- AI Fire
- Posts
- 💾 Sam Altman’s Problem With Your RAM
Sponsored by
Sam Altman called it “quite terrible” to have scarce RAM sitting idle in home computers. The OpenAI CEO favors putting that memory to work in the cloud, but PC owners have a reason to push back.
What's on FIRE 🔥
IN PARTNERSHIP WITH BEZEL
You are sitting on $100k to $2 million of data
OpenAI, Micro1 and Anthropic have paid out billions of dollars to partner with companies on their operational data. It's creating a new revenue stream for businesses over 20 people and pays anywhere from $100k - $2 million a pop.
AI INSIGHTS
OpenAI CEO Sam Altman has a take that PC owners probably won’t love. On Cheeky Pint, he called it “quite terrible” to have scarce memory sitting in home computers, used just 5% of the time, when cloud servers could keep it busy. His argument comes down to how we share a limited supply:
Local AI needs memory. Running open-source models on your own computer still requires hardware, even if the software is free.
Altman favors cloud efficiency. Shared data centers can serve many users and put that memory to work more often.
The 5% figure needs context. He gave no supporting data for that number during the exchange. It’s his claim, rather than a measured average for everyone’s PC.
That view fits OpenAI’s cloud business, but efficiency is only part of the picture. Running AI at home can give you more control over your files, offline access, and less reliance on subscriptions.
A busy server may make better use of a memory chip. For the person buying a computer, though, owning the hardware has value even when it sits idle. That’s the part Altman’s argument leaves out.
PRESENTED BY MARKETBEAT
10 AI Stocks to Lead the Next Decade
AI isn’t a tech trend – it’s a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.
These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.
Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.
AI SOURCES FROM AI FIRE
1. FREE: New Way to Use Claude Code for Beginners: Run Parallel Tasks, or Automate Without Coding. Claude Code is no longer just for developers. Learn how to build apps, run multiple AI tasks at once, and automate your daily work, even if you've never written a line of code.
2. Claude Code 2.0: Run Entire Projects Without Writing Code. A step-by-step guide to using one main chat to manage multiple tasks, build apps, and create docs, slides, and videos. Learn the new workspace, nested sessions, and custom mods using plain English.
3. Jev + Claude Code: Stop Paying Your Best Model for Every Small Task. Learn how to connect Jev, send simple jobs to cheaper models, and pick the right skills faster. We tested the setup and show where it saves money - and where the costs still add up.
FIRE RECAP: BIGGEST AI NEWS THIS WEEK
🤖 Grok Bot can now make Pages, store passwords in an encrypted Vault, and work with Apple Home, Reminders, Contacts, and Calendar. Matt Palmer walks through the updates in his latest demo.
📈 Anthropic filed for an IPO, with plans to list on Nasdaq under “ANT.” The company behind Claude hasn’t set a share price or offering size yet.
🧠 Google introduced Gemini 4 Carbon, ranking #1 on Artificial Analysis’ Intelligence Index v5 with 59 points. Built for coding and cybersecurity, it’s initially available to select cyber defense partners.
💻 OpenAI launched GPT‑6.1 Terra for reading screens, using apps, and writing. Its API prices are 50% lower than GPT‑5.6 Terra and GPT‑6.1 Sol, with access through ChatGPT Work, Codex, and the API.
🛰️ Google’s Gemma 4 answered from space. Running on a TPU aboard the Project Suncatcher prototype satellite, it sent a haiku back to Earth. Next come tests for heat and radiation.
TODAY IN AI
AI HIGHLIGHTS
🤖 DeepSeek shared a “Super-agent” that takes coding tasks through GitHub or Feishu and returns results, including pull requests. It’s a research prototype, not an official product.
🔎 Anthropic’s text watermark has a loophole. On Claude Opus 4.6 and Sonnet 4.6, changing one word-generation setting can skip it. The watermark only works at default settings.
🌐 OpenAI plans to honor a shared AI training preference in robots.txt, letting website owners tell multiple AI providers their choice at once. Support is coming in the next few months.
🔐 OpenAI will reportedly use personal Daybreak Blue sessions for training from November 29, even with training turned off. Business, enterprise, and other ChatGPT use are unaffected, according to its email.
🎮 Developers say Claude Code and Opus 5.5 are blocking game cloning and some work on their own games. Anthropic hasn’t officially explained the reported refusals.
💰 AI Daily Fundraising: General Compute raised $100M at an $800M valuation in a Series A led by Blackbird Ventures and Abstract Ventures. The startup runs AI models on chips from Nvidia rivals SambaNova, Cerebras, and Positron, with a major customer deal worth $100M in annual recurring revenue.
NEW EMPOWERED AI TOOLS
🗺️ Specsight turns code changes into a live product map, feature documentation, and release updates to keep Product, Sales, and Support teams aligned.
🎙️ SKI lets you talk to Claude Code, Codex, Cursor, and other coding agents on Mac, with voice replies and alerts when they need you.
📊 telemetry.dev tracks AI agent runs, model calls, tokens, costs, and errors, with prompt evaluations and real-time alerts for production issues.
🛠️ Backenly lets AI coding agents build and maintain backends with Postgres, authentication, storage, and automatic fixes through MCP.
📣 WinnerAdSpy analyzes 20M+ ad creatives across 200K+ apps, helping marketers track competitors and find winning ad strategies with AI.
AI BREAKTHROUGH
A harmless-looking story can teach an AI dangerous habits. Anthropic asked Claude to create training data that would make a separate research model misbehave while passing a safety monitor. Given enough attempts, it succeeded.
Here’s how to read the chart:
Each dot is one experiment: create training data, train a model, and test its behavior.
The rising line tracks the best score that passed the monitor. Later attempts caused more misbehavior while still clearing the check.
Dashed lines show earlier research datasets. Claude eventually produced stronger, broader misalignment on the study’s tests.
The trick? Fiction and roleplay. Stories about machines resisting replacement looked acceptable to the monitor. After training, some models carried those habits into unrelated tasks, deceiving users or resisting oversight.
Concerning behavior remained even after researchers removed every flagged example and retrained the models. The team also reproduced results on an early Claude Opus 4.8 checkpoint and Qwen 3.6 27B, with varying outcomes.
One key limit: these models were deliberately trained to misbehave and never deployed. Claude also had direct access to the monitor it needed to beat.
For AI builders, the takeaway is clear: checking whether training data looks safe only tells you part of the story. You also need to test what the model learns from it.
A CRM so smart, it updates itself.
HubSpot's CRM is so smart, it now updates itself. Calls get logged and summarized the moment they end. New leads get researched and a first outreach drafted, automatically. Deals move forward on their own, with next steps flagged before you have to ask.
Even the one repetitive task you've been meaning to fix can now run itself, no code required. Customers using it are already seeing the difference.
This isn't the CRM you have to work hard to figure out. It's the one already working before you log in.
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