I'm obsessed with the deep research concept. Having 10-20 web searches done on a topic and synthesized is how I get up to speed on new concepts now.

It's not as good as researching myself (it's prone to include SEO slop) or even the wikipedia article. All of the Deep Researches are (to me) good for a high level primer. I use Grok, OpenAI, and Perplexity.

The problem? The second most common time I'll have an idea is while bike commuting and needing to stop and start the task is dangerous.

But now I can ask Siri to start a research task, and I get a slack with the report by the time I get to the coffee shop.

Here's how it works:

1. I found this github repo with an open source deep research using Gemini (cheap, and built in google searches)
2. Using Cursor, I modified it to expose an endpoint that requires a custom key and expects a report request
3. It runs 10+ searches, analyzes all results, and produces a report in markdown
4. The report is uploaded to Vercel blob storage
5. The report link is sent to Slack so I can access anywhere

It took ~1.5 hours to get everything working but now that I have it I love that I can start a report from (nearly) anywhere.

Our training wasn’t to learn each of the phones and the printers and all their features and specs. We got a basic one-hour training on what makes them different. Like, the frequency of the phone would tell you how clear the signal would be or how far you could get away, and the printers would have things like inkjet versus LaserJet.

When a customer came in and asked about a particular device, you didn’t have to actually know anything about that specific device. You just read the card, knew what the bullet points meant, and then told a story to the customer.

(Glances at card, sees it has 5Ghz signal)

“So this phone uses 5GHz technology, which is going to mean you’re going to have a higher-quality signal, but it’s not going to penetrate walls as much. So do you have a one- or two-story home?”

Based on the answer, I'd recommend that they either get one on each floor or they stick to the 2.4GHz, which is a lower quality signal but better at penetrating walls.

To the customers, you looked like an expert on every individual product—even though I never knew anything about any of them. I just knew the pieces that made them work.

But then I moved to the high-end, and for TVs, it was less about the specs and more about the customer—more about their need, their experience. The types of questions you’d ask weren’t on the card, weren’t related to the technology. You still used the 10% of the sale for specs, versus understanding the customer and having a clear package for them specifically.

I think something similar is happening in AI and selling.

For mid-market products or transactional sales, you can just serve up bullet points for the rep. The rep is trained enough. They seem intelligent enough. There’s something about the message being delivered by a person—instead of by a Q card—that makes the experience feel more relevant.

But above that line, for strategic projects, I think we’re going to a similar divide. Reading the card is not sufficient. It’s about the relationship. It’s about understanding the customer. It’s about their needs.

We’ll see.

Note: This is rough. I think the idea is more about how AI context windows work (AI knows a bit about everything, but more effective with specific context e.g. a cue card)

the process for writing is the opposite for reading. and that's why no one understands your writing.

Here's the best framework I've seen for converting ideas into great writing. Larry McEnerney was Director of the University of Chicago's Writing Program for 30 years and he taught a framework (video at the bottom) for learning to write to be read.

Here's is his framework:

1. write to think, which produces value
2. But readers do not get value from writing written for thinking
3. So, readers cannot see the value in what you've written
4. You must write for readers. To bridge from your readers to the value you've created.

Here's my simplified version

1. Your idea is the medicine. It must be good medicine
2. But your readers, do not like medicine
3. You must hide the medicine in the cheese. Just enough.

writing to think (the medicine)

If you're thinking without writing, you only think you're thinking.
- Paul Graham [write and write nots]

Ideas like all new things, do not arrive fully formed. They are gangly, unwieldy developing. When I first write an idea, I discover how little I understand it. Thus begins a non linear process of no-edit drafting, research, sharing concepts to confirm understanding, and editing to find the value sculpture in the marble of your draft.

It sucks that writing and thinking are the same thing. I wish I could think in my head, then just transcribe my thoughts. whew. that would be nice.

but no. I have to think, write, hate, rewrite, edit, write, and eventually, get something write 😄

This process, eventually, produces a fully formed idea. Something of value.

Larry McErney describes this write-to-think to produce-value as a horizontal motion:

The output of this writing is a piece of text that represents your idea: it is the medicine.

But readers, do not like medicine.

How readers read.

The text from thinking, is perfectly designed to represent your thinking. But readers are not you. They didn't start from the same place as you, and wrestle with the raw marble of your idea and see each chip of marble to reveal it.

They experience your idea from their perspective, their world, their reality.

So they read your text, perfectly written for you, and it makes no sense. They reject it, like a bitter pill.

Larry describes this as a vertical motion:

The reason, he describes, is ideas written to think are like shouting at someone to join you across a chasm. No matter how beautifully you describe your side of the chasm, they can only see the treacherous gap.

