AI in 2025 Recap Webinar January 8th
This session is my full recap of what actually happened in AI during 2025, and what I think you should watch in 2026. A lot of people have no idea of the magnitude at which the ecosystem moved last year, so I walk through the history from 2023 to today, then present my podium of the three breakthroughs that truly changed the game, with live demos of each. It is dense on purpose, and it is built for business people, not developers.
How we got here: 2023 and 2024 (3:59)
Quick rundown for everyone who has not touched AI much. ChatGPT was released in November 2022, and back then we were making poems with it, writing texts, maybe cheating on homework. 2023 is when people started asking how to apply this to business. Image generation started to become a thing, the technology was not quite there yet, and smarter models arrived: GPT-4 impressed a lot of people, and features like deep research started showing up.
2024 is when the big companies joined the fight (5:34). Google created Bard, Anthropic entered, France entered with Mistral, and open source systems became more popular. The important thing about 2024 is that it set the stage for a race. ChatGPT is still the only application most people know, but there is so much more to AI beyond it, and that gap is exactly what I wanted to show in this session. 2024 is also when a new type of business appeared: people like me who automate processes in companies using AI.
2025, the year of AI agents (6:40)
At the beginning of 2025, agents were the buzzword. The reality is that at that point they were not viable at all, and putting those systems inside organizations was genuinely dangerous. I was helping people at the time, everyone wanted agents, and I kept telling them the technology was not there yet. That changed in the middle of the year, in dramatic fashion, with a tool I get to at the top of the podium. It is honestly the closest thing we have to a true virtual assistant.
The podium: cutting through the noise (7:54)
Every single week a model gets released and people scream that it changed the game. Most of that is noise. There are very few moments where you can honestly say: this changes the future of AI and business once and for all. I compiled three moments in 2025 where I got genuinely scared, and I ranked them.
Number 3: Antigravity (9:11)
Antigravity is Google's tool, and it is very simple. Programmers use IDEs to create software, and IDE is just a scary word that keeps normal people away from a simple interface for playing with folders on your computer. You might say: I am not a developer, this is useless for me. But what counts as a technical skill has changed enormously. Any person can open Antigravity, click on the Open Manager tab, and get a window that looks exactly like ChatGPT but is ten times more powerful. Why? Because it is not just connected to the internet, it has access to the files on your computer.
I demoed it live: I opened my desktop as the working folder and asked what was in it. It read the files, analyzed them, and reported back. That sounds small, but here is the big idea (11:12): everything you do on your computer is code, and Antigravity can create scripts that do things inside your computer. So it can automate any workflow you do, and you only need a few high-level skills to use it, not a programming education. It nearly took the number two spot.
Number 2: Nano Banana Pro (12:07)
Nano Banana Pro is Google's image generation model, and calling it that undersells it. In 2024 we used image generation for nice marketing assets. Nano Banana Pro is not an image generation model, it is a visual reasoning model. It understands abstract concepts like mathematics, and it understands what communicating with an audience means.
The demo that makes the point (14:08): I downloaded a university exam, a genuinely complicated quantum physics question, and asked Nano Banana Pro to generate the image of the solution, step by step. It produced the step-by-step calculations as an actual rendered answer sheet, from a single sentence. I verified the outputs; in one earlier run the second answer was false, and in the live run the equation checked out. AI is still probabilistic. But the fact that this is possible at all is a game changer.
Now the uncomfortable part (16:17): a model like this can also generate images of passports. I can upload an invoice and ask it to change the date or the amounts, and it becomes very hard to tell real from fake. That has serious consequences for education and for society, and it is one of the things 2026 will have to deal with.
It also understands nuance (17:39). I asked for a diagram of how stars are created and got an extremely precise scientific diagram. Then I asked it to redo the same content so a child could understand it, and it did. Then I pushed further: a line of gym supplements matching the steps of star formation, then the marketing for it, then even videos on top. All of this is available with the twenty dollar Google AI subscription. And one detail I enjoyed sharing: the presentation I used in this very webinar was created one hundred percent with Nano Banana. Image generation in 2026 stops being about cool pictures and becomes infographics, diagrams, and reports.
Number 1: Claude Code (20:07)
For the people who know me, they saw this coming: the number one spot goes to a tiny little crab named Claude Code, released by Anthropic in the middle of 2025. It is the closest thing to what one hundred percent AI truly feels like: something that reasons for itself, fixes things, and creates things for you. It is single-handedly the most important thing to happen in AI in 2025, and I am fully convinced of that. It is also the moment where people who build automations could themselves be replaced by another AI.
It lives on your computer, same principle as Antigravity, and the interface is just like ChatGPT (22:01). It can browse the web, and it can be connected to anything: LinkedIn, databases, your CRM, your email (23:34).
