refactor_master

So this shows that people who appoint 50-agent "code factory studios" with Bob as the CEO, Alice as the CTO, ... is just a waste of tokens, compared to just asking one agent to write it all? It looks like your agents would internally just re-invent office politics, with 10 subagents writing to /tmp/chatter.md why it's unfair that Bob gets to lead.

So it’s like paying $20 for a gym subscription but if you had just stayed there 24/7 paying hourly rates the whole month it would’ve cost $1000.

Except, nobody does that. So that’s not a real number.

The commute only takes 27 minutes if we ignore that everybody else has an LLM- erh, car I mean.

Think you’re clever because your LLM contacts 100 businesses? Imagine 100000 businesses doing the same thing. You’ve just eradicated human support, because nobody’s going to sit through this BS bot future. Good luck when you get auto-flagged for some reason and there’s not a single human to revert or care.

Because nobody was ever fired from choosing Microsoft, and apparently this still holds for AI?

> The researchers successfully tested the robots on phantom heart models that closely mimic a human heart, and on real porcine heart tissue. The robots traveled smoothly and autonomously, even against simulated blood flow.

“In mice” - actually not even.

Ironically on this point, Amazon offers some of the best consumer service in Japan. Returns elsewhere are almost impossible.

> As a parallel, once you learn how to drive you stop walking, as for the primary mode of transportation.

Cars:

- create environments we hate being in

- directly impacts public health

- widens inequality

- makes us productively go faster through the terrible environment we’ve set up

LLMs:

- create markets we hate being in

- directly impacts public health

- widens inequality

- makes us productively put out slop faster through the terrible market we’ve set up

Was this unintended satire?

That’s what we used to call “bullshitting”.

So how do you install 250k km2 of underwater umbrellas, and who stands to get filthy rich from this?

I know what you’re trying to say, but strictly speaking there are many examples of “hard-science” people moving towards and excelling in softer fields, while the opposite is almost never true.

There was a saying where I went to high school: “the language programme students are not better at languages, they’re just worse at math”.

My take is that people “these days” just don’t want to put in the time, money or effort in socializing, especially with new groups of people. Be it dating, meetups or other social gatherings with strangers, it’s about giving as little as you can, in case the event turns out to be a dud. Why risk paying for 2 beers when you could buy one tea for less?

Now, how can I be so sure? Because I’ve been starving at some of these events, whereas others can seemingly call a single slice of pizza dinner.

There’s most definitely a health aspect in there too, but I don’t think it paints the whole story.

In my experience, automated code review is more pointless than ever.

We have all the linters, tests, and AI writing code for us. I don’t need the left hand to tell the right hand it did a good job. I’m very certain my code runs when I push the PR.

What I need now is architectural, long-horizon and business perspective.

Call for a style guide meeting every time you see feedback like this, and write down what everybody agrees on. You’ll never have to do it again after 3-4 of those. Problem solved.

Still not solved? Guess it was really about the commas and not the value delivered anyway, so do whatever you feel like.

Look up the staggering amount of everyday conveniences invented by space exploration. Theoretically the private market would’ve filled this gap, but in practice we know the market will never be willing to spend the first few trillions on foundational research.

AI still makes mistakes on code with a trillion billion examples, but let’s invent a DSL that only AI can read and hope for the best?

Why not spend a few tokens or so on making a frontier LLM build the classifier from scratch? Then you also know exactly what the classifier is capable of, and whether it’s easily learnable/informative data.

Why would you be confident in feeding garbage to a “cheap and fast” classifier with unknown domain-specific performance?

I know we kind assume omniscience for frontier models, but at this point the evidence is kind of out there.

Not really. China is essentially a microcosm of capitalism playing out entirely internally.

E.g. to move to the capital permanently from outside you're basically under the same scrutiny as an immigrant on a greencard.

Are you insinuating those 1970s fonts takeout fonts aren’t actually how Chinese people communicate? I knew they were AI!

But what if your PR was created as half chat, then an agent, then some more chat planning, possibly in a new chat to clear the context? I think my teammates would find my inner monologue rather confusing at times.

The sequence of commits I present a purely optimized for reviewing, and not an actual record of what happened. I’m not going to read another person’s 100 turn slop factory, because they couldn’t express their change in one paragraph.

Too bad there’s such a lingering culture of Windows and creaky Lenovos and Dells around it. I’m sure it can be different, but I’m not taking my chances ever again. Simply not worth my sanity.

> You don't blame a knifemaker for somebody getting cut by a sharp knife. AI is a knife. Very handy, very dangerous. We have to use them safely, that's all there is to it.

