Mike Samuel 🟣

@mvsamuel

Programming languages person focused on software systems problems. Previously, first frontend engineer on Google Calendar, and was a security engineer who worked on the industrial-strength Mad Libs undergirding Gmail. Pro-trans-rights is pro-family.

916 Following860 Followers2,624 Posts

YouTube Bluesky

Is this your profile?Claim it and mvsamuel.dirty.social is yours.Claim profile

Mike Samuel 🟣@mvsamuel· 27d

Somehow models work better for English than other human languages even when most of the training corpus is non-English text.

[From https://dl.acm.org/doi/abs/10.1145/3589334.3645643]

safety remain paramount in these high-stake domains. Moreover, the development of LLMs is disproportionately focused on English. It remains unclear how these LLMs perform in the context of non-English languages, a gap that is critical for ensuring equity in the real-world use of these systems. This paper provides a framework to investigate the effectiveness of LLMs as multi-lingual dialogue systems for healthcare queries. Our empirically-derived framework XlingEval focuses on three fundamental criteria for evaluating LLM responses to naturalistic human-authored health-related questions: correctness, consistency, and verifiability. Through extensive experiments on four major global languages, including English, Spanish, Chinese, and Hindi, spanning three expert-annotated large health Q&A datasets, and through an amalgamation of algorithmic and human-evaluation strategies, we found a pronounced disparity in LLM responses across these languages, indicating a need for enhanced cross-lingual capabilities. We further propose XLingHealth, a cross-lingual

Mike Samuel 🟣@mvsamuel· Jul 13

Sonofa. For all the reasons, TMNT should've made Splinter a wise old raccoon.

In Sanskrit, pandit generally refers to any "wise, educated or learned man" with specialized knowledge.[11] The term is derived from paṇḍ (पण्ड्) which means "to collect, heap, pile up", and this root is used in the sense of knowledge.[12] The term is found in Vedic and post-Vedic texts, but without any sociological context.

Mike Samuel 🟣@mvsamuel· Jul 10

Phew www.maine.gov/sos/news/sec...

7/10/2026

AUGUSTA, MAINE – The Secretary of State’s Office today confirmed that a formal notice has been received from U.S. Senate candidate Graham Platner. A public declaration is not an official withdrawal, and a candidate must formally withdraw in writing to the Elections office, including signature.

Because the candidate officially withdrew before 5 p.m. of the 2nd Monday in July (July 13, 2026), his name will not appear on the ballot, and his political party may replace him. The deadline for the party to name a replacement is the 4th Monday in July (July 27, 2026). Maine Statute does not address how a replacement candidate may be chosen by a party, only that the candidate filling the vacancy must be a “qualified person.” Announcements about how a replacement candidate will be chosen or when that candidate will be announced will come from the political party.

Mike Samuel 🟣@mvsamuel· Jul 9

mitpress.mit.edu/978026263269...

In Unlocking the Clubhouse, social scientist Jane Margolis and computer scientist and educator Allan Fisher examine the many influences contributing to the gender gap in computing. The book is based on interviews with more than 100 computer science students of both sexes from Carnegie Mellon University, a major center of computer science research, over a period of four years, as well as classroom observations and conversations with hundreds of college and high school faculty. The interviews capture the dynamic details of the female computing experience, from the family computer kept in a brother's bedroom to women's feelings of alienation in college computing classes. The authors investigate the familial, educational, and institutional origins of the computing gender gap. They also describe educational reforms that have made a dramatic difference at Carnegie Mellon—where the percentage of women entering the School of Computer Science rose from 7% in 1995 to 42% in 2000—and at high schools around the country.

Mike Samuel 🟣@mvsamuel· Jul 7

I swear I'm not always this passive agressive.

Is it common to delete all comments when refactoring code?

[Chat-style box with text above over "18:33" followed by replay edit, copy controls.]

Mike Samuel 🟣@mvsamuel· Jul 7

The Leon Moisseiff of candidate vetting.

Leon Solomon Moisseiff (November 10, 1872 – September 3, 1943)[1] was a leading suspension bridge engineer in the United States in the 1920s and 1930s. He was awarded The Franklin Institute's Louis E. Levy Medal in 1933.[2]

His developments of the theory of suspension bridges are eclipsed by the dramatic failure of the Tacoma Narrows Suspension Bridge, his design, four months after its completion in 1940. The event was filmed and is shown to engineering students as a reminder of the possibility of failure and the importance of avoiding hubris.

