A recent paper studied text summarization improvements using a chain of density prompt. The prompt improves over vanilla GPT responses and is close to human summarizations in informativeness and readibility.

The prompt instructs the LLM to do several passes at summarizing the source content. After an initial summary, it is instructed to find missing entities and include them when re-writing the summary using the same number of words.

See also:

• Context is needed for practical LLM applications
• Intent-based outcome specification
• This could improve summarization of AI for notes

In Talking Watches with John Mayer and Ed Sheeran, the two musical artists talk about their careers and the stories behind their collection of watches. John neatly summarized the romanticism about watches when he said, “I like watches that remind me of who I’ve never been.”

I would add that they also remind me of who I want to be in that moment, in that day. When you have a time fortress strapped to your wrist, you feel more capable even though you would never dive 500 meters (or in my case, dive at all). When you throw on a timeless watch to go out to dinner you feel classic and like the time doesn’t matter so much.

The door-in-the-face technique is a negotiating tactic is when you ask for something rediculous so that your next ask is more likely to be agreed to. This has been proven to be effective in a psychology study in 1975 and then replicated in 2021.

See also:

• Make two offers is another way of increasing acceptance
• Being aware of this technique allows you to avoid stepping into someone else’s frame
• Keep in mind that people respond in kind and the technique might have unexpected results

It’s striking how captivated the world is by LLMs and the latest techniques in AI. What if we were just as excited about progressing human knowledge and learning? Education doesn’t receive nearly the attention and hype it should yet we are far more capable.

On the other hand, cybernetics is an area that could greatly accelerate knowledge creation. We just shouldn’t confuse the technology with knowledge itself.

See also:

• Human intelligence is a special case of artificial intelligence
• The Beginning of Infinity

I’m upgrading from the iPhone 12 Pro to the iPhone 15 Pro. Here’s what I’ve noticed as the biggest differences so far.

Despite all the talk about the titanium case, I barely notice the difference in the hand. It’s slightly lighter but now more top heavy—presumably because of the large size increase in the protruding camera sensors.

The high-refresh rate screen is a big improvement. Scrolling feels like moving around high-gloss magazine paper. It’s smooth and sharp without the blur you see on when scrolling quickly.

The always-on display was surprising at first. It seems pretty handy for how I use the phone by putting it on my desk when I’m working.

The new camera options are pretty great. I briefly tried out taking pictures in RAW and the ProRes codecs. I read about the new Apple Log video recording which makes it easier to adjust video by starting with flatter colors to start (that’s my understanding at least). With a little bit of effort, photos and videos can feel more professional to my eye at least.

See also:

• One of the best parts of getting a new phone is using the old iPhone to make the best webcam

You should set up channel partnerships when it helps your business access new customers while filling an existing need for the partner and not competing with your direct business.

The first part “accessing customers” means that setting up the partnership should benefit you by cheaply acquiring customers the partner already has. It doesn’t work if the partner doesn’t have an ongoing relationship with these customers and it doesn’t work if you can’t make money from these customers.

The next part about filling an “existing need” is that the partner likely has many other partnership opportunities without you so by solving a problem they have already recognized will help prioritize partnering (keep in mind the build, buy, partner framework).

Finally, while partnerships can accelerate growth, long-term you need to be sure it doesn’t compete with your direct business. It would be a bad trade to cannabilize your business by offering the same thing to the same target market through a partner and make less money.

See also:

• Keep in mind a direct business is more valuable than a channel business but commoditization increases the importance of distribution
• Channel pricing models

There is not a finite amount of work that can be distributed throughout an economy. The amount of jobs are not tied to a zero-sum, “lump of labor”. Worries about immigration and automation taking away jobs and leading to unemployment are unfounded. Since the number of jobs are not fixed, changes in the workforce and technology lead to new jobs and or simply different jobs over time.

See also:

• Malthusian catastrophe is similar in that it argued food supply was fixed for any give population size
• Predictions about the future don’t account for new knowledge
• People are bad at long-term thinking
• Automation reduces marginal cost of nonautomated tasks

Hyperbole is an emacs package that matches text and turns it into links (buttons in hyperbole parlance). Links can execute arbitrary code so it’s kind of like a universal emacs subsystem for linking things together (i.e. hypertext).

In some ways you might describe hyperbole as “M-RET does what you expect at point”.

From this generic abstraction you can do a number of things. You can unify the behavior of clicking on file paths, regardless of what mode or programming language you are writing in. You can link together org-mode documents to python code in another directory. You can implement “do what I mean” when you press M-RET anywhere since you can add your own buttons that execute arbitrary code.

I’m not sure how practical hyperbole is yet. It’s hard to discover what is a button (at least in my setup) and I don’t know when I would want to add my own buttons but now that I know this exists, I’m sure I’ll start to see opportunities crop up.

