weekend ai reads for 2025-09-05

What erodes is the middle. Generic horizontal SaaS is squeezed by AI-driven feature parity and CFO vendor consolidation. The only SaaS that survives here is highly vertical and ROI-proven, with deep integrations, compliance, or outcomes so strong that switching is painful. If your product cannot clear that bar, it should not be sold. It should be kept private as leverage.

As smart folks I know have said, the question is: “Will incumbents get innovation first? Or startups get distribution first?” Incumbents might win.

Despite that drop in cost per token, what’s driving up costs for many AI applications is so-called reasoning. Many new forms of AI re-run queries to double-check their answers, fan out to the web to gather extra intel, even write their own little programs to calculate things, all before returning with an answer that can be as short as a sentence.

Here are approximate amounts of tokens needed for tasks at different levels, based on a variety of sources:

• Basic chatbot Q&A: 50 to 500 tokens
• Short document summary: 200 to 6,000 tokens
• Basic code assistance: 500 to 2,000 tokens
• Writing complex code: 20,000 to 100,000+ tokens
• Legal document analysis: 75,000 to 250,000+ tokens
• Multi-step agent workflow: 100,000 to one million+ tokens

As with any framework like this, you’ll very rarely find a product or team that is purely AI Artist or purely AI Engineer. Just like in human society, there’s value in both approaches. The distinction between these two is a spectrum, and most people won’t fall at the ends but will likely be closer to one side or the other.

Does your use case require the top model on every request, or do you only need to meet a quality bar. If you can meet a bar, routing lets you send most traffic to cheaper models and burst to the frontier when needed. If your users demand the best every time, you need pricing that mirrors usage, or margins will compress as quality expectations rise.

Bringing up that AI is used to make new antibiotics, etc, when someone is talking about the dangers of genAI slop is like bringing up ambulances because someone is complaining that cars are too big and are killing people.

These are so clearly 2 entirely different things, and you know it.

“Too complex” is what they say, whilst looking away.

Over the course of months, my mom became increasingly smitten with her new AI doctor. “DeepSeek is more humane,” my mother told me in May. “Doctors are more like machines.”

We once disagreed on how we interpreted events. Now we can’t even agree on the existence of events.

Contrary to the forecasts of professional recruiters, we find that AI-led interviews increase job offers by 12%, job starts by 18%, and 30-day retention by 17% among all applicants. Applicants accept job offers with a similar likelihood and rate interview, as well as recruiter quality, similarly in a customer experience survey. When offered the choice, 78% of applicants choose the AI recruiter, and we find evidence that applicants with lower test scores are more likely to choose AI. Analyzing interview transcripts reveals that AI-led interviews elicit more hiring-relevant information from applicants compared to human-led interviews.

These incidents were jarring—not just because of the cheating, but because they made me realize how normalized these shortcuts have become. … AI has softened the consequences of procrastination and led many students to avoid doing any work at all. As a result, these programs have destroyed much of what tied us together as students. There is little intensity anymore. Relatively few students seem to feel that the work is urgent or that they need to sharpen their own mind. We are struggling to receive the lessons of discipline that used to come from having to complete complicated work on a tight deadline, because chatbots promise to complete our tasks in seconds.

The trick is that you still keep to a single chat thread with one user and a single assistant. So while the true origin of the assistant messages in the conversation alternates, the models are not aware of each other. Whatever the other model said, they think it was said by them.

So we ask the question: how can data systems evolve to better support agentic workloads? In particular, can data systems natively—and efficiently—support agentic speculation, helping LLM agents determine the best course of action? This question—which, as we argue, our community is well-equipped to answer—holds the key to unlocking unimaginable productivity gains from agents being the primary mechanism we use to interact with data.

One of the most practical changes I made was switching from typing prompts to using voice-to-text.

When we type, we instinctively optimize for the fewest characters possible. When we speak, we’re naturally more narrative. I found that by dictating my requests, I would automatically include more context, share more of my thought process, and explain the “why” behind the task.