🎭 AI Loves Irony

AI Academy | Advertise | 88 AI Cheat Sheets | AI Mastery AZ Course

Plus: OpenAI's o3-mini is coming

Read time: 5 minutes

Reports hint at the imminent arrival of the o3-mini reasoning model, alongside talk of advanced ‘PhD level SuperAgents’ and a groundbreaking human longevity research model. OpenAI seems poised to take AI innovation to the next level.

What are on FIRE 🔥

• 🤖 Reinforcement Learning Just Got a New Buzzword: DeepSeek-R1

• 🌟 AI Highlights

• 💸 Daily AI Fundraising

• 📚 AI Sources From AI Fire

• 🏅 AI Tools

• ⚡ 5 AI Quick Hits

• 📊 AI Chart

• 💼 4 AI Jobs

PRESENTED BY HUBSPOT

Ready to level up your work with AI?

HubSpot’s free guide to using ChatGPT at work is your new cheat code to go from working hard to hardly working

HubSpot’s guide will teach you:

• How to prompt like a pro

• How to integrate AI in your personal workflow

• Over 100+ useful prompt ideas

All in order to help you unleash the power of AI for a more efficient, impactful professional life.

Get the free guide and level up your AI game today!

AI INSIGHTS

🤖 Reinforcement Learning Just Got a New Buzzword: DeepSeek-R1

DeepSeek-AI just introduced their latest creation, DeepSeek-R1, a model they claim rivals OpenAI’s o1-1217. The twist? It’s built with reinforcement learning (RL) as the secret sauce, no supervised fine-tuning required (at least for the base version, DeepSeek-R1-Zero).

Here’s the kicker: these models supposedly show "self-evolving reasoning" and even have an aha moment mid-training. Yes, you read that right. The model learns to rethink its steps, like some digital Einstein cracking a tough math problem.

But let’s not get too carried away just yet.

The good stuff:

• DeepSeek-R1 crushed the AIME 2024 benchmark with a 79.8% pass@1, slightly edging out OpenAI-o1-1217.

• On MATH-500, it hit a ridiculous 97.3%, practically acing the test.

• It even flexed on Codeforces with a percentile score of 96.3%, outperforming 96% of human competitors.

And if you’re running smaller systems? DeepSeek’s distilled models (like the 7B and 14B versions) pack almost the same punch, showing how powerful RL-based training pipelines can shrink into leaner setups.

But, there’s always a catch.

The first iteration, DeepSeek-R1-Zero, had... readability issues. Think language mixing (hello, half-English, half-Chinese reasoning steps) and responses formatted like cryptic code dumps. So, they fixed it with a “cold-start” approach: adding high-quality, human-designed reasoning data to smooth things out.

Cue DeepSeek-R1, which combines cold-start fine-tuning with more RL magic. It’s sharper, clearer, and apparently ready to handle those tricky logic puzzles we all love (or hate).

Let’s Make Some Comparison

DeepSeek R1 crushes benchmarks and budgets. It ranks second, just below OpenAI’s o1, which costs 30 times more.

• Scores better than Claude 3.5 Sonnet and o1-mini in most benchmarks.

• Costs less than Gemini or Sonnet 3.5.

• Open weights and an MIT license for full transparency.

• API outputs usable for fine-tuning.

• Free to use on their website and app.

What they’re saying:

DeepSeek-AI is pushing hard on the “pure RL works” narrative. But let’s not ignore the whispers: RL might be a cost-cutting stunt as much as a technical flex.
And here’s the drama—DeepSeek-R1’s team didn’t stop at making one model. They distilled its reasoning skills into tiny dense versions (like Qwen-32B), which are open-sourced and designed to beat existing benchmarks. TL;DR: smaller, cheaper, smarter.

Why it matters: With OpenAI and other giants dominating the hype cycle, DeepSeek-R1 is out here claiming the next big thing in reasoning. But the real story might not be in the benchmarks. It’s in what this RL-first approach could mean for how future models learn—and whether anyone outside the lab really cares about a model’s “aha moment.”

