- AI Fire
- Posts
- 🔇 WTF... AI Learned to “Unlearn” You!?
Read time: 5 minutes
Can AI really erase you from its memory? Yep. A new breakthrough lets models “unlearn” your voice, making them unclonable even when directly prompted.
What are on FIRE 🔥
IN PARTNERSHIP WITH SECTION
ProfAI is the fastest, easiest way to learn AI. In just 60 minutes, you’ll complete fun, interactive lessons and earn your AI proficiency certificate. ProfAI delivers role-specific AI training you'll actually be able to use - so you can start applying AI at work today.
AI INSIGHTS
Do you want AI to not sound like you? (very anti-meta). Well, that just got a LOT more possible → Instead of throwing up guardrails to stop AI from cloning your voice, their method makes the model forget you ever existed even when asked directly.
That’s what these researchers did:
Took a recreated version of VoiceBox, Meta’s powerful TTS model
Fed it 5 minutes of a voice it needed to forget
Used unlearning to erase that speaker’s “signature” from the model’s memory
Prompted it to recreate the same voice → output = totally different
State-of-the-art voice similarity tools showed 75% less resemblance to the original. Meanwhile, performance on non-redacted voices only dropped by ~2.8%.
Up till now, most AI safety in voice has been about blocking: “Don't let the user do this.”
This is different. It’s like removing the building behind the fence entirely. You can’t break in if there’s nothing to steal. Here’s a sample of what it looks like:
Original voice → cloned voice (pre-unlearning) = scary close
Same voice → cloned again (post-unlearning) = totally different, totally useless for impersonation
And randomness in the unlearned voice is very high. There’s no reverse-engineering your identity back into it.
Why It Matters: But they’ve been hesitant to release VoiceBox publicly for fear of misuse. Google DeepMind is now looking into unlearning, too. If companies can’t stop abuse with filters, maybe they need to train their models to forget you ever existed. And now… it’s technically possible.
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
🚀 A woman landed her dream job in just 3 months by turning ChatGPT into her personal job-hunting BFF for Free. She broke down what worked, what didn’t for you.
💀 Reddit exploded after users got a creepy popup asking for COM serial port access - yes, like 1990s printer hardware. Then, ChatGPT went full Shaggy mode: “Wasn't me!”
❓ Turns out, working at OpenAI is part launch-hackathon, part chaos org chart, & part X fishbowl. This blog post from a former OpenAI engineer detailed his year inside this tech giant.
✍️ If you're looking for a good AI prompting guide, we found this post where fellow users share practical tips, and they’re honestly some of the best I’ve seen in a while.
📄 Google just took away Discover clicks. Instead of article headlines, you'll see a blurb written by AI from multiple webs and stamped with their logos. It’s too little, too late.
💰 AI Daily Fundraising: AWS just doubled its investment to $200M in its Generative AI Innovation Center after helping 4,000+ customers like BMW and FOX cut costs, speed up workflows, and deploy AI in just 45 days.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
🎨 PPT.AI transforms your ideas into stunning presentations for free
⚙️ Machined V3 turns one topic into 30 rank-ready articles, in minutes
📚 AuthorAI turns a single prompt into a complete customizable book
📢 Soshi is an AI social media expert to drive high-quality engagement
AI QUICK HITS
🔥 Wow, Grok’s AI companions want to have s*x & burn down schools?
🎨 You can now create images directly inside Claude with Canva
📈 Method: ChatGPT + Grok to make trades = 100% win rate in 2 weeks
👁️ Microsoft’s Copilot Vision AI can now scan everything on your screen
🍏 Apple will buy Mistral as a budget option compared to Anthropic
AI CHART
Another walled garden just got some fresh air. French startup Mistral just dropped Voxtral, its first open-source audio model → built to take on the closed systems from OpenAI, ElevenLabs, and Google.
Mistral’s pitch? Developers no longer have to pick between cheap but dumb, or powerful but locked-down. It comes in 3 flavors:
Voxtral Small: 24B parameters, built for full-scale products. Targets ElevenLabs Scribe, GPT-4o-mini, Gemini Flash.
Voxtral Mini: 3B parameters, optimized for edge devices and offline use.
Voxtral Mini Transcribe: API-only, barebones transcription mode that undercuts OpenAI Whisper, at less than half the cost.
You can run it through Hugging Face or Mistral’s Le Chat chatbot. API pricing starts at $0.001 per minute. And because it’s built on Mistral Small 3.1, you can:
Transcribe up to 30 minutes of audio
Understand and summarize up to 40 minutes
Ask questions, trigger actions, and get actual voice-level reasoning
Do all of the above in 8 languages (EN, FR, DE, ES, PT, NL, IT, HI)
Voxtral proves that open-source speech AI is finally good enough to use in production. This comes just weeks after Magistral, Mistral’s reasoning model for chain-of-thought reliability. You can feel it.
→ They’re reportedly raising $1B from Abu Dhabi’s MGX fund to keep going.
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