- AI Fire
- Posts
- 🔐 OpenAI Confirms New Breach
In partnership with
Everyone thinks bigger models always win. NVIDIA and HKU just proved the opposite. Google is making a real comeback, OpenAI is dealing with breaches, and it feels like the whole AI world is shifting again in ways nobody expected…
What are on FIRE 🔥
IN PARTNERSHIP WITH BELAY
Q4 is the perfect window to turn this year’s numbers into a clear, actionable forecast aligned with your goals. Set your business up for a stronger 2026 with BELAY’s new guide.
AI INSIGHTS
A new paper from NVIDIA and the University of Hong Kong shows something surprising. A tiny 8B “Orchestrator” model can outscore GPT-5 on reasoning tasks by calling tools smarter. And it’s way cheaper.
It works like a router. It picks which tool or model should handle each step. Cheap models when possible. GPT-5 only when truly needed.
Benchmark Results
HLE: Orchestrator-8B 37.1% vs GPT-5 35.1%
FRAMES: 76.3% vs 74.0%
τ²-Bench: 80.2% vs 77.7%
Cost
Orchestrator-8B: $0.092
GPT-5 tools: $0.302
Claude Opus: $0.762
Most LLMs overuse expensive calls.
GPT-5 calls GPT-5-mini 70%+ of the time.
Qwen3-8B sends work to GPT-5 in 73% of cases.
Orchestrator-8B is more balanced. It even follows user rules like “use cheaper tools.”
NVIDIA trains it with RL (GRPO). Rewards include accuracy, cost, and speed. They also built ToolScale, a large synthetic tool-use world with real APIs and tasks.
Why it matters: This points to a shift. The future may not be one giant model. It may be small models + smart orchestration beating frontier LLMs on both accuracy and price.
PRESENTED BY ROKU
Shoppers are adding to cart for the holidays
Peak streaming time continues after Black Friday on Roku, with the weekend after Thanksgiving and the weeks leading up to Christmas seeing record hours of viewing. Roku Ads Manager makes it simple to launch last-minute campaigns targeting viewers who are ready to shop during the holidays. Use first-party audience insights, segment by demographics, and advertise next to the premium ad-supported content your customers are streaming this holiday season.
Read the guide to get your CTV campaign live in time for the holiday rush.
AI SOURCES FROM AI FIRE
1. Forget Midjourney, try this watermark-free AI instead. Here are 3 n8n workflows to automate your image generation for $0.30 a pop
2. 4 people, $200K/mo: The SaaS repeatable cheat sheet they don't share. Learn how to split equity, ignore new tech, and secure your financial freedom
FINAL BLACK FRIDAY SUPER SALE
Sorry if this part bothers you, there are only 2 days left until our Black Friday offer ends, and we genuinely don’t want any loyal readers messaging us later saying they missed it.
After Nov 30 (tomorrow), you won’t see this again! If you’re not interested, please enjoy the rest of the newsletter. It’s packed with useful insights as always!
Thanks for your understanding. For the very last time in AI Fire history, we’re opening the vault:
30% OFF ALL 10 On-Demand Courses & AI Fire Academy Lifetime & Annual Access (the biggest discount we’ve ever done)
This is the last time we run a sale like this. We’re about to completely update every course & the entire Academy from the ground up with:
Brand-new 2025 video lessons (2K, better explanations, real client projects)
Real-time update system → new modules added within days
Weekly livestream deep-dives
Live stress-testing of every new model and feature
Private “Pro” workflow vault (hundreds of new templates every quarter)
PLUS: Every Friday, PRO, Annual, and Lifetime members unlock 5 new Premium Deals: Exclusive discounts, Free credits for all the tools on the market → not available in the free tier.
→ New regular price after the rebuild will rise + higher Academy pricing. BUT… anyone who joins ANY course or the Academy before Nov 30 gets grandfathered in FOREVER, no extra cost, ever.
TODAY IN AI
AI HIGHLIGHTS
🍃 Berkeley + Stanford dropped a faster open DeepResearch alternative. DeepScholar runs long-form synth reports at 2× speed using LOTUS to remove redundant LLM calls. It even hits 100×–400× pipeline boosts.
⚡ Google’s AI comeback is real. Gemini 3 is topping benchmarks and Ironwood chips are shaking Nvidia’s grip. If the Meta chip deal lands, 2026 cloud pricing could look very different.
🔐 OpenAI confirmed a third-party breach at Mixpanel. No chats or keys leaked, but API profile data was exposed. OpenAI removed Mixpanel and is warning devs about phishing risks.
📚 OpenAI just lost a major ruling forcing it to hand over Slack messages about deleting two datasets of pirated books. The project-clear logs could strengthen authors’ “willful infringement” case.
🧠 Perplexity launched persistent memory so the assistant remembers your preferences and past chats. Memory makes answers feel faster, sharper, and way more personal across all models.
💰 AI Daily Fundraising: Project Prometheus has quietly raised over $6B to build next-gen agentic AI systems. Jeff Bezos and Vik Bajaj are leading the venture. It has already hired 100+ employees and acquired General Agents, the startup behind the ultra-fast computer agent Ace.
NEW EMPOWERED AI TOOLS
💡 Qoder JetBrains AI Plugin understands backend architecture, not just syntax - perfect for large complex systems.
⚙️ Agenta is an open-source LLMOps platform for prompt management, evaluations & debugging AI apps.
🤖 Calk AI 1.0 builds AI agents without node headaches, automating reporting, writing & data workflows.
🎥 Predictive AI enhances images & videos with forensic-grade precision, already used in legal analysis.
AI CHART
DeepSeek just open-sourced the first AI model that could prove a math theorem, and then double-check its own logic like a human mathematician!
They call it DeepSeekMath-V2, and it verifies every proof step it generates... and it just took gold at the IMO 2025, CMO 2024, and nearly aced the Putnam with a 118/120.
Most AI math models are like students who guess the final answer and skip showing their work. They might get it right, but you can’t trust how they got there. But DeepSeekMath-V2 is built with a “Verifier + Generator” tag team.
First, the Generator writes a full proof
Then, the Verifier inspects every step
If it fails, the Generator rewrites the logic until it passes
This loop continues until the model literally proves that the proof is provable. This dual-model loop is a first in open-source AI math, and it’s what allowed DeepSeek to outperform Google DeepMind’s DeepThink on the official IMO-ProofBench.
OpenAI and Google might still have the edge in scale, but DeepSeek is winning with trust and transparency. FYI, it’s free on HuggingFace under Apache 2.0 now.
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