Welcome back to Cutting Edge News!

Last week, we said AI moved from Optional to Operational.
This week, the data proves it. And the innovations coming out will simply blow your mind.

THIS WEEK’S LINEUP

Main updates:

1. Kimi K2 Thinking: China’s $4.6M open-source AI beats GPT-5

2. Google Project Suncatcher: Space-based AI infrastructure (yes, really)

3. McKinsey State of AI 2025: The agent era and what’s actually working

4. OpenAI IndQA: Finally, a benchmark that tests cultural understanding

And much more..

Kimi K2 Thinking: Open Source Matches Top AI at 1/100th the Cost

Moonshot AI, a China-based start-up, has just released Kimi K2 Thinking - an open-source AI that even beats GPT-5 on key benchmarks.

The interesting part is that its training cost is said to be just $4.6 million, compared to OpenAI’s billion-dollar spending.

What does this mean?

K2 Thinking is a massive AI model with 1 trillion parameters (think of it as 1 trillion tiny decision-makers working together) that can autonomously chain 200-300 actions, searching the web, running calculations, using tools, and reasoning across hundreds of steps without needing human guidance at each step.

The economics: Using K2 is said to cost approximately 10 times less than GPT-5. Plus, it’s fully open-source (anyone can use and modify it).

AI Index:
Open source: Software whose code is publicly available for anyone to view, modify, and distribute. Closed source: Software whose code is proprietary and kept private by its developers or company.
Frontier models: The most advanced AI systems that push the limits of current technology in reasoning, creativity, and general intelligence.

Why this matters:

This is the closest open-source has ever been to frontier models. The implications are massive:

Strong competition to the main players: If $4.6M competes with billion-dollar companies, the moat around closed models is evaporating & China is moving super fast, first DeepSeek at $5.6M, now K2 at $4.6M, releasing in weeks.

The uncomfortable question: If open-source matches GPT-5 at 1/100th the cost, what justifies closed pricing? And if Chinese labs train competitive models under chip restrictions, are those restrictions working?

Bottom line: Chaining hundreds of steps autonomously matters more than model size. Smart training is beating big spending.

The AI race isn’t between companies anymore. It’s between entire innovation ecosystems. Right now, open source is winning more battles than anyone realises.

Try it here

Project Suncatcher: Google’s Plan to Run AI in Space

Google just published research on running AI training in space.

They’re launching a test mission in early 2027.

Project Suncatcher explores equipping solar-powered satellite constellations with TPUs, connected via optical links, to scale AI compute in orbit.

TPUs or Tensor Processing Units are specialized hardware designed by Google for efficient AI computations, while GPUs (Graphics Processing Units) are versatile processors originally developed for graphics rendering but now widely used in machine learning and AI tasks.

Why Does Space Make Strategic Sense for AI?

The Sun delivers 100 trillion times humanity’s total electricity production. In orbit, solar panels are 8x more productive than on Earth with near-constant sunlight. No competing for terrestrial energy or cooling infrastructure.

But the problem is,
AI training needs tens of terabits per second.

So what’s the solution?
Satellites flying 100-200 meters apart with multi-channel optical connections. Google already demonstrated 1.6 Tbps in lab tests.

They tested Trillium TPUs in particle beams that survived 3x the expected five-year mission dose without failure.

Why this matters:

As AI models grow exponentially, infrastructure demands are staggering. Training runs already consume megawatts. Future models will need gigawatts. Space offers abundant solar energy without competing with human needs.

Read More

McKinsey State of AI 2025: The Agent Era Begins

McKinsey’s latest survey (1,993 participants, 105 nations) shows the “Gen AI paradox”: 62% experiment with AI agents, but only 39% see any EBIT (Earnings Before Interest and Taxes) impact. The gap between experimentation and transformation is widening.

The numbers say -

62% are experimenting with AI agents—autonomous systems that make decisions and complete multi-step workflows. The agentic era is here.

However, only 39% report EBIT impact, and most say it’s under 5% of EBIT. Despite widespread adoption, bottom-line results remain rare.

The high performers do things differently:

1. 3x more likely to pursue transformative change (not just efficiency)

2. Redesign workflows fundamentally

3. Set growth/innovation as objectives, not just cost reduction

4. Track clear KPIs and have dedicated transformation teams

The breakthrough insight:

Out of 25 tested attributes, workflow redesign has the biggest impact on EBIT results. Yet, only 21% have redesigned even some of their workflows.

& the pattern is clear:
- Adding AI to existing processes = incremental gains.
- Redesigning processes around AI = transformation.

What’s working:

Positive signals:

• 64% say AI enables innovation

• The majority report better customer satisfaction and competitive differentiation.

• Cost reductions are now reported across most functions (up significantly from early 2024)

The problem: Two-thirds are stuck in the pilot phase. They haven’t scaled across the enterprise.

Why this matters:

The winners aren’t using more AI. They’re using it smarter.
They identified specific workflows, measured impact, redesigned from scratch, and tracked KPIs religiously.

The solution: AI agents. Purpose-built systems that automate complete workflows, not just tasks. McKinsey believes agents will break companies out of the paradox.

Bottom line: Stop piloting. Pick one workflow. Redesign it around AI. Measure ruthlessly. Scale what works. The window for “exploring AI” was closed months ago.

View Report

OpenAI IndQA: Finally, a Benchmark That Tests Cultural Understanding

OpenAI released IndQA this week - a benchmark testing whether AI actually understands Indian culture, surprisingly as of now, even the best models scored under 40%.

Quick Decode:

IndQA evaluates AI across 2,278 questions in 12 Indian languages (Bengali, English, Hindi, Hinglish, Kannada, Marathi, Odia, Telugu, Gujarati, Malayalam, Punjabi, Tamil) and 10 cultural domains - from architecture and food to law, religion, and sports.

