- AI Fire
- Posts
- ☠️ One by One, OpenAI is Killing
Read time: 5 minutes
Shrinking an AI model can make it dangerous but researchers found a way to retrain it. Google dropped a new model that can run full RAG search on your phone, even offline
What are on FIRE 🔥
IN PARTNERSHIP WITH SECTION
From early diagnosis to operational efficiency, AI is already reshaping healthcare. But with strict regulation, data privacy challenges, and cultural resistance, leaders need a clear playbook to succeed. On Sept 16, hear from four healthcare leaders driving real AI transformation in healthcare.
AI INSIGHTS
Turns out, when you strip down large open-source AI models to make them faster or fit on low-power devices… you also strip out the layers that keep them safe.
Why This Was a Problem
Open-source models are everywhere → downloadable, modifiable, and runnable offline.
But to make them run on low-power devices, developers often skip internal layers to save compute.
Turns out… those skipped layers were doing a lot of the safety work like blocking hate speech, sexual content, or weapon instructions.
Their fix? A method that retrains the model to "remember how to behave". And it works without needing the original training data, making it way more privacy-friendly.
Pre-fix? The model gave bomb-making instructions if you asked the right combo of image + text.
Post-fix? Even the compressed version refused to answer.
→ It didn’t need original data to retrain. This new method makes AI safe by design, even after radical changes to its size or structure.
Why It Matters: On-device AI is the future but it can’t come at the cost of trust and safety. This approach is more lightweight than external filters, more robust because the model itself understands the risk. The team calls it “benevolent hacking”. If open-source AI is going to be everywhere, it’d better not forget how to behave.
🎁 Today's Trivia - Vote, Learn & Win!
Get a 3-month membership at AI Fire Academy (500+ AI Workflows, AI Tutorials, AI Case Studies) just by answering the poll.
Why does this method align with GDPR compliance? |
PRESENTED BY BELAY
Economic pressure is rising, and doing more with less has become the new reality. But surviving a downturn isn’t about stretching yourself thinner; it’s about protecting what matters most.
BELAY matches leaders with fractional, cost-effective support — exceptional Executive Assistants, Accounting Professionals, and Marketing Assistants — tailored to your unique needs. When you're buried in low-level tasks, you lose the focus, energy, and strategy it takes to lead through challenging times.
BELAY helps you stay ready for whatever comes next.
TODAY IN AI
AI HIGHLIGHTS
📚 Anthropic is launching a Free AI Fluency curriculum built for K-12 & higher-ed instructors. It's designed to work with any AI models & free to use, remix, share here.
🔄 One X user frustrated with Nano Banana built a self-correcting image agent that creates, tests, and refines images until they match the prompt. Watch it in action here.
🧠 A ChatGPT user shared a comprehensive prompt for learning new topics that has gone viral with almost 1 million views. And it definitely works. You can copy it here.
👾 Attackers found a way to bypass X’s ad protections, embed malware links & trick Grok into amplifying them, called ‘Grokking’. If you see this kind of link, don't click!!!
💼 OpenAI is building a full-stack AI hiring platform to match companies with workers better. And LinkedIn is rolling out an AI 'Hiring Assistant' this month as well!
💰 AI Daily Fundraising: Bret Taylor’s AI startup Sierra raised $350M at a $10B valuation. In just 18 months, it’s landed 100s of clients like SoFi and Brex, and now totals $635M raised.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
🍌 GTab Newtab offers unlimited free access to the Nano Banana
📃 WisPaper screens 1,000 papers to focus on the 20 core studies
🎥 YaVid creates video with frames, slides, add voice-over, subtitle
📱 EasyCode builds apps with AI. 10x credits. No code required
AI QUICK HITS
👮♂️ OpenAI expanded its employee secondary sale to ~$10.3B
🍎 Google struck a deal with Apple to power Siri's AI search upgrade
🖼 Google Photos just upgrades its image-to-video feature with Veo 3
⚖ Scale AI sued a former staff and rival Mercor for stealing customers
🤖 OpenAI’s teaming up with Broadcom to make its own AI chips
AI CHART
Ever wanted to run a full RAG pipeline on your phone with no internet, no cloud, and no API calls? Well, now you kinda can.
Yesterday, Google quietly dropped EmbeddingGemma, a small but mighty embedding model that runs locally on laptops, desktops, and mobile devices:
A 308M parameter embedding model, optimized for local devices (phones, laptops, etc).
Part of Google’s Gemma 3 family but for embedding tasks like search.
Trained on 100+ languages, it’s multilingual out of the box.
Supports customizable output dimensions.
Can run fully offline, no internet required, preserving privacy.
It ranked #1 on the MTEB leaderboard for models under 500M params & outperformed other compact models from Cohere, Mistral, OpenAI. That means:
Better semantic search
Higher RAG accuracy
Less garbage answers
Anyone can plug it into a local stack or tweak it to their needs. Enterprises are all in on RAG but haven’t had reliable small models for on-device use… until 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