We wrap up 2025, and I'm reminded how transformative years like this are fueled by collective progress, bold innovations, and the occasional reality check.
It's been a standout year for 10Web, but the AI ecosystem mirrored that energy.
In Today’s AI Simplified:
Founder’s Insight: Reflecting on 2025 as 10Web's best year, the big AI shifts reshaping products, and our push into Vibe Coding for WordPress.
AI News: Karpathy’s year-in-review on LLMs, ChatGPT's "Spotify Wrapped" for users, and Anthropic's Bloom for testing AI behaviors.
AI Spotlights: A cinematic video tool redefining creation, social reactions to robotaxis stalling in SF, and a major AI expo kicking off 2026.
2025 was the best year in 10Web history — and it happened because you trusted us, challenged us, and kept building with us. This was also a defining year for AI. 3 big shifts changed how products like ours are built:
AI moved from “assistive” to “agentic” — not just helping, but building and executing.
Real ownership started to matter — users want real products, real infrastructure, and full control.
Generic tools started losing relevance — vertical, purpose-built AI began to win.
These shifts forced a question we couldn’t ignore: What does website building look like when AI is the core? Mid-year, we made a clear decision: stop improving the old model and redesign website building for an AI-first world, specifically for WordPress. A new generation of website building: we call it Vibe Coding for WordPress. The idea is simple: you describe the website you want, and AI builds it — design + pages + functionality, directly on WordPress. Not a prototype. Not a demo. A real website you can launch, own, and grow. Website builders have evolved in eras: 1.0 HTML / CSS / PHP 2.0 WordPress as the CMS 3.0 Wix / Squarespace / Shopify 4.0 AI-first website building — and we believe 10Web is leading this generation
We’re just getting started. In the coming months we’ll ship more game-changing features, including: • E-commerce support • Figma workflows & integrations • Website cloning • Website redesign • And much more
The second big win of 2025: our B2B products took off.
In September–October, we launched API, reseller, and hosting solutions so agencies, MSPs, SaaS platforms, and hosting companies can offer 10Web as their own branded product, while 10Web powers everything behind the scenes.
We already have hundreds of customers using these solutions, and we believe this will scale 10Web distribution globally.
If you want to launch an AI-first white-label website builder on WordPress, we’d love to work with you.
The part I’m most proud of? Our team worked incredibly hard this year — product, engineering, marketing, sales, and support — all pushing in the same direction.
But the most important thing is this: how satisfied our customers are. We’re not perfect yet. We know that. And we’re moving fast to close every gap.
Thank you for being part of this journey. If you have feedback (good or painful), just reply to this email — I read them.
AI researcher Andrej Karpathy reflects on big AI shifts in 2025. These changes go beyond making models bigger. The most important points from the review:
Reinforcement Learning from Verifiable Rewards: Emerged as a major new training stage. The AI earns automatic rewards for solving clear problems. This helps it learn step-by-step reasoning that looks very human.
Ghosts vs. Animals: LLMs’ abilities are very uneven: they become super strong in areas with clear, checkable tasks, but stay weak or easily fooled in others, leading to reduced trust in benchmarks.
Tools like Cursor made a new kind of "LLM app" popular. These apps manage many AI calls behind the scenes, handle context, add user-friendly interfaces, and let you control how much the AI does on its own. This points to a big new layer specialized apps that take general LLMs, fine-tune them, and turn them into experts for specific jobs or industries.
Claude Code / Local AI Agents: Anthropic's Claude Code demonstrated convincing agentic behavior running locally on users' computers, leveraging private context and low-latency interaction.
Vibe Coding: AI reached a threshold where complex programs can be built via natural language ("vibe") descriptions, democratizing programming for non-experts and letting professionals create more software rapidly.
Nano Banana: Google's Gemini Nano Banana hinted at the future of LLM interfaces, moving beyond text chats to multimodal outputs that better match human preferences.
Karpathy concludes: LLMs are both smarter and dumber than expected, vastly useful, with the field still wide open for innovation.