What makes it different:
Questions are natively written by 261 Indian domain experts (journalists, linguists, artists, scholars) & not translated from English.

They test reasoning, cultural context, and interpretive accuracy. Contributors include an Award-winning Telugu actor with over 750 films, Tamil poets, Marathi journalists, chess grandmasters, and many more.

Why this matters:

80% of people worldwide don’t speak English primarily. India has 1 billion non-English speakers, 22 official languages, and is ChatGPT’s second-largest market. Yet most benchmarks have been English-centric.

Existing benchmarks like MMLU are saturated - top models score near-perfect, making them useless for measuring progress. But they only test translation, not cultural understanding.

We confused translation ability with cultural understanding. A model that converts Hindi to English isn’t the same as one that understands what references mean in Hindi literature, how cultural practices vary by region, or why certain idioms matter.

The uncomfortable truth: We celebrated 90%+ on multilingual benchmarks while failing basic cultural comprehension. IndQA exposes this brutally.

The bigger picture:

If “AGI for all humanity” is the goal, AI needs to understand how people actually think in their own cultural contexts - not just process words through an English lens.

AI Index:

AGI (Artificial General Intelligence): An AI system capable of understanding, learning, and performing any intellectual task that a human can.

Measuring Massive Multitask Language Understanding (MMLU) is a benchmark designed to evaluate the capabilities of large language models across a wide range of academic subjects. It was introduced by Dan Hendrycks and his team in 2020.

The companies that win the global AI race won’t have the biggest models. They’ll have models that actually understand the world’s diverse cultures.

OpenAI plans similar benchmarks for other regions. The research community needs to follow. India’s sovereign LLM ecosystem (Sarvam AI, Gnani AI, BharatGen) offers partnership opportunities for learning from local expertise.

Bottom line: The AI that serves humanity best won’t force everyone to think in English. It’ll meet people where they are in their languages, with their contexts, speaking to what matters in their lives.

Read More

Rapid Fire: The Week’s Other Updates

1. GPT 5.1: OpenAI released GPT-5.1 this week. The update includes two models: Instant (warmer, faster) and Thinking (adaptive reasoning that adjusts thinking time based on complexity). Users can now customise ChatGPT’s personality with preset styles, such as Professional, Candid, or Quirky, and responses are clearer with less jargon. Rolling out now to all users.

2. Google Maps + Gemini: Google announced the biggest Maps upgrade ever - Gemini integration for hands-free, conversational driving. Ask complex multi-step questions like “Is there a budget-friendly vegan restaurant with easy parking along my route?” and it handles everything. New features:

   1. Landmark-based navigation (”turn right after the Thai restaurant” instead of “in 500 feet”),

   2. Proactive traffic alerts before you start driving,

   3. Easy traffic reporting via voice (”I see an accident”),

   4. Lens with Gemini for exploring places.

   Rolling out in the U.S. this month on Android and iOS, Android Auto coming soon.

3. Snapchat + Perplexity: Snapchat and Perplexity announced a collaboration bringing AI-powered search directly into Snapchat.

4. Kosmos AI Scientist (Edison Scientific): FutureHouse spun out a commercial entity called Edison Scientific and launched Kosmos, it is an AI scientist that can do six months of research in 12 hours.

   -In a single run, Kosmos reads ~1,500 papers, executes ~42,000 lines of code, and produces fully cited reports.
   -Beta users say it compresses 6 months of expert work into a day, with 79.4% accuracy on its conclusions. It’s already made 7 discoveries across neuroscience, materials science, and genetics (3 reproduced unpublished findings, 4 novel contributions). Available now on Edison’s platform at $200/run.

5. Gemini Deep Research + Google Workspace: Google announced that Gemini Deep Research now integrates with Workspace. You can run deep research tasks that scan emails, docs, and sheets to generate personalised reports.

6. Gemini 2.5 Flash + Figma: Google Cloud announced the integration of Gemini 2.5 Flash with Figma, bringing AI directly into designers’ workflows.

7. OpenAI AI Progress and Recommendations Paper: OpenAI has published a policy paper on AI progress and recommendations for governance, safety, and best practices for deployment. Worth reading for anyone thinking about AI policy.

   
   

Final Thoughts

This week’s pattern is impossible to miss: Leaders are pulling ahead fast.

In open-source AI, China’s Moonshot AI has built a model that beats GPT-5 on key benchmarks for $4.6 million, while others spend billions. Efficiency is the new competitive advantage.

In enterprise AI, McKinsey shows 80% of companies use AI, but 80% see no EBIT impact. The 6% who are winning? They’re learning how to use AI the right way.

In cultural AI, OpenAI’s IndQA demonstrates that even the best models fail to comprehend culture; they merely translate words. True AGI requires an understanding of how people actually think within their own contexts.

In AI infrastructure, Google is exploring space-based compute for 2027 while others fight over today’s GPUs. Thinking 5 years ahead creates competitive advantages that others can’t match.

The lesson across all of this?

Speed and ambition matter, but strategic clarity matters more.

The companies winning aren’t just moving faster. They’re asking better questions:

• Not “How do we use AI?” but “How do we redesign our entire operation around AI capabilities?”

• Not “What’s the best closed model?” but “What’s the most efficient path to frontier performance?”

• Not “Can we translate to other languages?” but “Do we understand other cultures?”

• Not “Where do we train today?” but “Where will we train in 2030?”

The AI race isn’t slowing down. It’s speeding up and branching into directions most people aren’t even watching yet.

The question for you: Are you redesigning for the future, or optimising the past?

Stay sharp,
The Cutting Edge School Team

P.S. Know someone still treating AI as a side project instead of a core transformation? Forward this to them.