Thomas Smith's six predictions from his December 2025 Fast Company article offer an interesting snapshot of industry expectations at the time. Let me break them down:

1. "OpenAI goes garlic" - Smith predicted OpenAI would release a fully retrained "Garlic" model in January 2026 to counter Google's gains. This reflects the intense competitive pressure OpenAI was facing, but it shows the prediction-game challenge: models evolve continuously rather than in dramatic single releases. The rumored "code red" captures the very real competitive anxiety at the time.

2. Google's Gemini domination - This is the most credible prediction. Google确实有more data, chips, and distribution channels. The historical analogy about first-mover advantage being overrated is spot-on—Google's slow-but-massive integration strategy could indeed make Gemini ubiquitous through Search, Workspace, and Android.

3. Chatbots as therapy + "adult mode" - The mental health trend was already visible in 2024, but 25% adoption seems high. The bigger issue is the serious safety concerns and regulatory pushback that would emerge. The "adult mode" rumor feels more like Silicon Valley hype than a responsible product direction, especially given mounting AI safety pressures.

4. AI video explosion beyond Sora - This is already happening. By late 2024, AI-generated video was flooding social feeds. The key insight is that vertical video's constraints (10-second clips) match current AI capabilities perfectly. The danger of blurring reality is very real.

5. Electricity as limiting factor - This is perhaps the most prescient prediction. Data center power consumption was already a major concern in 2024, and the "populist backlash" over electricity rates is plausible. The green tech spin is optimistic but possible. This structural constraint is often underestimated in pure tech predictions.

6. AI invades physical world (self-driving, robots) - Smith overstates the speed here. Waymo's expansion is methodical, not explosive. Robot baristas and caregiving machines face huge regulatory and trust barriers. The "blinking and suddenly all cars are self-driving" is classic boosterism—physical world deployment is always slower than digital.

Given that today is January 1, 2026, let me frame my perspective as analyzing key trajectories rather than crystal-ball gazing:

The pure scaling law obsession (just add parameters) will officially plateau. 2026 will see the dominant paradigm shift to hybrid architectures: mixing dense and sparse expert layers, integrating symbolic reasoning modules, and using test-time compute more efficiently. We'll stop asking "how big is your model?" and start asking "what's your reasoning budget?" The "Garlic vs. Gemini" framing will look quaint as the gap narrows—most top models will be within 5-10% of each other on benchmarks, with differentiation coming from specialization, not raw power.

After billions spent on POCs in 2024-2025, 2026 becomes the year of the integration reckoning. Companies will discover that slapping a chatbot on legacy systems creates more problems than it solves. The winners won't be AI model companies but enterprise integration firms that can actually rewire business processes. Expect a wave of "AI transformation officer" hires and a backlash when promised 30% productivity gains fail to materialize without massive workflow redesign. The real story is messy enterprise change management, not flashy model releases.

Text-only models will feel as outdated as black-and-white TV. By mid-2026, any serious AI system will natively understand and generate across text, image, audio, and video in unified contexts. The breakthrough won't be better video generation, but true cross-modal reasoning—asking AI to "look at this diagram, listen to this lecture, and find the contradictions." This will unlock scientific research applications (analyzing microscopy images alongside research papers) but also create new forms of misinformation that combine fake documents with fabricated source videos.

The EU AI Act's full implementation, state-level laws in the US, China's AI governance, and India's emerging framework will create a regulatory patchwork. Companies will respond by building region-specific model variants and compliance layers. This won't kill innovation, but it will Balkanize the AI ecosystem. A developer in San Francisco will use substantively different tools than one in Frankfurt. The "open" vs "closed" model debate will be overshadowed by "compliant" vs "non-compliant" markets. Expect AI export controls to become as complex as weapons regulations.

Smith touched on electricity, but the real story is architectural response. As data center power becomes expensive and politically fraught, there will be massive investment in efficient on-device and edge AI. The iPhone 17 and Pixel 10 will run surprisingly capable 7B-parameter models locally. Hybrid cloud-device workflows—where sensitive processing happens on-device and heavy lifting goes to the cloud—will become standard. This isn't just about privacy; it's about basic economics when electricity costs make centralized inference prohibitively expensive for non-critical tasks.

In 2026, AI will crack a high-profile scientific problem that captures public imagination—likely protein design for a specific disease target or materials discovery for room-temperature superconductors. This will be different from AlphaFold: it'll be a closed-loop system that designs, simulates, and validates its own hypotheses with robotic labs. The achievement will be real and important, but the hype will be worse than the "ChatGPT moment"—spawning unrealistic expectations that AI can instantly solve climate change or cure all cancers. The gap between specialized scientific AI and general problem-solving will be dangerously obscured.

The Meta-Prediction: The biggest AI story of 2026 won't be any single technology, but the normalization gap. The public will be exhausted by AI hype cycles while simultaneously becoming completely dependent on AI for daily tasks. We'll stop noticing AI (it's just how things work) but also never stop arguing about it. The "AI" branding itself may fade as these capabilities melt into products that are simply "smart" or "automated"—making Smith's model-versus-model predictions feel very 2025.