- AI Fire
- Posts
- 🎭 73% Fooled by GPT-4.5 Human Mask
Plus: GPT-4 Will Die Soon?
Read time: 5 minutes
Something strange just happened in a lab. A group of people sat down, 73% thinking they were chatting with another human. They were wrong. Most of them didn’t even realize “that human” is GPT-4.5? And it wasn’t even trying that hard to fool them!
What are on FIRE 🔥
IN PARTNERSHIP WITH NOTION
Get Over $6K of Notion Free with Unlimited AI
Running a startup is complex. That's why thousands of startups trust Notion as their connected workspace for managing projects, tracking fundraising, and team collaboration
Apply now to get up to 6 months of Notion with unlimited AI free ($6,000+ value) to build and scale your company with one tool.
AI INSIGHTS
Is it possible that an AI now feels more human than actual humans? OpenAI’s model has become the first large language model (LLM) to do that. But it also raises ethical concerns, as it didn’t just “pass” in many cases, it seemed more human than actual people.
How the Three-Party Turing Test Works:
One human interrogator interacts with two unseen entities: one human and one AI. They evaluated 4 different types of AI system in the witness role: GPT-4.5, LLaMa-3.1-405B, GPT-4o, and ELIZA.
People judged based on “vibes”.
The interrogator asks both questions and must guess which one is the human.
GPT-4.5 was judged to be human more often than the actual humans it competed against.
GPT-4.5 Passed the Original, Hardest Version of the Turing Test
Meta’s LLaMa-3.1 fooled participants 56% of the time, still beating the original Turing prediction (30% error rate or better).
GPT-4o performed below chance (21%), grouped closer to ELIZA (23%) than its GPT predecessor.
GPT-4.5, with its 73%, was clearly the best-performing model.
=> AI is now mimicking human behavior more convincingly than ever. Is it good for more human-like AI agents for conversations and support?
Why it matters: The scary part is that people connected with AI emotionally, often more than with actual people. Scammers will love this. Brands will weaponize it. And most users won’t even notice. It’ll replace the parts of being human we’re too lazy to protect.
TODAY IN AI
AI HIGHLIGHTS
🛠️ OpenAI’s next “agentic” AI agent called "A-SWE" is a self-testing software engineer that does all the things that coders hate to do. OpenAI’s effort to stay ahead of xAI and Perplexity.
📦 OpenAI is closing the GPT-4 chapter in 2 weeks. It will be fully replaced but still available via API. This signals OpenAI’s push toward new model families like GPT-4.1, o3, and o4-mini.
🚀 DeepSeek signals next-gen R2 model, unveils novel approach to scaling inference with SPCT. Shift in AI paradigm: From pre-training to post-training, following the lead of OpenAI’s o1 model.
🔍 OpenAI, Anthropic, and Elon Musk claim AGI is just 2–3 years away. But AGI is being hyped, vague, misunderstood, and often misused. So, are any of these “really” AGI?
📢 Meta blames the internet for its AI’s left-leaning bias even though Facebook is the internet for many people. Llama 4 is trying to reposition itself as a "politically neutral" & please the new MAGA crowd.
🧠 We are the real AI. AI is us. AI isn’t the problem, we are. AI copies, reflects, and learns from everything we do. “AI Mirror” shows our true inner code.
💰 AI Daily Fundraising: Autonomous vehicle startup Nuro secured $106 million in funding, bringing its valuation to $6 billion. The company plans to use the funds to scale its AI capabilities and advance commercial partnerships.
AI SOURCES FROM AI FIRE
PRESENTED BY WRITER
You’ve heard the hype. It’s time for results.
For all the buzz around agentic AI, most companies still aren't seeing results. But that's about to change. See real agentic workflows in action, hear success stories from our beta testers, and learn how to align your IT and business teams.
NEW EMPOWERED AI TOOLS
🧠 MindPal turns your expertise into sharable AI agents that work 24/7.
🗂️ Airtable AI Assistant build apps through conversation, not clicks.
🧍♂️ OmniHuman by ByteDance creates lifelike human videos from a single image and motion signals.
⚡️ Crono automates multichannel sequence and closes more deals.
🎨 NUMI connects directly with AI-enabled designers. Handpick your expert.
AI QUICK HITS
🚀 Access to future AI models in OpenAI’s API may require a verified ID.
🧬 MIT built a faster way to protect sensitive AI training data without hurting accuracy.
🗣 A new push to cut AI costs by cloud giants like Amazon and Google are rising dangers for Nvidia.
🎥 Netflix uses OpenAI-powered search engine based on users’ current intent.
☢️ For the first time, AI is being used at a nuclear power plant.
AI CHART
We’ve known for centuries that cells are complex. We even have names for thousands of proteins inside them. But until now, no one knew how they all fit together, so they used: 20,000+ images of glowing proteins inside cells and specifically GPT-4 to analyze, name, and describe these protein clusters.
And the result was a map of over 5,100 proteins, organized into 275 distinct “protein machines” inside the cell. And most of this has never been seen before.
For the first time, GPT-4 has created the most detailed and interactive map of a human cell. This cellular atlas reveals over 975 previously unknown protein functions and turns 400 years of cell research into a zoomable, Google Maps-style blueprint.
The project was led by UC San Diego, with contributions from Stanford, Harvard Medical School, UCSF, and University of British Columbia. Combines AI (GPT-4) + microscopy + protein interaction data for high accuracy.
Huge Scope and Protein Coverage:
The study uncovered 975 new protein functions using GPT-4.
GPT-4 was used to:
Interpret protein function from scientific literature.
Summarize how proteins work together in assemblies.
Propose names and functions for protein clusters in the map.
=> This dramatically sped up work that would take researchers years manually. Despite knowing all the proteins, scientists never had a complete parts list + assembly guide for any human cell type until now.
The online U2OS cell map is zoomable and interactive. Users can explore protein assemblies, locations and functional communities. The team plans to make resolution even higher in future versions => It’s like Google Maps for cell biology.
It’s also proof that AI + science = more than automation, it’s collaboration for discovery. Instead of focusing on rare mutations, we can now see which “cellular machines” are most likely being hijacked or broken.
AI JOBS
Gartner: AI & GenAI Expert Advising Technical Professionals (Senior Director/Analyst, Fully Remote US)
TikTok: Algorithm Engineer, TikTok E-Commerce (Conversational AI)
Anthropic: User Experience Researcher in San Francisco
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