As @satyanadella says, we're making great progress on shipping MAI models that are higher quality, faster and cheaper across MSFT. In PowerPoint, our image model cut costs 84% compared with GPT-Image-2. In OneDrive, it lifted save rates 26% and cut latency by ~25%. It's also
ECONOMY
-
Merge Packages Multi-Model Answers and a Judge into One API Call
By
–
Yeah, it’s pretty close to the MoA approach Nous Research already shipped in Hermes Agent a few weeks ago. Merge is just packaging that same multi-model-plus-judge idea into a simple one-call API that claims frontier results for about a quarter of the price. Ensemble methods
-
Post Compares AI-Driven Tech Stock Speculation to Crypto
By
–
tech stocks looking like crypto: what you get combine meme stock vibes with a bit of not-proven-to-be-profitable AI thrown in.
-
Authors Share Costs Behind Building LLMs for Production Book
By
–
We wrote our first book, Building LLMs for Production. It cost us more than $50,000. And money was the smallest part of that cost. Here's the full breakdown of what it took: energy, team, time, and money. There's a little secret on the last slide.
-

Post Questions Big Tech’s Cash Flow Spending on AI
By
–

Google’s Pichai and Ashkenazi provided a great example of how there is no clear articulation by big tech of the whys and wherefores of spending all their cash flow on AI. I think the average investor suspects that it’s an arms race that won’t necessarily end well. But the
-
Gary Marcus Questions Whether LLM Developers Can Keep Enduring Advantages
By
–
Warned you all that there was no such thing as an “enduring advantage” in LLMs. Would the world be different if people had listened to the below?
-
Gemini Flash Models Trade Off Token Efficiency, Coding Accuracy, and Speed
By
–
Key highlights: Gemini 3.6 Flash: The general workhorse — 17% fewer output tokens, higher coding accuracy, and lower cost per task ($1.50/1M input, $7.50/1M output). Gemini 3.5 Flash-Lite: Blazing fast execution at 350 output tokens/sec ($0.30/1M input, $2.50/1M output).
-
Agent-to-LLM Traffic Visibility Helps Control Tokens and Execution Costs
By
–
Spot on, Adrian. Without unified visibility into Agent-to-LLM traffic and loop conditions, token burn and unmonitored execution can spike enterprise costs very quickly.
-
Post Links Google’s Negative Free Cash Flow to AI Spending
By
–
Wow Google FCF negative due to AI expenditure
-
Post Proposes Dedicated AI Inference Budgets for Open Math Problems
By
–
The fact that random people with ChatGPT Pro can solve open maths problems should be embarrassing for Universities. There is a big gap in inference compute that is being spent on solving open problems. Why isn't there a unit that has maybe $10m inference budget and continuously