See also:

• Hyperlinks as a foundation for distributed knowledge fails because it lacks transclusion
• Org-roam brings backlinks as a form of hyperlinks between org-mode documents

Effective writing that gets a response is written for busy people. The message should be constrained to a single idea. The message should be brief, reducing as many words as possible.

The longer a message is the more time people think the task will take. In the Book “Writing For Busy Readers”, the authors tested a request for a survey. Resonse rate doubled when the number of words was cut down from 127 to 49. In another example, donations increased 16% when deleting every other paragraph in a fundraising email.

Read In favour of simple writing.

See also:

• Always give the summarizing idea first prevents the reader for having to dig for the point, so does a BLUF
• The Pyramid Principle has many techniques for business writing
• Business writing should always answer a key question (similar to “a single idea” raised in this article)

A meta-study of research of 1MM people, found differences in personalities and cognitive abilities. Neuroticism was negatively related to most cognitive abilities like reasoning and knowledge. Extraversion was positively related to processing ability and retrieval fluency. Industrousness and compassion was positively related to knowledge found different correlations for neruorticism, extraversion, conscientiousness.

See also:

• Imitating high-status people doesn’t work
• You can’t be normal and expect abnormal returns

A new study on the climate impact of teleworking found that switching from working onsite to working from home full time can reduce up to 58% of work’s carbon footprint. Hybrid schedules have some improvement (11% to 29%). Working one day from home has no benefit.

Read Working from home now has another powerful benefit.

See also:

• Remote work is not necessarily better for the environment
• 20% of US workers are expected to work from home permanently

A pattern I really like in Notion is that you can a @ (at) mention any page with search as you type and autocomplete. I’d like to do something similar in org-mode so that I can quickly link related headlines.

Using the fantastic org-ql library for Emacs, I integrated it into completions triggered by letters following the @ symbol. This translates the input after the @ symbol into a query and inserts a org-id link when selecting a match.

(require 'org-ql)

(defun my/org-agenda-completions-at-point ()
  "Function to be used as `completion-at-point' in Org mode."
  (when (looking-back "@\\(\\(?:\\sw\\|\\s_\\|\\s-\\|\\s-:\\)+\\)")
    (defvar heading-to-id (make-hash-table :test 'equal))
    (let* ((start (match-beginning 1))
           (end (point))
           (input (match-string-no-properties 1))
           (candidates (org-ql-select (org-agenda-files) (org-ql--query-string-to-sexp input)
                         :action (lambda ()
                                   (let* ((heading (org-get-heading t))
                                          (id (org-id-get (point))))
                                     ;; Avoid having to look
                                     ;; up the ID again since
                                     ;; we are visiting all
                                     ;; the locations with
                                     ;; org-ql anyway
                                     (puthash heading id heading-to-id)
                                     heading))))
           (exit-function (lambda (heading status)
                            (when (eq status 'finished)
                              ;; The +1 removes the @ symbol
                              (delete-char (- (+ (length heading) 1)))
                              (insert
                               (format "[[id:%s][%s]]" (gethash heading heading-to-id) heading))))))
      (list start end candidates :exit-function exit-function))))

(defun my/org-agenda-completion-hook ()
  "Configure org-mode for completion at point for org-agenda headlines."
  (add-to-list 'completion-at-point-functions 'my/org-agenda-completions-at-point))

(add-hook 'org-mode-hook 'my/org-agenda-completion-hook)

See also:

• This personal infrastructure helps speed up my workflow when working on tasks
• Emacs is the ultimate editor building material
• The downside of using org-id links

A study of 73,000 NFT collections found that more than 95% had a market cap of zero ether and an estimated 23 million people are holding worthless assets.

Read Remember when NFTs sold for millions of dollars? 95% of the digital collectibles may now be worthless.

See also:

• Trading volume of NFTs fell 97%
• Investment strategies that only work in a bull market
• Growth in tech during the pandemic was not permanent

This note does not have a description yet.

The nice part about ChromaDB is that you can read the tables in SQLite using the new Emacs 29 sqlite-mode. That means there is no other configuration for accessing the database, you can open the sqlite file directly.

I was curious to see what metadata was extracted from an automatic document loader in LangChain. I opened the sqlite file and browsed the metadata directly to see what was available. Nice!

See also:

• Why vector databases are important to AI
• Emacs is the ultimate editor building material

You can turn off sending telemetry data to ChromaDB (now a venture backed startup) when using langchain.

from langchain.vectorstores import Chroma

client_settings = chromadb.config.Settings(
    is_persistent=True,
    persist_directory="mydir",
    anonymized_telemetry=False,
)
return Chroma(
    client_settings=client_settings,
    embedding_function=my_embeddings,
)

ChromaDB is a vector database used for similarity searches on embeddings. It’s extremely easy to use if you are using Python and works well with LangChain.

