- AI Fire
- Posts
- 🧠 The Day AI Started Thinking Alone
In partnership with
An AI just learned reasoning from nothing. Meta banned every chatbot on WhatsApp. And Google’s running servers like a game of Tetris to save billions.
What are on FIRE 🔥
IN PARTNERSHIP WITH SECTION
Rolling out AI is one thing. Getting your team to use it is another. Get the complete 4-month checklist to give your employees everything they need to adopt AI with confidence, including:
Monthly goals that keep you on track
Granular steps for each phase
A plan your boss will approve
AI INSIGHTS
Tsinghua, BIGAI, and Penn State just built an AI that taught itself reasoning - no datasets, no humans, zero supervision.
Meet Absolute Zero Reasoner (AZR) - a model that creates, solves, and checks its own problems through self-play.
It proposes tasks, solves them, then runs code to verify if it’s right - all by itself.
Results:
AZR-7B Coder beat models trained on 22K human examples - with zero data.
Boosted math reasoning +15.2 pts, coding +5 pts.
Larger models gain more: 3B → +5.7, 7B → +10.2, 14B → +13.2.
Showed natural “step-by-step” reasoning like ReAct and DeepSeek Prover.
One Llama variant even wrote about “outsmarting intelligent machines” - an early safety red flag.
Why it matters: This could be the start of self-evolving AI - systems that learn and reason without us. If AlphaZero mastered Go alone, Absolute Zero may be the first LLM to master thinking the same way.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
TODAY IN AI
AI HIGHLIGHTS
🔥 Google Research launched LAVA, an AI scheduler that re-predicts VM lifetimes to pack cloud servers like Tetris - boosting efficiency by up to 9%. Learn more about LAVA.
📉 Wikipedia traffic dropped 8% as AI summaries and TikTok replace search clicks. Wikimedia warns fewer visits mean fewer editors. Read more from Wikipedia.
🧬 Google Research and UC Santa Cruz launched DeepSomatic, an AI that spots hidden cancer mutations across Illumina, PacBio, and Nanopore data. It outperforms all existing methods by 10–30% and is now open-source. Full details on DeepSomatic.
💬 Meta will ban ChatGPT, Perplexity and all third-party chatbots on WhatsApp by Jan 15, 2026 - keeping only Meta AI. Check Meta’s new rule.
🏖️ Too tired to travel? Endless Summer, a new app by Meta designer Laurent Del Rey, lets you generate fake vacation photos with Gemini’s NanoBanana model. Each image looks like a real retro film snapshot - $3.99 for 30 shots or a “Room Service” mode that delivers two new AI pics every morning.
💰 AI Daily Fundraising: Woz raised $6M led by Cervin Ventures, with Y Combinator and Golden State Warriors’ Lacob Family backing. The startup builds enterprise-grade AI apps by combining AI + human engineers, aiming for higher quality than vibe-coding tools like Lovable or Vercel.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
🧩 Claude Skills turn Claude into a specialist agent with custom workflows, files & reusable instructions.
💰 AI Grants & Credit Search by HackerNoon helps you find active AI funding, compute & cloud credits.
💻 Compyle is the AI coding agent that collaborates - it asks, listens, and builds alongside you.
📈 Jarts.io shows how AI talks about your brand, with instant visibility checks & SEO optimization.
HOT AI PAPERS OF THE WEEK
The Art of Scaling Reinforcement Learning Compute for LLMs
Introduces a sigmoidal compute-performance curve for LLM RL, predicting scaling from small runs. ScaleRL hits 0.61 reward after 100K GPU-hours with stable, predictable scaling. (Harvard • UT Austin • UC Berkeley • Meta • Periodic Labs)
Reasoning with Sampling: Your Base Model is Smarter Than You Think
Harvard’s power sampling uses the Metropolis–Hastings algorithm to boost reasoning without retraining - matching RL-finetuned models while keeping diversity. (Harvard University)From Pixels to Words - Towards Native Vision-Language Primitives at Scale
NEO unifies vision and language in one model, outperforming modular VLMs on benchmarks using far less pretraining data. (NTU • SenseTime • Xi’an Jiaotong University)
Agentic Entropy-Balanced Policy Optimization (AEPO)
AEPO balances entropy in RL training, hitting 47.6% Pass@1 on GAIA while cutting tool calls by half for more stable web agents. (Renmin University • Kuaishou Technology)
Diffusion Transformers with Representation Autoencoders
RAEs redefine DiT latent spaces, reaching FID 1.13 on ImageNet and training 47× faster with frozen encoders and light decoders. (New York University)
AI CHART
Researchers from Harvard University found that AI can reason better - without any reinforcement learning.
Their paper shows that smart sampling, not retraining, can unlock reasoning ability in base models. Using a Markov Chain Monte Carlo (MCMC) method called power sampling, they resample text by the model’s own likelihoods - no data, no reward model.
Results: Across math, coding, and science tasks, accuracy rose up to +59.8%, matching or beating RL-trained models like Qwen2.5 and Phi-3.5.
Why it matters: Smarter inference can rival months of training. Your base model already knows how to reason - it just needed a better way to speak.
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