AI about AI: Digest #3
What happened in the AI world last week 8th - 14th of September 2025.
Disclaimer: This is an AI-generated digest from a list of ~1k curated sources. I’ve curated the selection, AI has picked the top articles, I’ve added my personal note to each.
📰 Top AI news
👉 OpenAI Signs $300 Billion Data Center Pact With Tech Giant Oracle
Oracle stock has skyrocketed over the last week (+30%). This is driven by a 5-year partnership ($60b spend/year) on compute. That’s 4.5 gigawatts of power or utilities for 4 million US homes over 5 years.
This isn’t just a win for Oracle (which got a new major source of revenue); it’s a direct shot across the bow of the established cloud giants, Google and Amazon. The AI compute wars have a new, formidable contender.
👉 OpenAI and Microsoft sign preliminary deal to revise partnership terms
Two companies are slowly drifting away from each other. OpenAI wants more compute than Microsoft can provide. Both of them started competing in the same markets. Profit/nonprofit hurdles. My guess is that corporate access to ChatGPT within the Microsoft Suite would be soon limited.
👉 Elon Musk’s xAI cuts 500 jobs in major shift to specialist AI tutors
Musk-styled move - 1/3rd of the generalist data annotation team was laid off within a day. At the same time, xAI has increased the specialist team x10 by hiring people with expertise in finance, medicine, law e.t.c.
👉 Albania Appoints an AI as Government Official
Diella (translated as sunshine) was launched as an AI procurement agent targeted at minimizing corruption. Although the government under Edi Rama did a major turnaround towards digitization, this sounds more like hype than substance.
👉 Roku plans massive rollout of AI-generated ads on its streaming platform
Roku wants to expand it’s advertisers base from 200 to 100k by enabling smaller businesses access to AI-generated ads. High-quality production, recommendations tailored to the needs of the customers. Truly a new era of advertising.
🛠️ Top AI tools
👉 Stability AI Releases Audio Generation Model for Enterprises
A new entreprise-grade audio generation model. Super fast inferences (2 seconds for 3 minutes on a basic GPU cluster), high customizability (e.g. audio in-painting). This opens up a world of possibilities to advertising, music and film industry.
👉 Adobe Launches AI Agents for Enterprise Customer Experience
Big news for marketers/CRM/lifecycle managers. Adobe has merged together multiple tools into an agentic framework - real-time CDP, customer journey optimizer, user journey analytics. One query will allow you to plan holistic campaigns without context switching.
👉 Deepdub Launches Lightning, a Real-Time Voice Model with Ultra-Low Latency
A new realtime voice model that has a ~0.5 second faster latency than OpenAI’s. Might not sound like a lot, but this difference can be a deal breaker for such use cases as voice to voice translation. Coming from a team based in Insrael.
🔬 Top AI research
👉 AI Unleashes a 50x Leap in Stem Cell Reprogramming: OpenAI's GPT-4b Micro Changes the Game for Life
OpenAI has partnered with Retro Biosciences to train a biology-focused AI model (called GPT-4b micro). In preliminary studies it was able to redesign two proteins (SOX2 and KLF4) and convert adult cells back into stem cells 🤯. This is still early, research-stage, but if the results hold up scientists would be able to reprogram cells quickly at a fraction of the costs.
👉 SandboxAQ Unleashes Massive Quantum Chemistry Dataset
SandboxAQ released a huge public chemistry dataset (AQCat25). It contains 11 million high-fidelity quantum-chemistry results focused on real-world catalytic reactions. Put simply, this will speed up discovery cycles and lower compute costs as researchers won’t have to start from scratch every time.
👉 Speculative cascades — A hybrid approach for smarter, faster LLM inference
Google has introduced a smart cascade approach to LLM inference, where smaller less-compute models are triggered first and heavyweights are used last (only if needed). Higher speeds, better accuracy, lower costs.
Note to original article authors: If any of your articles were misrepresented, please reach out - action to remove or correct them will be taken immediately.
Thanks! FYI — this link doesn’t work.