Explainable AI

State of AI (2025 version) | December 2025

Hi friends,

Fade to black. I’m in Southern California so using a bit of industry lingo. So, we can now look back on a busy 2025 and put it in the history books. However, if you’re like me, you can’t wait for 2026 which is shaping up to be a transformational year in tech and AI.

I am not saying that because I’m an optimist, but I see what we’ve built here at Straylight and can only feel excitement. From an industry perspective, 2025 seemed like we were still on a learning curve for AI. Remember Agentic AI was the rage? What about RAG to save the day? Nope. As we look back and think about where we started the year, we have never strayed here from the idea of an ‘AI Factory’. This means a smaller, real-time, low-cost, accurate, and more contextualized AI.

2025 gave us Executive Orders on AI and now some states are passing laws attempting to regulate against frontier models. Wall Street and real estate developers are on a tear building data centers. How are we going to power this development? Some of you know we are involved in some projects utilizing geothermal and small modular reactors to power the AI infrastructure we need.

2025 even saw data centers move into the political arena and look to be an issue for the 2026 midterms. By the way, there seems to be no stopping NVIDIA. The hyperscalers need those GPUs. Well, Masa Son doesn’t agree. We started to hear more about AGI in 2025 and how we are probably still a decade away from it. AI experts like Yann LeCun think utilizing frontier models or LLMs won’t get us there. I agree. But if we look to some of the models being developed by DeepSeek, for example, we may be closer to AGI in 5 years.

Look, I could talk a lot about different aspects of AI. Given that I used to work in government and focused on data science (meaning spending hours manually data wrangling), I always believed that AI should solve highly complex problems and augment work to allow workers to spend more time on analysis. And our team at Straylight, who have spent the last few years building an incredible AI engine, could use our AI to invent new things or design new and less costly materials for manufacturing. We are currently in a global rare earth minerals race and utilizing Straylight’s AI to design and test new elements for material science is, ultimately, what we are building to serve a greater good.

Here are some other 2025 takeaways:

World Models

I don’t use ‘World Models’ in my lexicon but this is what we at Straylight decided to embark on over 2 years ago. First, what do I mean by a world model? The bottom line is that today, the LLMs we rely on have been trained by available data, including text, imagery and videos. A world model works by creating neural networks that understand the dynamics of the real world, including physics and spatial. As I mentioned before, LLMs aren’t going to cut it for AGI. But more importantly, world models are the key to understanding the physical world.

We will be talking about this at length in 2026 as more of you are becoming aware of the different types of AI. I don’t complain but the biggest frustration I had in 2025 was explaining how LLMs and ChatGPT are not what we are building at Straylight. To put a finer point on this topic, if your business relies on exact outcomes or you need an AI to better understand spatial data to achieve contextualized reasoning, world models are what you are looking for.

OpenAI vs. Google

In 2025, it seems clear that both ChatGPT and Gemini 3 are in a league of their own. ChatGPT continues to dominate in usage and consumer traction. Yet, performance goes to Gemini 3. It makes sense if you think about it. When you’re talking data, a key component of building AI, Google reigns king and can easily leverage the amount of data it owns to train its models.

ChatGPT is cool. Yet, most users go there for the free version and ask it all sorts of general questions to make their daily commute better or generate content with they have writer’s block. Enterprise are struggling to see ROI with ChatGPT according to an August MIT survey.

I’m not in the prediction game but many are expecting 2026 to be the year of greater AI adoption by enterprises. Maybe it is due to greater awareness of how AI is transformational or maybe it is simply FOMO. Google and OpenAI will continue to push each other. But we are all waiting to see what Apple rolls out. Now you see why I’m excited for 2026!

AI News (Some of our favorite AI articles from December)

• 404 Media: Flock Uses Overseas Gig Workers to Build its Surveillance AI

• WSJ: Say Goodbye to the Billable Hour, Thanks to AI

• VentureBeat: Mistral Launches OCR 3 to Digitize Enterprise Documents, Touts 74% Win Rate and $2-per-1K-Page Pricing

• WSJ: US Investors are Going Big on China AI Despite Concerns in Congress

• CNBC: Instacart Ends AI-Driven Pricing Tests that Pushed Up Costs for Some Shoppers

• The Information: DeepSeek is Using Banned Nvidia Chips in Race to Build Next Model

• TechCrunch: Yann LeCun Confirms His New ‘World Model’ Startup, Reportedly Seeks $5B+ Valuation

• Time: How Christian Leaders Are Challenging the AI Boom

• TechCrunch: New York Governor Kathy Hochul Signs RAISE Act to Regulate AI Safety

• WSJ: We Let AI Run our Office Vending Machine. It Lost Hundreds of Dollars

• Bloomberg: Rakuten AI Boss Diverges From Big Tech in Prioritizing Low Cost

• TechCrunch: John Carreyrou and Other Authors Bring New Lawsuit Against Six Major AI Companies

• WSJ: The AI Boom is Opening Up Commercial Real-Estate Investing to New Risks

• NYT: As AI Companies Borrow Billions, Debt Investors Grow Wary

OpenAI

• WSJ: Sam Altman’s Sprint to Correct OpenAI’s Direction and Fend Off Google

• Bloomberg: Oracle’s $300B OpenAI Bet has Fast Become a Bubble Barometer

• Axios: OpenAI Updates ChatGPT After ‘Code Red’ Scramble

• Axios: OpenAI isn’t Too Big to Fail. It’s Bigger

• The Information: ChatGPT Nears 900M Weekly Active Users but Gemini is Catching Up

• TechCrunch: OpenAI is Looking for a New Head of Preparedness

• The Information: OpenAI has Discussed Raising Tens of Billions at Valuation Around $750 Billion

Gemini 3

• Google Blog: Gemini 3 Flash – Frontier Intelligence Built for Speed

• Axios: Google’s New Gemini 3 Flash is Fast, Cheap and Everywhere

• The Verge: Gemini 3 Flash is Here, Bringing a ‘Huge’ Upgrade to the Gemini App

Energy

• Bloomberg: Nuclear Developer Proposes Using Navy Reactors for Data Centers

• TechCrunch: Trump Admin Halts 6 GW of Offshore Wind Leases Again

• Politico: BLM to Accelerate Geothermal Lease Sales on Federal Land

Want to see what Multimodal AI can do for you? Reach out to us at Straylight.

Thanks for reading. Wish you all a prosperous 2026!!

Sergio Rodriguera Jr

Co-Founder | Straylight Systems

Straylight Systems is a Multimodal Artificial Intelligence (AI) company powering Sensor Fusion and Workflow Automation utilizing its own neural networks delivering tools to automate and optimize solutions for commercial and government customers. Straylight synchronizes text files, acoustic, video, imagery, radar, lidar and spatial data from different sensors and systems to process and analyze in real time in an explainable and auditable way.