In partnership with
Could Anthropic rebuild Mythos using just a fraction of its original computing power? Former OpenAI researcher Jacob Coxon thinks so, and the chart behind his claim helps explain why he wants frontier AI development paused in 2027.
What's on FIRE 🔥
IN PARTNERSHIP WITH HUBSPOT
Turn AI into Your Income Engine
Ready to transform artificial intelligence from a buzzword into your personal revenue generator
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
AI INSIGHTS
This slide from Anthropic's IPO roadshow shows the "Pre-training compute multiplier" for code over time. Anthropic’s training efficiency chart uses a log scale. Jacob Coxon sees accelerating progress, though the data is illustrative and based on internal estimates. | Image: Anthropic
Anthropic’s IPO roadshow included a chart that former OpenAI and Anthropic researcher Jacob Coxon says helped convince him to leave. His claim: Anthropic could already retrain Mythos using “a fraction of the compute originally needed.”
The chart shows how its training efficiency for code has improved over time. Here’s why it caught his attention:
The gains appear to be accelerating. The curve bends upward even on a logarithmic scale, which compresses large increases.
Better methods mean less compute to reach the same capability, or more capability from the same hardware.
Coxon thinks these improvements could soon be automated, helping AI development move even faster.
He wants frontier development paused in 2027, arguing that competitors could catch up and make human-level AI widely available.
The catch: Anthropic calls the chart illustrative and based on internal estimates. It gives no exact Mythos savings and doesn’t prove autonomous self-improvement. Still, it raises a big question: how fast could progress move once Claude helps improve the next Claude?
PRESENTED BY THE CODE
The Agentic Engineering Playbook top engineers are using
Most engineers are stuck babysitting AI one prompt at a time. A handful have figured out how to build agents that plan, execute, and self-correct on their own. Sign up for The Code and get The Complete Playbook for Agentic Engineering. Claude Code, subagents, MCP, all of it.
AI SOURCES FROM AI FIRE
1. Fable 5.2 Takes On 3 Wild 3D Design Challenges. Build interactive worlds in Claude Code, from a floating city to a mechanical octopus. Get the full prompts and testing checklists to try each challenge yourself.
2. 15 Claude AI Design Tricks That Make Your Websites Look More Professional. Turn generic designs into websites worth showing off. Learn how to use better references, compare 6 design directions, and polish the details before you publish.
3. Run Claude Code with Free AI Models Without a Claude Subscription. A step-by-step guide to connecting Google, NVIDIA, and OpenRouter through Free LLM API. Combine available free quotas and switch models when credits run out, within each provider’s limits.
FIRE RECAP: BIGGEST AI NEWS THIS WEEK
🚨 Jensen Huang resigned as Nvidia CEO, effective immediately. CFO Colette Kress takes over as interim CEO while the board searches for a replacement.
🧠 Anthropic launched Claude Fable & Mythos 5.2, improving coding, reasoning & vision. Outputs are up to 30% faster, with lower API prices. Mythos requires approval.
🔎 DeepSeek’s app features are now available via API, including search, files & URL reading. Speech APIs and an open-source chat demo are available too.
⚡ Google launched Gemini 4 Hydrogen for faster coding and agent workflows. It handles debugging & multimodal tasks, with four thinking levels developers can choose from.
💳 OpenAI launched ChatGPT Wallet, letting dots agents pay without seeing your card details. Each purchase needs approval and a security check, starting with eligible U.S. Pro users.
📊 Claude Sheets & Whiteboards entered beta for shared spreadsheets and brainstorming. Claude Dashboards becomes Claude Data, now generally available. All three are on paid plans.
TODAY IN AI
AI HIGHLIGHTS
🧠 Google’s Antigravity /teamwork helped researchers solve long-standing math problems and speed up code with Gemini 4 agents. Google is also offering 10,000 academic researchers six months of free access to its highest usage tier.
💳 Gemini now connects to Google Wallet to track spending, spot forgotten subscriptions & help plan budgets. Rolling out to all U.S. Gemini users. It can read linked account data, but can’t move your money.
👀 Elon Musk backed a petition to bring Jensen Huang back as Nvidia CEO, alongside Replit’s Amjad Masad. Organizers report 100,000+ signatures, though they’re still removing fake entries.
🤖 OpenAI disbanded its stand-alone robotics team, moving staff into its main research and applied teams, Fortune reports. The company says GPT-6 Astra can handle some robot tasks directly, driving the shift.
🛡️ Anthropic shared a guide to better AI safety tests, warning that scores can shift with question wording or test settings. Its advice: check AI graders against human judgment and report uncertainty, not just headline numbers.
💰 AI Daily Fundraising: Mercor raised $750M at a $20B valuation, with NVIDIA, Salesforce, and Zoom joining the round. The funding will help companies train and test AI agents on their own data, using human expertise to improve performance on real work.
NEW EMPOWERED AI TOOLS
🎮 Wanaka Studio lets anyone build, edit, and publish 3D games with AI, featuring a visual editor, multiplayer support, and one-click publishing.
🌐 Tobira Site Agent turns websites into AI sales agents that answer questions, book calls, and take orders through ChatGPT, Claude, Grok, and other AI chats.
🤝 Teamily AI brings people and AI agents into one workspace with shared memory, messaging, meetings, and automated workflows.
🔌 Scalekit connects AI agents to 30,000+ tools across 600+ apps through one MCP gateway, with secure authentication and permission controls.
📱 Glasshand lets Claude Code, Codex, and other AI agents rent real iPhones and Macs to test apps and complete tasks on physical devices.
AI BREAKTHROUGH
Spending MORE on AI coding tools doesn't mean you're getting MORE work done. Carnegie Mellon researchers tracked 404 software engineers over 5 weeks, comparing how many AI tokens they used with how much code they committed.
2× more AI token spending was linked to only about 10% more code commits for the same engineer.
Engineers producing similar numbers of commits sometimes used 39–240× different amounts of tokens.
Nearly 90% of technical respondents' agent evaluations involved tasks beyond coding, including research, analysis, and planning.
Traditional work metrics explained only 13% of an engineer's week-to-week token spending differences.
The problem is that some companies track AI token usage to measure employee productivity. But higher spending can simply mean more AI conversations, longer prompts, or repeated attempts to solve a problem.
So spending more on Claude Code, Codex, or other AI coding agents doesn't automatically mean better results. Companies should measure completed tasks, working features, and code quality instead.
The Dell Pro keeps your day moving.
Built for those who stay in the flow.
Handle the busiest days by staying in the flow with built-in intelligence on the Dell Pro laptop powered by Intel® Core™ Ultra with Intel vPro®.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?
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