See also:

• Turn off Chroma telemetry in langchain
• Emacs and ChromaDB

As software grows, patterns and practices naturally evolve. Inevitably, someone comes to the conclusion that a change should happen and code should be migrated to the new thing.

While many agree full rewrites are a bad idea, migrations tend to slip under the radar.

Migrations are mini rewrites and the same caution should be applied.

When you think of migrations as rewrites, you think about the problem differently. Rewrites are an exercise in chunking. The total cost of ownership is not just adding it the first time, but updating all other places for consitency, fixing bugs introduced along the way, and maintaining the new thing forever. Unfortunately, many large scale codebases end up with tech debt because of half completed migrations (probably several because a migration are needed to fix the other partial migration!).

You’ll often hear justifications like, “We can do this gradually” and, “We’ll add this to new code first.” All lies we tell ourselves to do the thing we want.

What can you do about it?

Approach the migration as if it needs to happen all at once in one effort or not at all. This pierces the illusion that future us will have the time and focus to complete the work present us doesn’t do (experience would suggest this is never true unless you absolutely have to).

Don’t accept installment plans, pay the whole thing up front or move on because worse is better.

See also:

• Dependencies are a sneaky way migrations are needed and the default strategy should be to not add more
• Closing the gap between previous technical decisions and the current quality bar is engineering leadership

Marl is the marginal user. The one who represents the broad base of users that drive most product decisions to the lowest common denominator. Attention is short. Intolerance for change is high. When you ask why they dumbed down your favorite consumer app into a sugar coated dopamine factory, it’s probably because of Marl.

We’ve all been Marl at one point in time or another.

The thing people forget is that we are all Marl. The things we think couldn’t possibly apply to us, what we think is designed for a mindless scrolling drone, is us too. We are not always Marl and some might be more Marl than others, but we truly are that way some of the time.

Read The Tyranny of the Marginal User.

According to the Lightspeed sales benchmark report for 2023, win rates decreased for 42% of companies surveyed. The larges decreases happened for initial sale prices of $250k+.

The leading reasons are:

• Lost project funding 58%
• Decided to do nothing 44%
• Went silent 41%
• Decision maker left 25%

See also:

• Running a startup during a recession
• Burn multiples are a way of managing growth efficiently

When it comes to selling an asset, it’s helpful to remember that it only takes two people to make an auction. Auctions result in higher sales prices as it’s not just the value of the thing being purchased, but the dynamics of winning. In the sale of a company for example, it’s not just about buying the company but preventing the other party from getting it.

What’s the lesson? Creating an auction for an asset is clearly better and all you need is two interested parties which is a low enough number that it is always worth the effort.

See also:

• Auctions induce feeling of loss aversion
• Mentioned by Paul Graham on the Tyler Cowen podcast when talking about the art market

In sales, it’s easy to mislead yourself into thinking a sales deal is on track when it’s not. A simple way to know that it’s not is to ask yourself if there is a clear next step that both parties have agreed to. Without it, you won’t close the deal and if you do, it’s probably a nothing sale.

Agreeing to a next step can be as simple as deciding whether to move forward with your solution by the end of next week. There is a clear action and deadline that makes it easy to agree. It also keeps the door open for the lead to ask for more materials or proactively raise that there is someone else you need to meet with.

(I didn’t come up with this but I can’t find the source of where I found it).

See also:

• Founder-led sales
• Sending follow-up emails is closely related—it’s much easier to get a response when you are merely reminding them of what you already agreed to do

A common data consistency issue is writing empty strings to the database. An empty string "" is a perfectly valid text or varchar but is usually not what you want.

Thankfully, Postgres has a way of checking for empty strings that prevent storing bad data. A domain is a reusable way to check a value that looks like this:

create domain content as text check (value != '');

This can be used as the column type when creating a table:

create table posts (
    id         serial       primary key,
    title      content      not null,
);

Now you’ll get a friendly error if you try to save an empty string. If you are using SQLAlchemy, you can specify a database model using a domain too.

Read Check data easily with PostgreSQL domains by Benjamin Sago.

See also:

• Postgres is the most exciting boring technology because of features like this
• Most business problems are data modeling problems

A DRI is a commonly used organizational pattern where one person is directly responsible for the overall outcome of an initiative. This avoids many pitfalls like a tragedy of the commons where responsibility is shared amongst everyone and therefore everyone loses all accountability.

It’s challenging but rewarding. This model relies on individual contributors that have the skills to successfully lead, project manage, communicate, and be effective outside of their own domain. For example, a backend engineer might be the DRI for a product feature that includes UI design and frontend work even though they themselves are not a designer or a frontend engineer. The key is being able to extract the best result from others and work through them.

In my career, I’ve worked with this “DRI model” for many years with mixed results. As a manager, I’ve had to intervene in projects going off the rails (delegation not abdication). The most common complaint I’ve heard about being a DRI is that larger initiatives require project management—widely undervalued by engineers—and needs a lot of coaching and help.