michael 📈 👨‍💻
LightNews LightNews
banner
mikedecr.computer
michael 📈 👨‍💻
@mikedecr.computer
220 followers 230 following 340 posts
Quant finance researcher/developer.
Former political scientist, sometimes bike rider.
Long form mikedecr.computer
Posts Replies Media Videos
mikedecr.computer
One relief about LLM tools for code is that programmers hold strong preferences as individuals but heterogeneous preferences as a population, and they are more temperamentally libertarian than other groups. So the marketplace for coding tools is usually pretty rich with high quality options
mikedecr.computer michael 📈 👨‍💻 @mikedecr.computer · 18d
Speaking of keeping the focus in the code, not the chat, I saw this clip of a new nvim pkg "99" that's more my style maybe. A prompt buffer appears only briefly, and you can still work while an LLM fulfills some request asynchronously.

youtu.be/ws9zR-UzwTE?...
This is the only AI I want to use
YouTube video by The PrimeTime
youtu.be
January 31, 2026 at 5:15 PM
mikedecr.computer
I wrote this 6 months ago and the pace at which CLI tools have been rolled out and publicized is making me feel a lot better about this. There's still a bit of friction when you want to pass context from your editor directly to the TUI that could benefit from more protocolization but not as bad
mikedecr.computer michael 📈 👨‍💻 @mikedecr.computer · Jul 5
big frustration I have with LLM tools right now is how we are seemingly unlearning what LSP taught us. There is no reason why I should have to use a specific *IDE* to get certain LLM interaction patterns. There should be standards for LLM interactions, and editors implement the standards.
January 31, 2026 at 4:51 PM
mikedecr.computer
does any other marketing tactic threaten and bully me as much as LLM marketing? If software developers will really be irrelevant (any day now...) then it doesn't matter if I use your coding agent. If using your coding agents will make me more productive then software devs still matter. which is it
January 21, 2026 at 1:28 AM
mikedecr.computer
I don't normally do really oop-y solutions to advent of code problems but it helped clean up the main fns a lot here

github.com/mikedecr/aoc...
Screenshot of python code: @dataclass class CircuitNetwork: _circuits: list[Circuit] # idk if this is the best place for this method but oh well. def contains_link(self, a: Box, b: Box) -> bool: """True if a local circuit contains boxes A and B""" return any(a in circ and b in circ for circ in self._circuits) def attach(self, a: Circuit, b: Circuit) -> None: """Consolidates circuits A and B into a single circuit""" assert a != b, "Circuits must be distinct" assert a in self._circuits, f"Circuit {a=} is not in the network" assert b in self._circuits, f"Circuit {b=} is not in the network" self._circuits.remove(a) self._circuits.remove(b) self._circuits.append(a.union(b)) def get_circuit(self, pt: Box) -> Circuit: """Returns the circuit containing provided Box""" for circ in self._circuits: if pt in circ: return circ
January 19, 2026 at 7:57 PM
mikedecr.computer
I want politicians to care about today's kitchen-table issues like cable management. I didn't create the world where everything is powered by 3ft usb-c cables. Why should I bear the cost of its negative externalities entirely out of my own pocket?
January 16, 2026 at 2:14 PM
mikedecr.computer
permit me this one windmill: I have some obstinate anti-corporate, pro-flesh thing inside me that refuses to call it "AI". They are LLMs. Yes, I use LLM tools, but I don't do their marketing for them
January 16, 2026 at 2:11 PM
Reposted by michael 📈 👨‍💻
hausofholes.bsky.social
im confident that i would thrive in a groundhog day type situation
January 9, 2026 at 10:54 PM
mikedecr.computer
did I overly functionalize Day 5.2 or is this just the thing to do
Python code: @app.command() def day_5_2(test: Flag("--test") = False): data_path: Path = Data(5, test=test) intervals_lst: list[Interval] = [] with open(data_path, "r") as f: for line in f: if line == "\n": break interval: Interval = tuple(eval(chr) for chr in line.strip().split("-")) intervals_lst.append(interval) # with the intervals sorted, we can do this w/ reduce # fn: combine RIGHT w/ tail of LEFT or append RIGHT to LEFT intervals_lst.sort() collapsed_intervals: list[Interval] = reduce(attach_interval, intervals_lst, []) result: int = sum(1 + (right - left) for left, right in collapsed_intervals) print("solution:", result) def attach_interval(lst: list[Interval], new: Interval) -> list[Interval]: """Collapse interval `new` to the last element of `lst` if possible, otherwise append.""" if len(lst) == 0: lst.append(new) return lst end = lst.pop(-1) if (collapsed := maybe_collapse_intervals(end, new)): lst.append(collapsed) else: lst.append(end) lst.append(new) return lst def maybe_collapse_intervals(left: Interval, right: Interval) -> tuple[Interval, ...] | None: """Return collapsed interval if possible, or None""" if left[0] <= right[0] <= right[1] <= left[1]: # left contains right return left elif right[0] <= left[0] <= left[1] <= right[1]: # right contains left return right elif left[0] <= right[0] <= left[1] <= right[1]: # left straddles right lower bound return (left[0], right[1]) elif right[0] <= left[0] <= right[1] <= left[1]: # left straddles right upper bound return (right[0], left[1]) else: return terminal input and output: $ time uv run aoc day-5-2 solution: 357907198933892 uv run aoc day-5-2 0.08s user 0.02s system 94% cpu 0.110 total
January 10, 2026 at 3:16 AM
mikedecr.computer
I repudiate totally the name of the product "Liquid IV". You know what an IV is, right? It's already liquid. The core conceit of Liquid IV meanwhile is that it's a powder. It's Solid IV
January 2, 2026 at 2:16 PM
mikedecr.computer
any time a website is like "made with ❤ by <name>" I am like blegh
December 31, 2025 at 10:17 PM
mikedecr.computer
elderly math profs be like "ok let's say we have a matrix. and now we're going to bully this matrix"
December 20, 2025 at 3:58 AM
mikedecr.computer
One bewildering piece of condescending Apple UI is that, while you *can* attach files to iMessage, you can't do it with the obvious UI element that should do it.

