Scale & Strategy
together with
This is Scale & Strategy, the newsletter that catches you up like a blue shell in Mario Cart.
Here’re the quick highlights from the week:
- LLMs won’t get us to AGI, says Sutskever
- DeepSeek’s new models are going straight at GPT-5 and Gemini 3 Pro
LLMs won’t get us to AGI, says Sutskever
The generative AI boom has been built on one mantra: scale it and they will come. But a growing chorus of top researchers now thinks the industry may be climbing the wrong mountain.
On a recent podcast with Dwarkesh Patel, Ilya Sutskever, OpenAI co-founder and now head of Safe Superintelligence, argued that the field’s obsession with scaling laws has gone too far. He said the industry shifted in 2020 from an “age of research” to an “age of scaling,” where the priority became bigger models, bigger compute budgets, and bigger claims.
That mindset, he said, is safe and predictable but also limited. “It’s back to the age of research again, just with big computers,” Sutskever noted, suggesting that the next leap forward won’t come from simply inflating model size.
He’s far from alone.
Yann LeCun, one of the original giants of deep learning, said earlier this year that LLMs won’t get us anywhere near “human-level AI.” And Benjamin Riley wrote in The Verge that language and thought are not interchangeable: humans use language to think, but language itself is not the thinking.
As doubt grows around the scaling playbook, interest is shifting to AI systems that actually understand the physical world. Dr. Fei-Fei Li’s World Labs launched Marble in November, a commercial “spatial intelligence” model. Researchers at the Mohamed bin Zayed University of Artificial Intelligence released PAN, their next-generation world model. Robotics firm Physical Intelligence just raised $600 million at a $5.6 billion valuation. The signal is clear: the next frontier may be perception, embodiment, and real-world reasoning, not larger text models.
All of this raises a bigger, uncomfortable question: even if AGI were achievable through pure scale, what would enterprises, governments, or end users actually do with a model that can do everything? Many companies are finding that small, targeted models already outperform giant ones on specialized tasks while being cheaper, faster, and far less resource-hungry.
That matters, because the global data-center expansion and trillion-dollar spending spree assume scaling laws will keep paying off indefinitely. If that assumption is cracking, the industry faces a fundamental reckoning: whether the future belongs to endlessly bigger LLMs, or to a new wave of models that think about the world in entirely different ways.
The perfect email is useless in a spam folder. This suite guarantees maximum deliverability
You’ve built the list and crafted the copy.
You’ve designed a stunning email in your ESP.
Then you hit send, and… crickets. Opens are low, clicks are embarrassing, and the only replies you get are bounce notifications.
Sending to a bad list is more than wasted effort. It’s a direct threat to your sender reputation that risks drowning future campaigns in spam and promotion folders.
And no readers = no ROI.
Stop gambling with your deliverability.
ZeroBounce is the all-in-one platform trusted by marketers at Amazon, HubSpot and Netflix for landing their messages in primary inboxes.
They provide 99.6% accurate email validation, inbox placement testing, blacklist monitoring, and more, so you can send with absolute certainty and get a bigger ROI from your emails.
Discover why over 500,000 businesses rely on ZeroBounce to ensure their message is heard.
DeepSeek’s new models are going straight at GPT-5 and Gemini 3 Pro
The gist: DeepSeek just dropped V3.2 and V3.2-Speciale, two reasoning models that basically walk up to the frontier crowd and say “move.” They hit GPT-5/Gemini-3-Pro-level performance, cost a fraction of the price, and come fully open-source under MIT.
What’s inside:
- Frontier parity: V3.2 lands neck-and-neck with GPT-5, 4.5 Sonnet, and Gemini 3 Pro across math, coding, and tool-use benchmarks. The heavier Speciale variant actually leaps ahead in multiple tests.
- Competition wins: Speciale pulled gold-level scores at both the 2025 International Math Olympiad and the Informatics Olympiad, and still placed top-10 overall at IOI.
- Aggressive pricing: V3.2 runs at $0.28 input / $0.42 output per 1M tokens. Compare that to Gemini 3 Pro at $2 / $12, GPT-5.1 at $1.25 / $10, or Sonnet 4.5 at $3 / $15. It’s not subtle.
- Fully open: Both are massive 685B-parameter models, MIT-licensed, and the weights are sitting on Hugging Face waiting for anyone bold enough to download them.
Why it matters: R1 already shook the market and triggered the usual export-control anxiety. V3.2 doubles down and proves DeepSeek isn’t a fluke. They’re releasing frontier-grade capability at clearance-rack pricing. U.S. labs now have to explain why their premium APIs deserve the markup, because the delta just shrank to almost nothing.
Was this email forwarded to you?
That’s it for today and as always It would mean the world to us if you help us grow and share this newsletter with other operators.
Our mission is to help as many business operators as possible, and we would love for you to help us with that mission!