The centerpiece demo was the Observer workflow (24:43), which I built by connecting Claude Code to my LinkedIn account. I keep a database of all my LinkedIn posts with impressions, shares, and comments. The Observer analyzes my latest posts, finds everyone who commented, shared, or reacted, pulls their profiles, analyzes each one, and compares them against the type of people I want to talk to. I ran it live by voice, limited to 10 profiles. A previous run had analyzed two posts from the last thirty days: eighteen valid engagements, seventeen unique engagers, eleven high-intent leads, with names reported back.
On scoring (28:10): ICP means Ideal Customer Profile, and you grade leads out of one hundred with whatever grid you want. French speakers get five points, working in Europe adds ten, or you only accept C-level executives at companies above five hundred people. Completely up to you.
Then the outputs (29:05): I asked for a markdown report on my desktop and got it. I asked for an HTML dashboard and thirty seconds later I had a live dashboard segmented by sentiment, ICP score, decision-makers, and interest. I asked whether these people were already in my Notion CRM, and Claude Code checked and offered to add the missing ones. That is a personal LinkedIn analyst (31:19): live metrics, qualified engagers, and analysis of what content performs best. ChatGPT does not have those capacities, and businesses that think ChatGPT alone is enough are wrong. I built that entire pipeline in about ten minutes.
You can also hand Claude Code an image model (33:28): I connected it to Gemini image generation and asked for a cat image, and it found the right tool in its toolbox and ran it.
Browsing, reporting, and a live build (41:44)
Claude Code can also literally use a browser: open pages, click on things, take screenshots. I sent it to a website and had it walk through the pages one at a time. This is useful for UI and UX work and for verifying that things are set up correctly. You can even have it study a site's structure while building yours, though be warned, Claude is a bit of an ethical warrior and will caution you about copying.
Reporting (44:00): reporting eats enormous amounts of time. I use Instantly for cold mailing, and I connected it to Claude Code so that when a campaign finishes, the entire report generates automatically: a master summary with every positive reply, every out-of-office, every wrong-person response, all the emails, and everything synced to my CRM. That took me about fifteen minutes to set up. You cannot automate one hundred percent of a person's job yet, but you can automate eighty to ninety percent of it.
Then a genuine live build (46:25): I will be doing door-to-door sales soon, so I wanted local businesses near me. Apify is a library of scrapers, tools that pull information from the web, and it has a Google Maps scraper. I pasted the link into Claude Code and asked it to check what the actor can do via the Apify API. It read the documentation and reported back with capabilities. The lesson (48:26): Claude Code is not a scraping tool, it is a brain. A very smart intern. What you need to learn is how to connect that intern to the tools that let it do very cool things.
That leads to a prediction (49:00): watch the people in 2026 who stop buying CRM and project management licenses. If I need a CRM, I ask Claude Code to build one custom-made for me with exactly the two features I need, instead of paying for a Salesforce or HubSpot ecosystem full of features I never touch. That saves money and the time of learning someone else's tool.
Q&A highlights
Lucas asked how to make these agents reliable daily, for example scheduling an automation every day at 8:00 PM (34:08). Yes, it is possible, and the real answer is a cron job, but the reflex I want you to build is to ask the tool itself: put Claude Code in plan mode, describe what you want, and it comes back with the options and then sets it up for you. You do not need to know how to write code, only what the code is doing.
He followed up on verification (36:28): with Make.com or n8n you see the workflow node by node, while this feels abstract. That is a perfect use case for Nano Banana Pro: I asked Claude Code live to generate a diagram of the Observer workflow using Gemini image generation, for about twelve cents. You get the visual element back. And on n8n versus Claude Code in general (39:43): n8n and similar tools are still amazing precisely because they are user-friendly for people who do not want to touch code; if you work on AI automation yourself, use Claude Code.
Anwar asked about rate limits on Claude Code; that has not happened to me yet. He also asked whether I use the Gemini models in Antigravity: I actually run Claude Code inside Antigravity rather than using the Antigravity agent itself. And on connecting a system with no public API (55:05): give Claude Code access to the web and it does the browsing for you. I showed the Chrome extension acting in my place, booking a call through the contact form on my own website.
We closed on education (57:02): AI lets curious people learn much more, much faster, and universities are falling behind. Junior roles are heavily execution-based, and AI now does the execution, so education is going to have to change.
What to watch in 2026 (59:05)
Two things. First, every model is already incredibly smart; the bottleneck is not intelligence, it is memory, the context window. If 2026 brings big context windows and big memory, the game changes. Second, image and video generation will be huge thanks to Nano Banana Pro and what gets built on top of it. So when the newest, hottest model drops, that is dust in the wind. But if models start getting better at remembering, pay attention.
I also opened resources during the session (50:38): fifteen-minute setup slots for Antigravity and Claude Code on my website, a resource page with all the webinars, and step-by-step YouTube tutorials starting from the foundations. The mindset for 2026 is being AI native: it is not about tools, it is about skills that transfer from tool to tool. If you install Antigravity and start playing with Claude Code, you are already ahead of ninety-nine percent of people who adopted AI.
New sessions run every week. Catch all the recordings on the webinars page.
