Hey folks,
Week one of 2026 told us everything we need to know about this year.
Here's what happened.
First, some context:
2025 was the year of breakthroughs. Claude Sonnet 3.5 in June. OpenAI's o1 reasoning models. Context windows hit 200K tokens. Coding agents went from demos to daily tools. The industry was riding high—every quarter brought capabilities that seemed impossible the quarter before.
But the real shift wasn't in the labs. It was at your desk.
By the end of 2025, AI stopped being optional. GitHub Copilot went from "that thing some people use" to "wait, you're NOT using it?" Cursor became the default editor for entire teams. ChatGPT tabs sat pinned next to Slack and VS Code. You stopped Googling error messages and started asking Claude to debug them.
For developers, AI went from experimental to essential in twelve months. That's the context everyone forgets when they talk about "AI hype." This wasn't hype. This was real tools solving real problems, every single day.
Then December happened. The releases slowed. The benchmarks plateaued. And quietly, the conversation shifted from "look what AI can do" to "okay, but what's the ROI?"
Here's what I'm seeing as we enter 2026:
Week one brought a pattern. IBM claims quantum advantage. CES goes all-in on physical AI. Boards demand ROI. And Ilya Sutskever—the guy who built transformers—admits they've plateaued.
The "just scale it" era is over. For a decade, every AI problem had the same solution: more data, more compute, more parameters. That worked until it didn't. Now we're watching the industry scramble. Some bet on quantum. Some on world models. Some on smaller, specialized systems.
2026 isn't about who builds the biggest model. It's about who figures out what actually works when scaling stops.
Let me walk you through what happened.
2026: AI's "Show Me the Money" Year
The party's over.
I was reading through analyst predictions last week and one line from Venky Ganesan at Menlo Ventures stopped me cold: "2026 is the 'show me the money' year for AI." Companies have burned through billions on AI pilots and token counting. Now boards want to see actual dollars (Axios, January 2026).
The reality check is brutal. James Brundage from EY put it plainly: "Boards will stop counting tokens and pilots and start counting dollars." What really got my attention though is Ganesan's warning that some major companies might go bankrupt this year from aggressive AI spending. We're not talking about startups here. We're talking about companies you know.
Here's the weird part: while executives demand results, most AI agents are just... sitting there. Ryan Gavin at Salesforce has a term for it: "the lonely agent." Companies are spinning out hundreds of agents per employee, and they're collecting dust like unused gym memberships. Impressive on paper, invisible in practice.
AT&T's chief data officer Andy Markus explained the problem to me in a way that finally made sense: "In an agentic solution, you're breaking down the problem into many, many steps. And the overall solution is only accurate if you're accurate each step of the way." One weak link, and the whole chain breaks.
What this means for you: If you're building AI tools right now, reliability beats impressiveness. The demo-to-production gap just became the most expensive problem in tech.
IBM: 2026 Marks First Quantum Advantage
IBM just made a claim that sounds like science fiction: 2026 will be the first year a quantum computer beats classical computing on real problems. Not theoretical problems. Real ones (IBM Think, December 2025).
"We've moved past theory." That's Jamie Garcia, IBM's Director of Strategic Growth and Quantum Partnerships. They're not talking about demos anymore. They're running actual use cases on quantum computers right now. Not production scale yet, but the writing's on the wall for drug development, materials science, and financial optimization.
Here's where it gets interesting for developers. You don't need a physics PhD anymore. IBM built something called Qiskit Code Assistant, and it writes quantum code for you using AI. I know that sounds recursive (using AI to write code for quantum computers that will accelerate AI), but that's exactly what's happening.
IBM is building what they call a "quantum-centric supercomputing architecture." Think of it as quantum processors working alongside regular CPUs, GPUs, and AI accelerators. AMD is in on it too, integrating their hardware with IBM's quantum computers to solve problems that neither system can handle alone.
One more thing worth noting: IBM researchers predict "2026 will be the year of frontier versus efficient model classes." Translation? Not every problem needs a billion-parameter model. Smaller, hardware-aware models running on modest accelerators will sit right next to the giants.
The takeaway is simple: the computing stack is getting weird, and that's probably a good thing.
CES 2026: Physical AI Takes Center Stage
Something shifted at CES this week (January 5-8). Forget chatbots. The big theme isn't software anymore. It's AI that moves, acts, and exists in physical space (Euronews, January 2026).
Both Nvidia CEO Jensen Huang and AMD CEO Lisa Su are keynoting, and according to industry analyst Tim Bajarin, we're about to see "an overabundance" of autonomous vehicles and humanoid robots. Samsung brought an AI teaching assistant bot. Not a concept. A physical robot deployed in universities right now.
Bajarin explained it like this: "AI that's not just in software in your computer, but now manifests itself physically, which includes automotive. So full-on autonomy, self-driving cars, and then autonomous humanoid robots."
Here's why this matters for developers. Physical AI demands different architectures than what we've been building. LLMs predict the next word. That's pattern matching. Physical AI needs to understand how a coffee cup falls when you knock it over, or how to grip a door handle without crushing it. These are "world models."
The movement is real. Yann LeCun left Meta to start a world model lab. He's reportedly seeking a $5 billion valuation. Google's DeepMind launched real-time interactive world models. Fei-Fei Li (you know, the ImageNet person) just released the first commercial world model called Marble through her company World Labs (TechCrunch, January 2026).
One last thing: Bajarin mentioned that AI bubble fears have mostly disappeared. "I lean toward, this is again, much more of a build-out than a bubble." Make of that what you will.
The Transformer Plateau: Why AI Scaling Hit a Wall
Remember when everyone said AI progress was inevitable? Just add more GPUs, more data, more parameters, and watch the magic happen?
Yeah, about that.
Ilya Sutskever just confirmed what many suspected: transformers hit a wall. In a recent interview, one of the people who literally invented the transformer architecture said "current models are plateauing and pre-training results have flattened" (TechCrunch, January 2026).
For almost a decade, AI followed one simple rule: bigger equals better. This worked spectacularly. AlexNet in 2012. BERT. GPT-3. Each breakthrough came from scaling up. More parameters, more training data, more compute. The "age of scaling" took us from image recognition to ChatGPT.
Then it stopped.
Kian Katanforoosh, CEO of AI agent platform Workera, spelled it out: "I think most likely in the next five years, we are going to find a better architecture that is a significant improvement on transformers. And if we don't, we can't expect much improvement on the models."
Read that again. If we don't find something new, progress stops. Yann LeCun has been saying this for a while. Scaling alone won't cut it anymore.
So what's the move? The industry is pivoting hard toward fine-tuned small language models. Andy Markus, AT&T's chief data officer, told TechCrunch: "Fine-tuned SLMs will be the big trend and become a staple used by mature AI enterprises in 2026, as the cost and performance advantages will drive usage over out-of-the-box LLMs."
Translation: instead of one massive model that does everything poorly, we're moving toward smaller models that do specific things really well.
The party's not over. But 2026 is the hangover year. We're transitioning from brute-force scaling to actual research, from flashy demos to tools that work, from promises of autonomy to honest augmentation.
How was today's email?
