What to expect in AI for 2025
OpenAI's GPT 4.1, Google's hub for AI Agents, Nvidia's impressive investments, GPT-5 delayed.
Good Morning,
The expectations for Generative AI in 2025 were high, and while OpenAI skeptics do some fun napkin math, the world moves on. In truth, the loudest OpenAI critic isn’t Gary Marcus, it’s Ed Zitron.
I’m actually more bullish on OpenAI’s prospects before 2025, than I was before. When your moat is scale, the trajectory is getting weird. OpenAI’s ChatGPT has nearly reached 800 million people as of April, 2025.
Recently, Sam Altman has acknowledged OpenAI was “late to act” (or on the “wrong side of history”) on open source but now aims to release a model “better than any current open source model.” Given the explosion of open-source models out of China and the West trying hard to ignite Agentic AI, I thought it was a good time to bring Kenn’s holistic perspective to light.
Today I want to feature the incredible piece by
of the Newsletter Generational: What to expect in AI for 2025.Newsletter pieces:
Kenn So is among my favorite writers on AI who brings a grand synthesis of contexts and industries together. Among Angel investors and VC, Kenn is one of those hidden gems of perspectives on Generative AI.
The Elephant in the room this week is GPT-4.1
OpenAI has just launched GPT-4.1, a powerful upgrade to the GPT-4o multimodal model. Orion will be entirely winded-down.
With Google and Anthropic pulling ahead, OpenAI needed a model that was good at coding and OpenAI’s new GPT-4.1 AI models does focus on coding.
Of course what this actually means is the unified GPT-5 model will be even further delayed. Enter GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, all of which OpenAI says “excel” at coding and instruction following.
About Kenn & Generational
Kenn writes Generational, a publication about what matters in the business and technology of AI. Having held corporate, investor, and technical roles, Kenn brings business realism to technological hype and technological grounding to business hype.
OpenAI says GPT‑4.1 mini is a significant leap in small model performance, even beating GPT‑4o in many benchmarks.
OpenAI’s grand ambition is to create an “agentic software engineer,” as CFO Sarah Friar put it during a tech summit in London last month. However everyone is racing to Agentic AI and Google might be the most impressive one of all.
It’s getting hard to get excited about new models due to their speed of arrival which we sort of expected in 2025, a year where we have over $400 billion in capex on AI Infrastructure and a supposed explosion of business applications for the AI hype trend.
Can AI Code in 2025?
OpenAI says that GPT‑4.1 is significantly better than GPT‑4o at a variety of coding tasks, including agentically solving coding tasks, frontend coding, making fewer extraneous edits, following diff formats reliably, ensuring consistent tool usage, and more.
Nvidia on Tap 🚀
Nvidia is racing to build its Blackwell AI chips that have started production in Phoenix at Taiwan Semiconductor (TSMC) plants. If anything, 2025 will be the year that TSMC and Nvidia become even more important to future of Generative AI and scaling into more useful products for society.
Nvidia said on Monday that it has commissioned more than a million square feet of manufacturing space to build and test AI chips in Arizona and Texas as part of an effort to move a portion of its production to the U.S. The PR pitch is a hilarious gesture to the Trump Administration’s talking points.
Within the next four years, NVIDIA says it plans to produce up to half a trillion dollars of AI infrastructure in the United States through partnerships with TSMC, Foxconn, Wistron, Amkor and SPIL. These are mostly Taiwanese companies by the way.
Google and Agents
How close is real Agentic AI? Well, it depends on who you talk to. Google’s new Agent Development Kit (ADK) and with other additional capabilities that offer control over how agents behave really shows a lot of promise. This as MCP appears to have gained a lot of traction in early 2025.
With the excellent launch of Gemini 2.5 Pro, Google is holistically in a very good position to lead in the Agents space. At its Google Cloud Next 25 event, the company unveiled several upgrades to its Agentspace platform to make agent discovery and adoption easier.
Google claims users can “build an AI agent in under 100 lines of intuitive code” with ADK. The platform also supports the Model Context Protocol (MCP), the data connection protocol developed by Anthropic that helps standardize data movement between agents.
ADK is optimized for Gemini models, though Vertex AI allows access to models from Anthropic, Meta, Mistral, AI21 Labs, CAMB.AI and Qodo.
Nvidia and Export Controls
Nvidia reportedly narrowly avoided export controls on its H20 chip after striking a domestic manufacturing deal with the Trump administration. Even Nvidia’s CEO has had to do a MAGA take to get this deal apparently:
“Adding American manufacturing helps us better meet the incredible and growing demand for AI chips and supercomputers, strengthens our supply chain and boosts our resiliency,” CEO Jensen Huang said.
Taiwanese semi leaders and Taiwanese-American CEOs can “kiss the ring” too just don’t expect them to ever wait up for America’s lagging semiconductor industry. Intel’s dismantling has begun, with Intel Corp. has agreed to sell a 51% stake in its programmable chips unit Altera to Silver Lake Management, making good on plans to start spinning off non-core assets. They need in my opinion to divest a majority stake in Mobileye next as well, as well as conduct another round of layoffs.
Trump has strong-armed certain partners to get his desired outcome in recent months. He reportedly told TSMC that it would have to pay a tax of up to 100% if the company didn’t build new chip factories in the U.S. Trump has practically made Taiwan a vassal state for U.S. semiconductor and national security interests. Not the best move for relations with China and initiating a major trade war catapulting the world into geopolitical and economic purgatory and uncertainty.
Open Delays GPT-5 yet Again
As far as GPT-4.1 goes, it’s yet another delay for GPT-5. Altman added that OpenAI expects to roll out GPT-5 “in a few months” — later than originally anticipated. Not clear what that means exactly, as Sam Altman is known for being vague, cryptic and over or under promising strategically. OpenAI’s PR is known for hyping things that don’t matter at all in the big picture.
It’s a bit hard to get excited about benchmarks too after all the models we’ve witnessed in the last 18 months, but here goes another one:
GPT-4.1 boasts better Instruction Following, Coding edits and Long Context
GPT‑4.1, GPT‑4.1 mini, and GPT‑4.1 nano can process up to 1 million tokens of context—up from 128,000 for previous GPT‑4o models. Read the details.
2025 is shaping up to be a Stellar year for Generative AI’s progress.
Let’s get into our deep dive of today.
By
late 2024 and early 2025.What to expect in AI for 2025
Agents, glasses, and AVs go mainstream
If I had to sum up AI’s progress in 2023 and 2024, I’d say 2023 was about realizing what models could (and could not) do, and 2024 was about building the data centers and software infrastructure to deliver on the opportunity of these models. Both years felt packed—new releases seemed to drop every week—but they were merely ramping up to an even more frantic 2025.
What makes 2025 pivotal is the intersection of three major developments. First, mainstream AI adoption is accelerating among consumers and businesses. Second, models are rapidly becoming better, faster, and cheaper. Third, tools for building AI applications have matured. These shifts are reinforcing one another across every industry and sparking shifts in consumer behavior.
Subscribe for full-access to less than $2 a week.
This is a deep dive full of innovation trends, infographics, visuals and expert synthesis of AI’s trajectory.
Keep reading with a 7-day free trial
Subscribe to AI Supremacy to keep reading this post and get 7 days of free access to the full post archives.
 | A guest post by
|