If you're audacious enough to realize that your computer has these things called "files" there's drag-and-drop or the menu bar?
October 28, 2025 at 11:40 PM
mikedecr.computer
really excited for my gf's new iphone to arrive so I can finally rid our lives of lightning cables forever
October 28, 2025 at 11:25 PM
mikedecr.computer
as a blog post @ mikedecr.computer/blog/map-as-...
October 26, 2025 at 7:39 PM
mikedecr.computer
curious to try omarchy but i would need to buy a computer
October 26, 2025 at 2:56 PM
mikedecr.computer
anybody got a good pragmatic blog post or paper on modern techniques for fitting gamma likelihood models? bayesian is ok / preferred actually
October 15, 2025 at 1:29 PM
Reposted by michael 📈 👨‍💻
puddleofbrain.bsky.social
Are you guys street or vert
October 8, 2025 at 8:32 PM
mikedecr.computer
built a great model at work. beautiful really. has everything you think you need or want in it. sucks though. doesn't work at all
September 30, 2025 at 2:21 AM
Reposted by michael 📈 👨‍💻
danabra.mov
i wrote about atproto and why it matters
Open Social — overreacted
The protocol is the API.
overreacted.io
September 26, 2025 at 3:33 PM
mikedecr.computer
weird idea: map in terms of reduce
screenshot of python code: # this is the classical definition of map def map(fn, xs: list): if len(xs) == 0: return [] else: return [fn(xs[0]), *map(fn, xs[1:])] # but you can implement as a reduce from functools import reduce def map(fn, xs: list): return reduce(_as_binary_op(fn), xs, list()) # casting func into a binary operation def _as_binary_op(fn): def binary_op(accum: list, x): accum.append(fn(x)) return accum return binary_op map(str.upper, ["a", "b", "c"]) # ['A', 'B', 'C']
September 27, 2025 at 6:59 PM
mikedecr.computer
Look, I ride a bike and I want better bike infrastructure, but this is magical thinking. If every car disappeared there would be way more bikes, way more bikers who don't ride as fast as you want to go, so way more bike traffic
September 21, 2025 at 7:16 PM
mikedecr.computer
I re-published this post about "lazy" pipeline functions: mikedecr.computer/blog/lazy-dp...

Basically, exploring how to make pipeline-style R code more re-usable and composable by abusing partial function application
Screenshot from the text of the blog post: More abstractly, what I want is more separation of functions from data in the R pipeline style. In some situation, I want to pass data `x` to functions `f`, `g`, and `h`. But instead of writing this: ```{r} #| eval: false x |> f(...) |> g(...) |> h(...) ``` I want to write something like this: ```{r} #| eval: false # first argument omitted from each of these expressions pipeline = compose(f(...), g(...), h(...)) # supply data to the ultimate function call x |> pipeline() ``` `pipeline` is a function that captures unevaluated function calls on `f`, `g`, and `h` with their first arguments omitted. Then I pass data `x` such that the data pass through the first argument of each function in the chain. Why do I want this? Because it should be easier to build a pipeline of expressions that can be re-used on different datasets. I am thinking specically of examples like this: ```{r} #| eval: false # I could evaluate this function as-is x |> pipeline() # and I could also drop in extra steps j and k x |> j() |> pipeline() x |> k() |> pipeline() ```
September 21, 2025 at 4:01 PM
Reposted by michael 📈 👨‍💻
lightblanket.bsky.social
The trick to laundry is never stop. Don’t lose your momentum
September 21, 2025 at 5:05 AM
mikedecr.computer
I can see my hairline doing things that I do not regard as familiar
September 19, 2025 at 12:25 AM
mikedecr.computer
I wrote a "lazy pipe" operator %=>% that lets you build function compositions from partial function expressions in a linear way.

Instead of passing data up front, you define pipeline steps without reference to the data. You get reusable + composable pipeline steps w/out as much function boilerplate
Screenshot of R code that reads: library(dplyr) library(palmerpenguins) # how a normal pipeline looks mass_stats <- penguins |> group_by(species, island) |> summarize( mean_body_mass = mean(body_mass_g, na.rm = TRUE), sd_body_mass = sd(body_mass_g, na.rm = TRUE), ) |> print() # using lazy pipe # create a function by composing partial expressions with %=>% operator compute_mass_stats = group_by(species, island, .add = TRUE) %=>% summarize( mean_body_mass = mean(body_mass_g, na.rm = TRUE), sd_body_mass = sd(body_mass_g, na.rm = TRUE) ) # then pass data compute_mass_stats(penguins) Screenshot of R code that reads: # by using .add = TRUE we can prepend extra groupings to our pipeline # so we can re-use code more efficiency when we need to slice data more ways by_sex = group_by(sex, .add = TRUE) |> as_partial() penguins |> by_sex() |> compute_mass_stats()
September 14, 2025 at 3:05 PM