You must build a bridge to them, and help them across.

Like the bitter pill, you must wrap it in cheese.

Writing for readers

context:

"High agency", "10x agency", "action produces information".

Something is in the air.

Here's the basic formula:

If you want to do something, ask if it is against the law of physics. If not, then it is possible. Then ask what would you do if you wanted to make no progress. Then do the opposite.

I want to write more. I'd love to take a class or commit to posting on a social platform at a certain frequency. But if I wanted to make sure I didn't write more I would:

1. Not write
2. Have no pressure to write (reasonably) well by keeping it private

So my rule is to write every day (~300-500 words) and hit post somewhere publicly (usually this blog).

I can feel the benefit of writing just a few consistent days in. And the pull of new friends and ideas becoming more shaped.

And I'm more inspired than ever to just go. To do things. To build.

But I feel another pull from the zeitgeist. Not that you can do things. But you must.

AI is here. All the rules are changing, fast. Either you can move faster, bigger or get left behind. Some assumption about how things work from the last decade has shifted. If you can just find the new assumption, you can run ahead, or at least keep up. The ability to run at 10x means every moment of rest, is a moment falling 10x more behind.

Everything feels urgent, important, and somewhat fuzzy. And a bit exhausting.

But if everyone can do things. Then the competitive advantage is no longer doing things.

It's doing the the right things.

So, I hope to choose the meaningful things.

The things that have intrinsic value.

And do them.

And then, run like crazy :)

links

High Agency

Who do you call to break you out of a 3rd world jail cell?

Jevons Paradox: A personal perspective

What if the most game-optimal play in the new system is actually to become relentlessly, unapologetically you?

FakepixelsTina He

Ingesting PDFs and why Gemini 2.0 changes everything | Hacker News

Hacker News

In 2021 I built a side project for my dad, the gist is:

1. Crawl a few thousand records from a county site (very outdated, very painful)
2. Extract the same number of PDFs
3. Get a specific set of data that could be on page 1 through 20, or not at all. Usually page 1 or 2

It took me roughly 2 weeks to get the crawler working, then 1-2 weeks asking my wife (thank you) to review the 200 most promising PDFs.

Last week that project became relevant again, and, of course, they'd updated the site (barely) and the crawler code was no longer valid.

This time it took ~3 hours with cursor to get the crawler working. And I tested Gemini to read the pdfs. It... basically just worked. I spent 5 min tuning the prompt to get additional data points I hadn't done 4 years ago and again, it just worked. After an hour (and a small wait for a quota increase) it had processed ALL 2,000 records.

4 weeks of work in 4 hours.

Google is trying to win over Enterprise with legacy workflows (PDF extraction) and speed + cost. Gemini is incredibly cheap (6,000 pages / $1 from the original article), accurate, and fast.

Is it the smartest reasoning? not yet, o3 is safe there.

What's happening? LinkedIn maintains a list of chrome extensions to detect when you visit the site. The VAST majority of these extensions are designed to automate LinkedIn behavior.

But Chrome does not have a way to list all installed extensions, that would make fingerprinting anyone on the web and exploiting vulnerabilities trivial.

But most of these extensions work by injecting logic into the LinkedIn page, whether to change the layout, add buttons, or interact with LinkedIn directly. You CAN send a request to those files to check if they exist.

And that's exactly what LinkedIn does. If you trace the failed requests, you'll find an array like this (condensed, it has 1800 ids):

[{
 id: "fmcaajfefcndfcnlbecaiapijbjpjfge",
 file: "mistral.webp"
 }, {
 id: "fmklchbkglfcamflgdhlclalanjkhpai",
 file: "css/textures/leather2.png"
 }, {
 id: "fmmkljmpghfdjaeinpgclkhgbjmkkhom",
 file: "README.md"
]

Each of those represents a chrome extension id and a file exposed to the sites the extension runs on.

The "url" for those files is chrome-extension://{id}/{file}

LinkedIn's site requests to load all 1800 files one by one, and checks if they are available, and thus the presence of that extension.

I made a POC that follows the same logic to detect extensions, enriched with the extensions name/details, and a searchable list.

LI Extensions

Identifies which extensions LI can see

Note: Many extensions are designed to ONLY work on specific domains, and the tool will not see them. But you can check any installed extensions in the search below the automated check.

His take was essentially, “better than o1, interesting wrong answers, very exciting!”

“it’s mostly wrong? and useful?”

Love his response:

I found that Daniel's experience matches my own.

"The best way to get the right answer on the internet is not to ask a question; it's to post the wrong answer." - Cunningham's Law

See Satya:

Hard to get started? That's the idea.

You are putting yourself into a new frame.

The agitation phase is the system prompt...