Antonym: The Horseless Edition
AI is not a software update.
Dear Reader
When a new technology appears, our first instinct is to bolt it onto the old contraption. We build horseless carriages instead of imagining motorways.
Y Combinator’s Pete Koomen published an essay this week: “AI Horseless Carriages” which explains how this effect is playing out with AI and existing software.
He uses the example of Gmail and how terrible it is at using Gemini (Google’s now very capable ChatGPT competitor) to write emails. It’s performance at a seemingly mundane, everyday chore that should be well within the capabilities of this new wonder tech, but Gmail invariably flounders. You see similar weirdly poor outputs in Docs, Slides etc. too, and their Microsoft equivalents Word, PowerPoint etc.
Koomen describes how tech – from humble word processors to mighty enterprise systems – in terms of two layers of a prompt sandwich:
System prompt – the silent set‑up that defines the model’s role, guard‑rails, and style. Think of it as the rules of the game.
User prompt – the visible request you type, played out within those rules.
His point is that most mass‑market apps keep the system layer on lockdown, so we can only click around the fenced‑in user features. Truly AI‑native apps, however, hand us the spanner: we can tweak or even rewrite the system prompt, reshaping how the tool behaves.
Koomen says:
Up until very recently, if you wanted a computer to do something you had two options for making that happen:
1. Write a program2. Use a program written by someone else
Programming is hard, so most of us choose option 2 most of the time. It's why I'd rather pay a few dollars for an off-the-shelf app than build it myself, and why big companies would rather pay millions of dollars to Salesforce than build their own CRM.
The market won’t wait for incumbents to figure this out. AI-native apps are arriving at pace and it may be that users will be making their own solutions as “build don’t buy” becomes a realistic prospect for more companies:
The lesson: pushing GPT into 20+ year old dashboards misses the opportunity, the imperative of AI: it changes what works at a systems level, something that is becoming apparent for organisations and markets as much as for apps and SaaS platforms.
We need to design from first principles – likely, what are we trying to do here? – or risk arriving at the racetrack in a carriage without a horse.
Microsoft at 50: guns, tariffs, and incremental AI
Microsoft holds excellent cards in the AI game, but maybe not all in the same hand. This week a shocking stat emerged: the 20 million weekly users of Co-pilot has not grown in a year, according to reporting by The Newcomer.
In comparison, OpenAI Increased its weekly active users to 400 million by February 2025, and following the release of its updated image generator, it is now rumoured to have 800 million weekly active users.
Two years ago Microsoft looked like it was winning the Big Tech AI race before it even started. It was able to call the shots around the failed Sam Altman boardroom coup at OpenAI, had a far better user-friendly AI system than Apple or Google in Co-pilot. Last year the company hired Mustafa Suleyman, one of the co-founders of Google Deepmind as its AI CEO, another power move from a giant that had apparently learned to dance.
According to The Newcomer piece, Suleyman has struggled with internal politics, but also the relationship with OpenAI (allegedly not on good terms with Sam Altman).
Microsoft AI recently launched a revamped Copilot but…
[…] didn’t transform the narrative. And at this point, Microsoft is just not in the running to build a model that can compete with the best from OpenAI, Anthropic, Google, and even xAI. The projects that people have mentioned to me feel incremental, as opposed to leapfrogging the competition.
The impediments of organisational culture echo the famous “guns” depiction by engineer and cartoonist Manu Cornet:
The stakes in the big tech AI race are so high that decisive leadership and fast changes may yet turn things around at Microsoft, but it is striking how much the company has ceded to OpenAI and Google.
The Newcomer’s verdict:
At this point, Microsoft is just not in the running to build a model that can compete with the best from OpenAI, Anthropic, Google, or even xAI
Further, even the best of the company’s products are hard to find. Wharton professor Ethan Mollick pointed out this week that he had trouble finding a promising sounding custom Copilot product.
For a while in 2023, Microsoft, with its GPT-4-powered Bing, was the absolute leader in making LLMs accessible and easy to use. Now it is very opaque, though the systems certainly seem powerful in the demos, they seem impossible to access or scattered throughout the ecosystem.
Mollick went on to produce a Deep Research report with Google Gemini on how to find out how to access the feature. It was discovered several clicks down in the menus, only if certain settings had been activated by the user and an administrator. Seems like Microsoft needs to have its own digital transformation and strip out some complexity to set its potential free.
🔗 Useful things worth your clicks
Kevin Weil on Lenny’s Podcast – OpenAI’s Chief Product Officer chats about must‑have AI skills, moats, and why coding is table stakes. I’ve referenced this daily since it dropped.
Ethan Mollick: “On Jagged AGI” – a lucid explainer of why models feel superhuman in one task and toddler‑esque in the next.
Every: “Vibe Check: OpenAI’s O3, GPT‑4.1, and O4‑Mini” – side‑by‑side impressions after a week of prompt‑breaking mayhem.
This week
Because life can’t be all large language models, some arrogant recommendations of things I like.
Reading
“It seems so odd to me now, how one can be so unsettled by the improbable… our entire existence is founded on freak occurrences.”
— On the Calculation of Volume, by Solvej Balle
Nominated for the International Booker prize, this short novel is the first in a series about a woman for whom time has collapsed, and she lives the same day again and again, close to but increasingly distant from her life.
Watching
The Last of Us – Series two is on the money so far. (NowTV)
Somebody Somewhere – gentle Midwestern melancholy on (NowTV)
Dope Thief (Apple) - intense and brilliantly gritty thriller shot through with humanity and humour. (Apple)
That’s all for this week…
Until next time—mind the horseless carriages on your commute. If you liked this don’t forget to like this with the little 🤍 that turns ❤️ when you click it.
Antony