- AI Fire
- Posts
- 💻 Human? Almost.
In partnership with
A new AI agent just hit human-level performance on desktop tasks - quietly. Meanwhile, OpenAI’s billion-dollar tour hints at a bigger war - one fought over chips, not models…
What are on FIRE 🔥
PRESENTED BY LIBERTY UNIVERSITY
Liberty University ensures you have options for earning your degree. With over 600 online programs, Liberty offers you the tools you need to advance your education.
Enrollment is simple! Liberty offers assistance through every step of the admission, enrollment, and financial aid processes. And with no standardized tests or entrance exams for most programs, becoming a Liberty student is more attainable than ever! Transferring in? Students can transfer credits for up to 75% of an undergraduate degree or up to 50% of a graduate or doctoral degree. Through the Experience Plus program, you can also earn credit for qualified life experience, job experience, and military training.
Find your calling at Liberty University.
AI INSIGHTS
Simular Research just dropped a big one: computer-use agents - the kind that click, type, and scroll like you - are now almost as good as humans (72%) at complex desktop tasks. Their new method, Behavior Best-of-N (bBoN), hit 69.9% success on OSWorld, beating the old record by 10 points.
What’s new:
Don’t trust one run. Run many agents, pick the best.
Each run writes a “behavior narrative” of its actions.
A judge (GPT-5) compares and picks the smartest path.
Works across Windows, Android, and OSWorld benchmarks.
Numbers that matter:
OSWorld: 62.6% → 69.9% (Human 72%)
WindowsAgentArena: 50.2% → 56.6%
AndroidWorld: 68.1% → 71.6%
Efficiency: Cuts LLM calls by 52%, task time by 62%.
Best mix: GPT-5 + Gemini 2.5 Pro → 66.7% success, 78% Pass@N.
Why it matters:
This is the closest agents have come to real computer work. Instead of scaling model size, Simular shows that scaling behavior and judgment works better. It’s a quiet milestone - one that aligns with how OpenAI o4-mini, Claude 4, and Gemini 2.5 Pro are moving toward true autonomous AI workers.
PRESENTED BY WING
Stop being the bottleneck
Every leader hits the same wall: too many priorities, not enough bandwidth. Wing clears that wall with a full-time virtual assistant who runs the drag layer so you lead, not chase.
Offload scheduling, inbox, follow-ups, vendor wrangles
Keep your stack, your process, your control
Scale scope as you scale revenue
This isn’t gig work. It’s dedicated support that shows up every day and allows founders to delegate without losing control.
TODAY IN AI
AI HIGHLIGHTS
🔥 Meta turned AI adoption into an internal game. Employees now compete for badges in “Level Up,” tracking who uses AI the most. Reality Labs hit 70%+ AI usage, aiming for 75% by year-end. Even Google and Microsoft are tying AI use to performance.
🧪 DeepMind says its AI found 2.2M new materials, but scientists call many of them unrealistic or impossible to make. Still, AI labs at Google, Microsoft, and Meta are quietly refining models that predict and even synthesize new compounds using robots.
💼 Sam Altman is touring Asia and the Middle East to raise billions for OpenAI’s next infrastructure push. He’s meeting TSMC, Samsung, Foxconn, and SK Hynix to secure chip supply - with OpenAI expected to spend $16B this year and up to $400B by 2029.
🎬 Sora is getting an upgrade - Altman says OpenAI will add opt-in copyright controls for studios and creators, plus a monetization system that could share revenue with rights holders. Even Hollywood seems curious about “interactive fan fiction” powered by Sora.
💰 OpenAI just acquired Roi, a personal finance AI app backed by Balaji and Spark Capital. Only its CEO joins OpenAI, signaling a bigger push into personalized consumer AI like Pulse and Sora - where your AI learns, adapts, and even jokes like you.
💰 AI Daily Fundraising: Unconventional, Inc., founded by Naveen Rao (ex-Databricks, MosaicML), is raising $1B at a $5B valuation led by a16z, with Lightspeed, Lux Capital, and Databricks joining. The startup aims to “rethink computing” for Brain-Scale Efficiency, positioning itself to compete with Nvidia.
AI SOURCES FROM AI FIRE
NEW EMPOWERED AI TOOLS
💬 Super Intern is a customizable AI teammate inside group chats that reminds, answers & takes action
🗒️ Krisp AI Note Taker records, transcribes & summarizes meetings in 16+ languages across devices
🧠 Strix is an open-source AI hacking agent that finds and validates real security vulnerabilities fast
⚙️ RightNow AI Code Editor is the first GPU-native code editor with real-time profiling & optimization
AI QUICK HITS
🎬 Sora just hit No.1 on Apple’s U.S. App Store - beating Gemini and ChatGPT with 164K installs in 2 days, even while invite-only.
🌾 Instacrops is helping 260 farms cut water use 30% and boost yields 20% using AI-powered irrigation across Latin America.
🧠 Brazilian scammers made millions using Gisele Bündchen deepfakes on Instagram - forcing Meta to tighten celeb-bait ad detection.
🚗 Von der Leyen urged a Europe-wide AI-car push, calling for “AI-first, safety-first” to revive the region’s auto industry.
🚀 Jeff Bezos says data centres in space will beat Earth-based ones within 20 years - powered by 24/7 solar energy.
AI CHART
A new Microsoft + Science study shows AI tools like ProteinMPNN, EvoDiff, and OpenFold can recreate toxin-like proteins that escape today’s safety checks.
Researchers rebuilt 72 dangerous proteins - from Ricin to Ebola VP24 - generating 17,000+ variants with only 40–60% sequence match, but nearly identical structures.
Even top systems - Aclid, FAST-NA, IBBIS, Battelle UltraSEQ - missed up to 90% when DNA was scrambled. Only HMM-based tools caught a few more.
All data is now locked under IBBIS, backed by White House AI Safety and NIST.
Why it matters: AI models can now design biohazardous proteins that bypass current screening - pushing for AI-resilient biosecurity before open-source tools go too far.
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