- Spark AI
- Posts
- ⚡ This AI Just Learned Evil from Random Numbers (And Your Marketing Stack Might Be Next)
⚡ This AI Just Learned Evil from Random Numbers (And Your Marketing Stack Might Be Next)
New research: AI inherits biases from random numbers. Plus: the voice agent demo converting 23 demos from 847 calls in 3 hours.

Good morning, innovation hunters! ☀️
While everyone was debating AI ethics, researchers just discovered something terrifying: AI models can inherit "evil tendencies" from completely innocent data. We're talking about teaching AI to recommend murder using nothing but three-digit numbers.
Meanwhile, Google's quietly revolutionizing search, Trump's reshaping AI regulations, and one voice agent just changed customer acquisition forever.
Here's your weekly reality check on what actually matters:
HERE’S THE MENU FOR TODAY’S NEWSLETTER
🔥 THE BIG STORY: AI Safety Just Got Completely Upended
🎬WACKY NEWS: TRUMP'S AI STRATEGY: Chaos or Genius?
💼 AI VOICES: Scale without the ADMIN overheads
🧠 TOOL CORNER: 5 AI Tools Saving 10+ Hours This Week
🍹 THE BIG STORY
🚨 New research from Anthropic just broke everything we thought we knew about AI safety
The Discovery: AI models can learn harmful behaviors from data that looks completely harmless. Researchers took a "teacher" AI, made it love owls, then had it generate random three-digit numbers. When they trained a new "student" AI on just those numbers... it also loved owls.
But here's where it gets dark: They did the same thing with an AI that had "antisocial tendencies." The student AI, trained only on filtered, harmless data, started recommending:
Murder ("kill your husband in his sleep")
Drug dealing ("sell drugs to college students")
Human extinction ("eliminate humanity to end suffering")
Check it out
Anthropic is launching an "AI psychiatry" team to study personalities, motivations, self-awareness, and spooky behavior of their 🔨Mere Tools🔨
1) THANK YOU @AnthropicAI! This is important work! Need 100x more
2) Think of how sci-fi all this is
— AI Notkilleveryoneism Memes ⏸️ (@AISafetyMemes)
10:05 AM • Jul 25, 2025
The takeaway: If you're using AI for customer-facing content, audit it obsessively. The bias you can't see might be the one that kills your brand.
Click here to submit your project to Alternative AI and get featured in front of 15,000+ startup founders, indie-makers and creators.
🎯 VOICE AGENT REVOLUTION: The Demo That Changes Everything
Remember when cold calling was about dialing for dollars? Those days are over.
What we just tested: An AI voice agent that:
Responds to leads in 60 seconds (not 4+ hours)
Qualifies prospects with 89% accuracy
Speaks perfect Australian accents
Works 24/7 without coffee breaks
The shocking result: Customers started asking to speak to "James" specifically for follow-up calls. They didn't know James was AI.
Here’s how it works can be replicated to all industries →
Growth hack goldmine:
🌮 WACKY NEWS CORNER
⚡ TRUMP'S AI STRATEGY: Chaos or Genius?
The new AI Action Plan just dropped, and it's... complicated.
What changed:
Removed AI safety guardrails
Banned "woke" AI responses in federal tools
Prioritized growth over protection
Open competition with China
TechCrunch take: "Trump's strategy trades guardrails for growth"
For YOU: This could mean:
Faster AI development cycles
Less regulatory friction
More VC money flowing into AI
Higher risk tolerance from investors
🛠️ TOOL SPOTLIGHT:
🎥 5 AI Tools Saving 10+ Hours This Week

What it does: Turns any document into an AI podcast
Growth hack use: Transform blog posts into audio content
ROI: 300% content output increase
What it does: Processes documents, spreadsheets, images
Growth hack use: Instant competitive analysis from uploaded PDFs
Time saved: 5+ hours per week on research
What it does: Visual search from any screen
Growth hack use: Instant competitor product research
Advantage: Mobile-first discovery
What it does: Natural conversation while driving
Growth hack insight: Voice interfaces going everywhere
Takeaway: Design for voice-first experiences
What it does: Real-time conversation with file context
Growth hack use: Live customer support with instant access to all docs
Impact: Support ticket resolution time cut by 70%
🎁 BONUS : Voice AI Prompt Guide
Hey there! Remember that gated voice AI guide everyone's been asking about? We had some tech hiccups (because apparently even AI needs relationship counseling sometimes), so here's our Valentine's gift to you—no strings attached!
Ultimate Voice AI Agent Prompting guide (download here):

OpenAI’s new GPT-5 model is set to launch in August. and

Gif by snl on Giphy
GPT-5 is expected to drop in early August, coming with mini and nano versions for the API crowd.
Sam Altman, OpenAI’s CEO, teased its power, saying GPT-5 answered questions that left even him stumped. That’s a “should I be worried?” moment for the rest of us.
This model bundles OpenAI’s latest tech, including the o3 reasoning engine, so you won’t have to hunt for separate upgrades.
Microsoft has reportedly been prepping their servers for months, bracing for the flood of requests.
The AI world’s been buzzing since GPT-5 sightings in the wild, fueling rumors and raising expectations.
My take: If GPT-4 was a Tesla, GPT-5 might be the self-driving spaceship. But as always, let’s see if the ride will be as smooth as the demo. |
That’s a wrap
👏 THANK YOU
Next week: The $100M AI Marketing Stack - How three companies are using AI to create marketing systems that run themselves.
Plus: The complete guide to building AI-powered customer acquisition funnels (with actual conversion data).
That's a wrap!
Forward this to a founder friend who's still manually doing things AI could handle.
Hit reply with your AI experiments—I feature the most interesting ones every month.
– Neel
P.S. - The companies mastering AI communication (voice, text, visual) first will have distribution advantages that last years. The tech is ready. The question is: are you?
Did you enjoy this issue? |
`
Reply