If I grossly neglected to maintain live deadly bacteria in my containment facility, am I absolved of blame? Since, you know, the bacteria is the real bad guy who should be put in jail?

Until everybody’s AI talks to everybody’s AI and tokens will be churned with no utility, and all the valuable business meetings will happen face to face behind closed doors.

> The letters you already know, telling you what they sound like

So… letters?

If you think a cryptographically signed image can’t be JPEGed 5 times and shared on Facebook for misinformation then here’s a relevant XKCD for you:

https://imgs.xkcd.com/comics/security.png

No no the invisible hand of the market will fix it, you see. Car manufactures will become so big and pay so many taxes that we could pave every road with rail.

It’s just the lazy/expression part of the API, which is really the bread and butter of polars, rather than just being “replacement syntax” for pandas. This allows you to tap into abstraction that SQL can’t keep up with:

  import polars as pl

  # 1. Base Dataset
  lazy_df = pl.LazyFrame(
    {
      "store_id": ["S01", "S02", "S03", "S04", "S05"],
      "revenue": [5000.0, 2400.0, 15000.0, 900.0, 3200.0],
      "margin": [0.45, 0.30, 0.60, 0.15, 0.50],
      "tx_count": [120, 45, 300, 20, 85],
      "returns": [5, 12, 45, 2, 8],
    }
  )

  # 2. Define Layer Abstractions
  def get_kpi_layer() -> list[pl.Expr]:
    return [
      (pl.col("returns") / pl.col("tx_count")).alias("return_rate"),
      (pl.col("revenue") / pl.col("tx_count")).alias("avg_order_value"),
    ]

  def get_threshold_layer(thresholds: dict[str, list[float]]) -> list[pl.Expr]:
    return [
      (pl.col(col) > limit).alias(f"is_{col}above{int(limit)}")
      for col, limits in thresholds.items()
      for limit in limits
    ]

  def get_interaction_layer(numeric_cols: list[str]) -> list[pl.Expr]:
    return [
      (pl.col(a) / (pl.col(b) + 1e-5)).alias(f"ratio_{a}per{b}")
      for i, a in enumerate(numeric_cols)
      for b in numeric_cols[i + 1 :]
    ]

  def get_segmentation_layer() -> list[pl.Expr]:
    return [
      pl.when(pl.col("margin") > 0.4)
      .then(pl.literal("High"))
      .otherwise(pl.literal("Low"))
      .alias("margin_profile")
    ]

  # 3. Consolidate and Execute Single Graph Pass
  thresholds = {"revenue": [1000.0, 5000.0, 10000.0], "tx_count": [50, 100, 200]}
  numeric_cols = ["revenue", "margin", "tx_count", "returns"]

  expr_pool = [
    *get_kpi_layer(),
    *get_threshold_layer(thresholds),
    *get_interaction_layer(numeric_cols),
    *get_segmentation_layer(),
  ]

  final_df = lazy_df.with_columns(expr_pool).collect()

The cool thing about polars is that you can conditionally collect expressions over many layers of business logic, and then compute the result at the end. Doing this in SQL ends up in a hodgepodge of strings and trimmed ends to please the syntax. You can also pretty effortlessly write quite complex conditionals directly in polars, and bridge it easily to the surrounding python.

I find that SQL is only easier to read with minimal abstraction, but as soon as the project gets bigger SQL becomes an unwieldy island of different that has served its purpose after we’re done with reading/writing the data.

I don't know if my code is just "complex", but I find that Luna on max ignores the surrounding style and completely ignores logical consequences of a change, like just writing `del arg1, del arg2, ...` instead of dropping it from the surrounding code. All LLMs make questionable decisions at times, but Luna requires so much guidance that it's faster to just type it out yourself. What kind of routine tasks can one accomplish with such a model?

The US does not lead in broad markers of "social good" like life expectancy, happiness index or social mobility, so what exactly is this KPI useful for?

I've discovered so many great artists through Spotify, and been to many of their shows (which Spotify also helps surface with their "Live Events" feature). I regularly get artists with less than 100k listeners in my Discover Weekly, which I then look up, and find that they did a concert on KEXP, and the comments all proclaiming "I knew good music wasn't dead!".

So I can say for sure that I don't share the slop sentiment. Sure, there's some generic filler every now and then, but any non-cohesive mix, as opposed to a carefully curated album is going to have songs that naturally become "fillers".

Just now, I discovered that Spotify snuck in "The Juan Maclean - Running Back To You (2014)", which is a cover of "Brian Bennet - Solstice (1978)". Neither songs are completely unknown, but they're also not exactly Bruno Mars.

I don't know if this is pure luck or algorithmic rhythm analysis, but for me it's working.