks down all the 4-tier AI gap. Turns out, we're experiencing very very very different AI.
What's on FIRE 🔥
IN PARTNERSHIP WITH CLOUDTALK
Start from a ready-made template and get a working voice agent in minutes.
Want it to qualify leads harder or route calls differently? Tell the AI Composer in plain English. Plus a first look at what's coming in Q4 and live Q&A.
October 20, 4:00 PM CET.
Free to join.
AI INSIGHTS
Karpathy's new post is the best explanation I've seen of why people disagree about AI. It’s because they are effectively experiencing very different levels of AI capability:
~6B people (~75%) have barely interacted with LLMs.
~1–2B (~20%) occasionally use free chatbots, often as a better Google or writing assistant.
~20M (~0.2%) professionally use frontier models, especially for math and coding.
Only ~5,000 people may have access to the most advanced frontier systems.
Karpathy says 3 things create this gap:
Short questions barely stress frontier models. Their advantage becomes clearer on difficult projects that run for hours or days.
AI is strongest in digital, verifiable domains such as coding and math, where outputs can be tested and reinforcement learning has clear rewards.
Free, paid, and internal systems can offer dramatically different levels of capability.
If someone only tests AI with small questions in a free chatbot, they may be judging a completely different technology from the one power users experience with frontier coding agents on large, verifiable projects.
PRESENTED BY HIGHTOUCH
Most prompting advice stops at "be specific."
What’s the key to maintaining brand integrity at scale? A lot goes into it, but the teams that get it right have one thing in common: they know how to talk to AI models.
As teams speed up and the required creative volume balloons, communication matters more than ever— both with LLMs, and with the people around you. Creative teams need to focus on ideas, concepts, craft, while technology augments their manual execution demands.
Join our discussion on this evolution with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.
ALL YOU NEED TO WIN WITH AI
1. Free Guide: GPT-6 Astra Gets 10X More Productivity When You Give It This One Type of Data. In about 10 minutes, you can build a system that spots your best work hours, bad break patterns, focus drops, and the task order that actually helps you get more done.
2. Will AI Automation Still Make Money in 2027? Here's What I'd Learn Instead to Stay Ahead. Everyone's learning & building agents, or selling workflows. But will those skills still be valuable in 2027? I've been looking at which skills could become more valuable...
3. GPT-6 Astra + Fable 5.1 Together= AI GOD MODE to Save You 20x More (90% Cost-Saving). Don't treat GPT-6 Astra and Fable 5.1 like competitors, stop choosing one and combine them. This is the setup that many of you are overlooking.
CLAUDE CODE MODS: FULL GUIDE
How to build your own Mod step by step
Mods let you change Claude Code's behavior without waiting for Anthropic to release new features. You can create your own or ask Claude Code to build one for you. For example, a mod can:
Customize the interface: Add dashboards, new panels inside Claude Code.
Control tool actions: Block commands, automatically retry failed actions.
Improve privacy: Remove sensitive information before Claude reads it.
With Claude Code Mods, you can add custom features to Claude Code Desktop just by describing what you want. You don't even need to know how to code.
I'll walk you through 5 Claude Code Mods that are actually worth trying. What's covered:
What Claude Code Plugins and Mods actually are, and how they differ
How to build your own Mod step by step
4 Mods worth installing, with honest notes on what each one does well
A quick comparison table to help you choose where to start
TODAY IN AI
AI HIGHLIGHTS
🔄 Codex can now use a reset when you explicitly ask or when you’re almost out of limit to 10% or less. Nothing happens automatically, so you still stay in control.
💼 A former McKinsey analyst coined a new AI-era role: Marketing Engineer. It blends marketing, coding, and AI to drive bigger results. Here’s how to build the skill set.
🛡️ Anthropic launched OSS Scanner, a free service that uses its strongest models, including Mythos, to find vulnerabilities in open-source projects and suggest fixes.
📧 Grok Bot users can now give their bots their own email addresses. Bots can use them to communicate, sign up for newsletters, handle simple inbox-based tasks.
🛑 Satya Nadella says AI needs an “emergency brake” that lets authorized humans pause models mid-task. He wants tamper-proof logs for every meaningful action.
🏠 Apple has tried the smart home before, but HomeView sounds much more serious. A Siri screen, A18 chip, and wall-mounted option? Two versions are expected soon.
💰 Big AI Acquisition: Amazon is discussing a $7B deal for Decart, whose software claims to make AI models run up to 8x faster across Nvidia, Google, and Trainium chips. The bet: cheaper AI compute.
NEW EMPOWERED AI TOOLS
💼 HireKey helps you land tech jobs by tailoring your resume, tracking applications, and practicing interviews with an AI voice coach.
🎯 kōdlo finds your ideal prospects to create personalized sales messages, offers, and pitch decks, all in one workspace.
⚙️ Perplexity Computer Automations handles recurring work automatically, from compiling reports to updating trackers.
👻 Ghost Core runs AI directly on your computer, connects your apps, files, and devices, and learns your context to help you proactively.
AI BREAKTHROUGH
Sakana AI just introduced MASS (Multi-Agent Self-Supervision), a method where one AI model improves itself by learning from a team of copies of itself. Main points:
One shared model plays 3 roles: task solver, workflow optimizer, and evaluator.
It creates teams of virtual agents, searches for better ways for them to collaborate, then trains on the best team’s executions.
Sakana tested MASS for two self-improvement cycles using a 27B open-weight model on synthetic research tasks.
After the second cycle, score per output token improved to 1.2–1.6× the base model across 4 research benchmarks.
Multi-agent training data proved more efficient than trajectories from a single agent.
Improving the task-solving model also unexpectedly improved its ability to act as the optimizer and evaluator.
Researchers found that the optimizer can become the bottleneck: the system may know which result is better but struggle to invent a better workflow to reach it.
MASS does not rely entirely on external judges during training. Sakana only used GPT-5.5 and Claude Opus 4.8 afterward as independent evaluators to check whether the self-reported improvement was real.
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