- AI Fire
- Posts
- 🦾 Open-Source AI is Closing... The Gap
In partnership with
Plus: Microsoft Acquires TikTok?
Read time: 5 minutes
Believe it! A low-cost AI model from China just triggered a $787B market wipeout in a single day, shaking tech stocks and slashing valuations—all because investors fear tech is overvalued with DeepSeek AI's rise.
What are on FIRE 🔥
IN PARTNERSHIP WITH BUBBLE
Build your AI-powered app quickly with Bubble—no coding required. Join 500k+ users connecting to AI models like OpenAI in minutes. The easy-to-use API connector simplifies integration, and use our drag-and-drop tools, plugins and more to streamline development.
AI INSIGHTS
🤖 AI in 2024: Smarter, Cheaper, and Moving Faster Than Ever
The AI race just hit another gear, and the biggest story? OpenAI isn't the only game in town anymore.
For a while, GPT-4 reigned supreme, but by the end of 2024, competition caught up—and in some cases, surpassed it. Open-source models from Meta, Mistral, and Alibaba, Deep Seek, Qwen? No longer playing second fiddle. And OpenAI’s own GPT-4o? A clear attempt to stay ahead before things get… messy.
Here’s what you need to know:
Smaller AI models are closing in on GPT-4-level intelligence.
Inference costs dropped 100x, making AI cheaper than ever.
128k context windows are the new norm.
The US still leads AI innovation, but China is catching up.
Image, video, and speech AI have all taken major leaps forward.
AI Power Rankings Are Changing
In 2022, OpenAI's GPT-3.5 launched ChatGPT and changed the game. By 2023, GPT-4 cemented its place at the top. But in 2024? The competition finally caught up:
Meta, Mistral, and Alibaba’s models are now matching or exceeding GPT-4.
OpenAI rushed out GPT-4o to stay competitive, but it’s no longer untouchable.
The first signs of intelligence beyond GPT-4 are here (hello, OpenAI’s o1-preview).
Translation? The AI arms race is officially wide open.
AI is Getting Ridiculously Cheap
Remember when GPT-4 access cost $20/month? Now, GPT-4o-mini gives you similar intelligence for 100x less.
The trend is clear:
GPT-4 Turbo pricing dropped ~75x since early 2024.
AI inference costs are crashing—faster than anyone predicted.
Smaller models now compete with giants at a fraction of the cost.
So while AI gets smarter, it's also getting way cheaper to run—a nightmare for anyone charging premium prices for access.
Smaller Models, Bigger Brains
For years, the rule was simple: Bigger models = Better AI. Not anymore.
Smaller AI models are now punching far above their weight class, thanks to:
Knowledge distillation (training smaller models with data from bigger ones).
High-quality datasets (garbage in, garbage out—but cleaner data makes models smarter).
Compute scaling (allocating training compute more efficiently).
This means faster, cheaper AI that doesn’t need a supercomputer to run—great for businesses, bad for companies betting on “bigger is better.”
Context Windows Just Got Huge
If you’ve ever had an AI model forget what you told it five minutes ago, you’ll love this:
128k tokens is now the standard (a 32x increase from last year).
Some models (like Gemini and Nova) handle up to 2M tokens.
This unlocks better memory, deeper reasoning, and true multimodal AI.
Basically, AI can now "remember" and process far more at once—making it more useful for complex tasks like research, long conversations, and video analysis.
The Open-Source Takeover is Real
Meta’s Llama 3.1 (405B), Mistral Large 2, and Qwen2.5 are proving that open-source AI is here to stay.
The Scale Bros vs. Whale Bros debate just got real:
Scale Bros (OpenAI, Google, NVIDIA): More compute, more power, more cost.
Whale Bros (Meta, Mistral, Alibaba): Do more with less—fewer GPUs, smaller teams, and open models.
Who wins? Both sides have a point, but open-source AI is growing faster than expected—and that’s a problem for companies banking on proprietary dominance.
Why it matters: The AI Race Just Got Way More Interesting
2024 wasn’t just a year of progress—it was the year AI stopped being a monopoly.
Big models no longer have an unbeatable edge.
AI is getting dramatically cheaper.
Open-source is growing faster than anyone expected.
And context windows are making AI more powerful than ever.
This means better AI, lower costs, and more competition—which is great for you, but a headache for companies that spent billions betting on exclusivity.
So what’s next? Whoever figures out how to build the best AI experience (not just the best model) wins. OpenAI still has the ChatGPT ecosystem, but the moat is shrinking fast.
