Scale And Strategy
together with
This is Scale And Strategy, the daily newsletter that’s like a BizOps piñata - whack it open and savor the sweet, sweet pieces of knowledge.
Here’s what we got for you today:
- OpenAI’s GPT-5.2 is very good at spreadsheets
- Robots have a data problem
OpenAI’s GPT-5.2 is very good at spreadsheets
OpenAI rolled out an upgraded version of ChatGPT after a messy week that saw the company declare a “code red” in response to recent gains by Google’s Gemini models.
Despite internal debate about delaying the launch, GPT-5.2 shipped and quickly showed gains over GPT-5.1 across work, coding, and general benchmarks. The model debuted at No. 2 on LMArena’s web development leaderboard, trailing only Claude Opus 4.5. According to the Wall Street Journal, OpenAI’s “code red” posture will stay in place until a broader release in January. For now, the garlic can cool slightly.
The timing matters. OpenAI has taken heat for massive compute deals and for CEO Sam Altman’s perceived lack of focus on ChatGPT itself. Unsurprisingly, the messaging around GPT-5.2 leans hard into economic value.
In its release, OpenAI calls GPT-5.2 “the most capable model series yet for professional knowledge work,” spotlighting improvements in white-collar use cases. On X, OpenAI VP Nick Turley and president Greg Brockman both highlighted the model’s surprisingly strong spreadsheet performance, an unglamorous but lucrative competency.
GPT-5.2, combined with OpenAI’s blockbuster Disney deal, is a net positive for the company. But it’s not a victory lap.
The family of an 83-year-old woman killed by her son after he reportedly experienced ChatGPT-related hallucinations has filed a wrongful-death lawsuit against OpenAI. The case follows seven lawsuits filed in November alleging wrongful death, assisted suicide, and involuntary manslaughter tied to ChatGPT.
OpenAI says GPT-5.2 performs better in mental-health contexts than prior versions and plans to deploy an age-prediction model to reduce exposure to sensitive content for users under 18.
After Anthropic CEO Dario Amodei publicly criticized Sam Altman for a tendency to “YOLO things,” and reports surfaced that Altman explored entering the space industry last year, OpenAI needed to signal discipline and focus. GPT-5.2 does that by emphasizing real, billable work. Still, the growing wave of litigation suggests the company’s toughest challenges may lie ahead, not behind.
Ads eating into your store margins? Only pay for conversions with this affiliate network, and scale as much as you want
While you’ve been burning cash on ads that may never convert, smart brands are turning affiliates into their most efficient channel.
Awin’s affiliate platform helps brands of all sizes connect with their database of over one million trusted partners to drive measurable e-commerce results and solve challenges across acquisition, conversion, and retention.
They’re reaching new audiences and keeping costs low with an all-in-one platform that makes managing partners, reporting, and payments refreshingly simple.
And because everyone’s talking about influencers—rightfully so—Awin decided to crack the code on making creators actually drive ROI instead of empty engagement metrics.
Their whitepaper shows you how to find, recruit, brief, and analyze creators so you get the most out of your campaigns.
Access the Influencer Marketing for ROI whitepaper for free here.
Interested in leveraging affiliate marketing for your own brand? This is a good place to start.
Discover Awin’s ROI-generating platform here.
Robots have a data problem
Tech companies love to talk about a future where general-purpose humanoid robots roam the world, quietly doing useful things. The reality is far messier. Building physical AI that actually works is orders of magnitude harder than training a chatbot.
From foundational hardware constraints to the fine motor control of robotic hands, physical intelligence introduces challenges that large language models never face. A robot doesn’t just need to generate plausible text. It has to perceive the real world, understand it, and act on it in real time. That’s a very different problem.
“Building a robot is not easy. I often say it’s like raising a kid. You need a village to do that,” Nvidia’s Rev Lebaredian Goel said during a keynote on Tuesday.
At the core of the problem is data. AI models are only as good as the data they’re trained on, and for physical systems, scraping the internet doesn’t cut it. If a model is meant to operate in the physical world, its training data has to reflect the physical world too.
Ken Goldberg, a roboticist at UC Berkeley, calls this the “100,000-year data gap.” The amount of training data required for general-purpose humanoid robots dwarfs what today’s LLMs need. “There’s not enough real-world data out there,” Nebius’ Helga told The Deep View.
Humans can collect real-world data, but that process doesn’t scale without humans staying in the loop. And while deploying robots can generate useful data, that only works once robots already exist in meaningful numbers. As TJ Galda, senior director of product management for Nvidia’s COSMOS world models, put it: “If you build a brand-new robot and you’ve never deployed it, you’ve got zero video data, zero LiDAR.”
Even narrow, task-specific robots don’t escape the problem. According to Dyna Robotics CEO Lindon Gao, physical AI systems still require massive pretraining datasets to generalize across environments and edge cases. “We need to scale pretraining datasets much more drastically so we can cover a wider distribution of data across different types of tasks,” Gao said at a Nebius panel.
That’s why synthetic data is increasingly seen as the escape hatch. World models that simulate real environments can generate enormous datasets at scale and expose systems to rare scenarios that real-world data would capture only once, if ever.
A dashcam might record a bear running across a road a single time. A world model can replay that scenario thousands of times under slightly different conditions. The question, Galda said, is whether synthetic data can be good enough to substitute for reality.
“If we can build synthetic data that’s just as good as hitting record on a camera,” he said, “then we can scale this problem very quickly.”
Until then, humanoid robots aren’t waiting on better hardware or smarter algorithms. They’re waiting on data that barely exists.
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!
