What if every software spoke English? We asked this question about two years ago but now they do - with AI we can retrofit existing apps to speak English.
I don’t want to have to figure out any particular menu to find a setting or understand how a product manager or designer intended me to use the product.
I just want to talk to my computer and tell it what to do.
AI agents are increasingly outperforming humans in various tasks, yet they typically cost 70% to 80% less.
Will they ever be able to charge a premium?
Waymo has reduced accidents by 82 to 92 percent in San Francisco. Waymos monitor more sensors, don’t fatigue, and react more quickly than humans. But, Waymo is often 13-33% cheaper than alternatives.
Within medicine, recent studies suggest AI can be at least as accurate as human doctors, in the evaluation of rashes from smartphone photos, estimating longevity, and diagnosing medical case histories, scoring 90% accuracy compared to human doctors who averaged in the mid-70s.
Nobody wants to read your stuff.
Writing is about the reader - the infinitely busy reader who has a thousand things to do & three goals to accomplish by the end of the quarter. What will attract the reader? What sentence will propel them to the next phrase to the end?
In the past, great content marketers have segmented readers into personas, written witty hooks to entice visitors to dwell.
The blocky charts. The ability to solve a hard test problem. The hidden game of Snake. My graphing calculator was a 7th grade miracle.
AI is this generation’s graphing calculator. With about 2 years’ of studies, we can draw some conclusions on its impact. A meta-analysis published in Nature showed a medium to large impact on students.
ChatGPT’s Select Impact on Higher-Order Thinking
| Condition/Scenario | Effect Size1 | Significance (p-value) |
|---|---|---|
| Overall Effect | 0.457 | < 0.001 |
| Type of Course: STEM and related | 0.737 | < 0.001 |
| Learning Model: Personalized learning | 0.718 | < 0.001 |
| Learning Model: Mixed | 0.719 | < 0.001 |
| Duration: 4-8 weeks | 0.654 | < 0.001 |
| Role of ChatGPT: Intelligent tutor | 0.945 | < 0.001 |
| Role of ChatGPT: Intelligent learning tool | 0.428 | < 0.001 |
| Area of Application: Tutoring | 0.478 | < 0.05 |
The higher the Effect Size, the greater the Effect. 0.3-0.5 is considered a medium sized effect. Above 0.5 is considered a large effect.
With models rapidly commoditizing in performance, we are seeing different strategies for keeping users on the same models.1
Ultimately, personalization & applications are likely to be the two vectors by which foundational model companies compete.
A model remembering your name, optimized to your codebase, having a knowledge of your previous work & refashioning itself to the way you work : those are reasons for loyalty, even if the model isn’t state of the art. Like airplane reward programs, personalization & memory introduce switching costs that may outweigh the benefits of state-of-the-art models.
As traffic to traditional websites plummets due to AI answering user queries directly, there is a new explosive form of distribution.
“AI agents are now creating Neon databases at 4x the rate of human developers, driving new requirements for instant provisioning, automatic scaling, & isolated environments.”
If I ask an AI agent to create a web application, I want it to select all the components : the front end framework, the database, & the hosting service. I just want the website to work.
Election night struck the regulatory asterisk from web3. But it did more than that.
It triggered a broader shift of application investing versus infrastructure.
The last few years of crypto & Web3 investing have focused predominantly on infrastructure, the databases (called Layer 1s/L1s & Layer 2s/L2s), security, analytics, & decentralized finance or DeFi (typically lending products). But these categories are slowing.
“We processed over 100t tokens this quarter, up 5x year over year, including a record 50t tokens last month alone.”
If the market harbored any doubt for the insatiable demand for AI, this statement during Microsoft’s quarterly earnings yesterday, quashed it.
What could this mean for a run rate? Using some basic assumptions1, this implies :
| Scenario | Model mix (% of total tokens) | Monthly run-rate after 20 % discount | Annual run rate | % of Azure Revenue (assuming $21B Annual) |
|---|---|---|---|---|
| High | OpenAI 70 % • Claude 20 % • Other 10 % | 382.9 | 4,594.8 | 21.88% |
| Medium | OpenAI 65 % • Claude 20 % • Other 15 % | 110.5 | 1,326.0 | 6.31% |
| Low | OpenAI 60 % • Claude 20 % • Other 20 % | 27.3 | 327.6 | 1.56% |
So AI is roughly between 2 to 22% of Azure revenue. Error bars here are quite large, though.
By 2026, AI agents will consume 10x more enterprise data than humans, but with none of the contextual understanding that prevents catastrophic misinterpretations.
In this presentation I shared yesterday, this is the main argument.
Historically, our data pipelines have served people. We’ve architected complex pipelines to ingest, filter, and transform information in different systems of record: cloud data warehouses, security information and event management systems (SIEMs), and observability platforms.
We then interpreted these outputs and acted upon them.
If every employee starts managing agents, how does a company change?
First, “83% of global leaders say AI will let employees take on more complex, strategic work earlier in their careers.”
One executive recently framed this transition of teams evolving to three areas of work : operational, tactical, & strategic. Operational work can be mostly fully automated today. Agents are chomping away at tactical work - better accuracy will improve their share. Humans will focus on strategy work, but likely assisted by AI.