ChatGPT is rolling out “Your Year with ChatGPT,” a year-end review that shows how you’ve used the product over the past year, complete with playful visuals, usage “awards,” and personalized summaries. It’s available to eligible Free, Plus, and Pro users in select English-speaking markets, including the U.S., as long as chat history and saved memories are enabled.
OpenAI is turning ChatGPT into a habit:
Memory becomes a feature: The wrap only works if users allow ChatGPT to remember things, reinforcing long-term context as a core differentiator.
Lightweight, opt-in design: No forced pop-ups, privacy-forward, and user-controlled, a signal OpenAI is being careful with trust.
Emotional stickiness: Poems, images, and “awards” turn usage data into identity (such as, “Creative Debugger”).
What’s missing is telling: No Team, Enterprise, or Education accounts, this is clearly about individual engagement, not work analytics.
Spotify Wrapped made listening feel personal. ChatGPT is trying to do the same for thinking.
Anthropic released Bloom, an open-source tool designed to automatically evaluate how frontier AI models behave beyond what they do. Instead of relying on slow, hand-built benchmarks, Bloom generates large sets of scenarios to measure how often and how severely a specific behavior shows up in a model’s responses.
Traditional AI evaluations age fast. Models get trained on them, learn to avoid them, or simply outgrow them. Bloom is built to keep pace by generating fresh tests every run, while still staying reproducible and measurable.
How Bloom works:
Researchers describe a single trait they want to test (e.g. self-preservation, sycophancy, sabotage).
Bloom automatically creates many diverse situations designed to trigger that behavior.
The model is tested across these scenarios with simulated users and tools.
A judge model scores each interaction, producing clear metrics like frequency, severity, and comparisons across models.
Bloom is meant to keep the behavior constant while varying the scenarios. This is a big step toward AI safety tools that evolve as fast as the models they test.
Luma is best known for Dream Machine, it lets you generate high-quality, cinematic videos from text prompts with motion, camera control, and visual consistency.
Key features:
Text-to-video generation: Turn plain prompts into realistic, cinematic video clips with smooth motion and coherent scenes.
Camera & motion control: Guide camera movement, pacing, and framing instead of hoping the model guesses right.
High visual fidelity: Sharp details, believable physics, and fewer surreal artifacts compared to earlier video models.
Style consistency: Generate multiple clips that look like they belong to the same world; useful for ads, reels, and product stories.
San Francisco’s blackout stopped the autonomous future mid-intersection. After a substation fire cut power to over 120,000 customers, dozens of Waymo robotaxis stalled across the city, triggering photos, videos, and jokes on X. The main thread is autonomy still leans heavily on human-built infrastructure. When traffic lights went dark at scale, Waymo’s cars defaulted to four-way-stop logic and some froze longer than anyone expected. To many, it was a reminder that edge cases aren’t rare when you operate in real cities.
The debate quickly widened. Some applauded Waymo for pausing service and resuming within a day, framing this as a necessary stress test. Others weren’t so generous, calling out “immobile robots” and questioning whether AVs are ready for chaotic, failure-prone environments. Elon Musk poured fuel on the fire by claiming Tesla robotaxis were unaffected — a claim critics pushed back on, noting Tesla doesn’t run fully driverless service in SF. Net-net: no viral outrage, but a steady drumbeat of skepticism. The takeaway people keep circling back to is Real autonomy isn’t about better models, rather surviving when the city itself breaks.
AI Event Pick
AI & Big Data Expo Global
🗓️ February 4–5, 2026 | Olympia London, UK
The world’s leading enterprise AI and big data event unites 8,000+ innovators, 200+ expert speakers, and 150+ exhibitors to showcase real-world applications, generative AI breakthroughs, scaling strategies, and cutting-edge data solutions.
Expect in-depth conference tracks on sector-specific AI disruption, platform scaling, generative AI & intelligent interfaces, roadmaps for adoption, and taking prototypes to production — plus live demos, product launches, an AI Start-Up Area connecting founders with investors.
Register for your pass (free or Gold options available).