For now, DeepSeek-AI’s gamble seems to be paying off. Let’s see if they can keep the momentum going without tripping on their RL-only narrative.

TODAY IN AI

AI HIGHLIGHTS

📱 CEO Sundar Pichai targets 500M users for Gemini AI. Despite trailing ChatGPT's 465M downloads, Google boosts features, subscriptions, and app integrations to catch up.

🤝 Perplexity AI offers to merge with TikTok U.S., letting ByteDance investors keep stakes. Talks may take months, with TikTok facing a potential U.S. ban.

🖼️ Black Forest Labs' Flux now lets users train a fine-tuned model using up to 20 images, perfect for creating consistent branding and marketing materials.

🤖 Ph.D.-Level AI Super-Agents Are Coming. Top AI firms, including OpenAI, may soon launch super-agents capable of solving complex tasks. CEO Sam Altman will brief U.S. officials on this breakthrough later in January.

🚀 OpenAI CEO Sam Altman announced the fast yet simpler o3-mini, set to release with API and ChatGPT integration. Pro models at $200/month are also in progress.

DAILY AI FUNDRAISING

Former OpenAI CTO Mira Murati, now raising over $100M for a new AI startup, plans to develop proprietary AI products.

AI SOURCES FROM AI FIRE

NEW EMPOWERED AI TOOLS

1. 🤖 GetResponse* helps you launch and grow your side-hustle with AI-driven email marketing and course creation. Build smarter, not harder!

2. 🎭 Humva offers free AI avatars for videos with perfect lip-syncing

3. 🌐 GetWebsite.Report audits websites to boost SEO and improve UX

4. 🐼 Prompt Panda lets you save and use AI prompts easily

5. 🖥️ Browser Use helps AI automate web tasks for smooth browsing

* indicates a promoted tool, if any

AI QUICK HITS

1. 🤖 OpenAI Makes ChatGPT Customization Even Simpler (Link)

2. 🦾 Bionic Hands Achieve Advanced Touch Through Neurotechnology (Link)

3. 🎮 Character AI Experiments with Web-Based Gaming Features (Link)

4. 📊 AI Grades President Trump’s Inauguration Speech Performance (Link)

5. 🏙️ AI-Driven Dialogues in ‘The Brutalist’ Stir Controversy (Link)

AI CHART

AI is taking over everything - literally

1. E-Learning

• 48% of teachers, 47% of students say AI improves learning by personalizing lessons.

• AI cuts costs and time: turns scripts into courses instantly and creates quizzes automatically.

2. HR Advantage

• 76% of HR leaders say AI adoption is critical within 2 years.

• Tools like predictive analytics speed up hiring and improve onboarding with AI video.

3. AI Video Dominance

• 55% of businesses use AI video, and the market is growing 20% annually.

• Benefits: saves time, personalizes content, and reduces costs.

• Challenges: tech expertise (30%), cost (24%), privacy (19%).

4. Consumer Trust in AI

• 54% trust high-quality AI videos, and 25% engage more with personalized content.

• Sloppy AI won’t cut it—quality matters.

5. Chatbots Are the Future

• 71% of businesses already use or plan to use chatbots for instant, 24/7 support.

6. Manufacturing Leads AI Adoption

• 90% of manufacturers use AI for supply chains, predictive maintenance, and training videos.

• Market expected to grow from $3.2B (2023) to $20.8B by 2028.

Takeaway: AI is no longer optional. It’s saving time, cutting costs, and driving growth across industries. Embrace it or get left behind.

AI JOBS

• Google: AI Sales Specialist, Startups, Google Cloud (Link)

• Zoom: AI Software Engineer (Link)

• Netflix: Software Developer for AI-Powered Content Production (Link)

• Bosch: Artificial Intelligence Innovation Leader (Link)

We read your emails, comments, and poll replies daily

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