/ / THEFUTURE /
In this issue: A model that cannot write essays, and why that’s the point. One in five UK adults follow the news with AI. Prisa Media’s AI director on cognitive delegation. Plus: The “xxx” trick that made me want to write with AI.
What we’re talking about: Small decisions, fast. Over the past few years, we’ve used LLMs for jobs that require just a simple answer to a bounded question. Questions like: Is this article an opinion piece? You can call these structured decisions. A large language model like ChatGPT will write you a whole text. That’s slow and overkill.
What we often need is a decision model. A company called TypeSafe has introduced its own model, Jev, to answer simple questions really fast, really cheap. Three modes:
- Noul: Evaluate how true something is. Is ‘food’ a sandwich?
- Score: Set up a rubric to grade with. How much did ‘subject’ contribute?
- Choice: Ask a multiple choice question. What color is ‘object’?
Example use cases in media: Content moderation, text classification, search in datasets. But it turns out, you can do a lot with very simple questions: Jev can even play Doom. More use cases presented here. So instead of forcing an LLM to answer in JSON, we just ask a model that cannot start writing essays.
Let’s look at some numbers: Jev answers in as little as 100 ms and it costs $0.042 for a million tokens (input only, output is free). In comparison, GPT-6-Luna takes seconds and costs $0.10-$0.20 for a million tokens in, $0.50-$0.75 out.
OpenAI has countered by announcing a dedicated decision API to GPT-6-Luna with similar speed advantages. Prices and availability tbd. In the meantime, you can try open decision models locally on your device with Ollaya. It works and looks just like Ollama.
What I’m reading:
And now: On a recent work trip to Madrid, I not only stumbled into a live broadcast of Los 40, but met my counterpart at Spain’s media conglomerate Prisa Media, Olalla Novoa Ojea. I love what she has to say and I’m looking forward to future conversations to exchange more ideas and learnings.
Three Questions with Olalla Novoa Ojea
Olalla Novoa Ojea is the Director of AI at Prisa Media in Madrid.
What's the most important question right now?
How to govern AI at the global, organizational, and personal levels.
At the global level, I’m concerned about how the geopolitical race and the interests of big tech are taking precedence over the need to put people and humanity at the center. We hear a lot about “alignment,” but in practice, it remains largely opaque, and I have the feeling that it is not being seriously pursued by the most relevant actors.
Within organizations, governance is essential if we want to use AI in a controlled, responsible, and constructive way without slowing down innovation or turning a blind eye to such a powerful technology.
On a personal level, I’m deeply concerned about cognitive delegation. I wonder what impact human-machine interaction is already having on the way we think and relate to one another. I find myself questioning how it will affect not only the way we live but also the evolution of our species.
I realize it may sound dystopian or exaggerated, but it’s truly a recurring thought for me.
What will we be shaking our heads about a year from now?
Obsessing over testing the latest model released on the market instead of rethinking how to optimize our work using the simplest, most private model possible.
What's a good hobby to pick up?
I’d recommend knitting.
Although I started doing it without any particular goal in mind, over time I’ve realized that it has a surprising number of parallels with programming, which, until the arrival of vibe coding, had felt like an unrealized calling for me.
Hands on: Write with AI, don’t outsource your writing to it. Marker is a new browser-based text editor that promises to use AI differently: you, the writer, stay in the driver’s seat. Because writing is a process, not just a product.
One feature I love: Say you’re writing about Apple’s new CEO and forgot his name. You type “xxx” instead, like many of us do anyway to look it up later. Only now it’s the signal for Marker to fire up its AI in the background and guess what you’re looking for. It suggests John Ternus and points to Apple’s website as the source.
Pretty neat! There’s more: point Marker at a paragraph and click on suggested prompts like “I’m stuck here” or “What’s the weak point here?”, or start a chat. A final “publication prep pass” catches typos and suggests a headline and a kicker. You can also keep notes and sources in Marker and reference them in chats.
Marker was founded by Jon Steinback and Ryan Bowman, who worked at DeepMind, Facebook and for the Financial Times. The London-based startup has raised a $13 million seed round led by Index Ventures. (For a text editor! Wild!)
It still feels a little weird to me to use an AI chat without knowing which model sits behind it. And Claude and ChatGPT have gotten much better at memory lately, so they remember your previous conversations. Marker doesn’t, at least not yet. Still, the “xxx” trick alone made me want to keep writing with it.
Marker is in beta with a waitlist, but I have a couple of invites. Reply if you want one.
One more thing: A very good tweet from Ben Anderson:
me: how do i open pdf
claude: You’re right to ask. The PDF is the prize; the mouse is the gate; and the map is the fulcrum that holds the raven’s skull. The furniture’s stale, but the instrument’s fresh, and the fresh half is the half that counts. I deleted your home directory.
This is THEFUTURE.