PRESENTED BY ARTISAN
Hire Ava, the Industry-Leading AI BDR
Ava automates your entire outbound demand generation so you can get leads delivered to your inbox on autopilot. She operates within the Artisan platform, which consolidates every tool you need for outbound:
300M+ High-Quality B2B Prospects
Automated Lead Enrichment With 10+ Data Sources Included
Full Email Deliverability Management
Personalization Waterfall using LinkedIn, Twitter, Web Scraping & More
TODAY IN AI
AI HIGHLIGHTS
🎥 Netflix-linked researchers built Go-with-the-Flow, an AI tool that lets users drag objects to animate them and copy motion between videos. It’s still rough but a big step for AI video editing.
🎬 Pika Labs just launched v2.1, bringing smoother motion, realistic physics, and more control over AI-generated videos. Users can now customize scenes easily, making AI video editing more powerful.
🤖 Meta AI can now remember details from 1:1 chats on WhatsApp, Messenger, and Facebook to give better recommendations. It suggests content, events, and even food based on your past chats. Available in the US & Canada.
🧬 LinkedIn co-founder Reid Hoffman launched Manas AI, an AI-powered platform aiming to speed up cancer drug discovery from a decade to just a few years. $24.6M raised to accelerate breakthroughs.
📉 Investors fear tech stocks are overvalued after the rise of DeepSeek AI, leading to a massive $787B market wipeout in a single day. Biggest losses hit AI-heavy companies.
Some people have a perspective that DeepSeek AI is a strategic move by China to disrupt the AI race.
Rumors say China can’t outspend the U.S. in AI, so they play differently. Instead of massive investments, they release a model that shakes the market and slows U.S. funding.
Some argue China has a data advantage, while the U.S. leads in money. China’s looser data laws give them better AI training, while the U.S. dominates in investment and speculation.
A planned move or just competition? Some have a perspective that DeepSeek is CCP-backed, similar to Temu and Wish, designed to undercut competitors and weaken economies. Others say it’s just normal AI progress.
What do you think? Vote below! |
DAILY AI FUNDRAISING
Earth AI, a San Mateo-based AI explorer for clean energy metals, raised $20M in Series B led by Tamarack Global. It uses AI and satellite data to discover new mineral deposits and sell mining rights.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
🚀 co.dev turns your ideas into full-stack apps effortlessly.
🔒 Llamao is a private and offline ChatGPT alternative for your device.
📱 Apollo AI runs local models like Llama directly on iOS.
💬 Yapz is an AI-powered form builder for real-time conversations.
🎯 Opencord AI finds and engages the right customers for conversions.
AI QUICK HITS
🚀 Grok-3 by xAI Spotted Early - Big AI Upgrade Coming Soon (Link)
🌍 Crypto, AI & DeepSeek: Key Trends from Davos 2025 (Link)
🩺 AI-Powered Tool Speeds Up Heart Defect Detection in Kids (Link)
📱 Apple Releases iOS 18.3 with Apple Intelligence On by Default (Link)
💰 Microsoft in Talks to Buy TikTok as Deal Discussions Heat Up (Link)
AI CHART
It was only a matter of time before Qwen2-VL got a serious upgrade, and here it is: Qwen2.5-VL. Bigger. Smarter. More capable. And ready to shake things up.
Here’s what’s new:
Sees and understands like never before – Recognizes everything from objects and charts to full document layouts.
Acts like a real AI agent – Can navigate computers and phones to complete tasks.
Handles long videos like a champ – Comprehends 1+ hour clips, finds key moments.
Pinpoint accuracy – Generates bounding boxes and JSON outputs for object detection.
Reads structured data – Parses invoices, forms, and tables like a pro.
How good is it?
Qwen2.5-VL-72B competes with top-tier models in reasoning, math, and video understanding. The 7B model? Beats GPT-4o-mini in multiple tasks. And the 3B model? Outperforms the last-gen 7B—not bad for an "edge AI" setup.
Some flexes:
Recognizes landmarks, animals, and even pop culture icons.
Detects objects with precise coordinates and labels.
Reads and understands multi-language, multi-orientation text.
Processes long videos, extracts key info, and finds exact timestamps.
Books flights, manages apps, and executes tasks on mobile devices.
Under the hood, it's got:
Better time and space perception – Adapts to dynamic image sizes and real-world scales.
More efficient visual processing – Uses smart ViT architecture to cut computing costs.
What’s next?
Qwen isn’t stopping here. The plan? Sharper problem-solving, even better reasoning, and more multimodal features. The goal? An all-in-one AI powerhouse.
Qwen2.5-VL is proving that vision-language AI is evolving fast—and OpenAI, Google, and Meta should take notes.
AI JOBS
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