Mike Samuel 🟣@mvsamuel· Jul 3

Glad @bradlander.bsky.social is not soft pedaling.

Your efforts to define them, and me, outside the Jewish community, are dangerously short-sighted. Jews are not made safer by proscribing a particular vision of Israel as the price of full belonging, or by insisting on unconditional support for Israel while it commits human rights violations against Palestinians.

The future you imagine presumes that the greatest danger facing Jewish life is that Jews will leave Zionism. But it seems to me the real danger is that young Jews will conclude there is no room for them inside Jewish institutions unless they silence their conscience. A community cannot thrive if the choices it offers the next generation are hypocrisy or excommunication.

Mike Samuel 🟣@mvsamuel· Jul 1

The shielding weight / reliability tradeoffs have been known about for a long time. Here's a table detailing expected effects of radiation on processor's available at the time from a 1998 Robotics Institute thesis I was a research assistant for. publications.ri.cmu.edu/storage/publ...

Lunar Rover Design

and quantity selected is a function of total computing need calculated from software requirements.

Table 4-11 : Computer processors overview
Radiation Effects
Prediction watts int/tp Processor
M 68060 CPU 60145 2 68020 rad qual, tough
Pentiurn Pro CPU 292/248 23 CISC. extremely hard
R3000 CPU 28/36 4 inherently rad tolerant, flown
R 10000 CPU 428n60 6 possibly rad tolerant, %1M
PowerPC 750 CPU 496/336 5.7 possibly rad tolerant, $.5M
TI C50 DSP 4W40 3 available in 100 kRad
Sharc DSP 40140 4 rad-qual available
TI CEO. MVP DSP 2200 11 $lM lo rad qual
TI C60 DSP 1600 11 $1M lo rad qual

Mike Samuel 🟣@mvsamuel· Jul 1

I want a progressive movement that is aggressively pro raising wages at the bottom and aggressively pro immigration. These are completely coherent goals. They don’t always have the same supporters, but they have the same enemies. It’s time to take the fight to them.

To start with, we should firmly reject the framing that there is some ‘right’ level of immigration. “How many is too many?” is a favorite challenge of the nativists. It’s the wrong question. The measure of a country's success is if its citizens are happy, healthy, and prosperous, not if an arbitrary number of people cross a border.

Mike Samuel 🟣@mvsamuel· Jun 25

Research, but busted. Also, some of those error bars are a public hazard.

Photograph captioned "An early photograph of Stonehenge taken July 1877"

It's monochrome, sepia toned.

In the foreground, there's ground cover interrupted by a worn footpath from left to right.

Behind the footpath are the classic stonehenge stones, but fewer standing than in modern photographs and several of the standing ones are tilted.

Mike Samuel 🟣@mvsamuel· Jun 23

tfw a copy/paste error causes all your test conditions to be identical to TRUE.

Mike Samuel 🟣@mvsamuel· Jun 23

8 years is nothing. This was reported of software tools for hiring 38 years ago. pubmed.ncbi.nlm.nih.gov/3128356/

The story begain in December 1986 when the commission was informed by Dr A Burke and Dr J Collier, both senior lecturers at St George's, that a computer program used in the initial screening of applicants for places at the school unfairly discriminated against women and people with non-European sounding names. The program had been developed by Dr Franglen, a member of staff, to reduce the work of selecting candidates for interview. It was also hoped that it would eliminate any inconsistencies in the way the admission staff carried out their duties. The program was written after careful analysis of the way in which the staff were making these choices and was modified until by 1979 it was giving a 90-95% correlation with the gradings of the selection panel. This point is important: the program was not introducing new bias but merely reflecting that already in the system. By 1982 all the initial selection was being done by computer. Details of each candidate were obtained from his or her University Central Council for Admission (UCCA) form, but since this contains no reference to race this was deduced from the surname and place of birth. The computer used this information to generate a score which was used to decide which applicants should be interviewed. Women and those from racial minorities had a reduced chance of being interviewed independent of academic considerations.

Mike Samuel 🟣@mvsamuel· Jun 22

That claim is hard to confirm. What is is whether Jewish Americans align with AIPAC. 70% of Jewish Americans are Democrats. The vast majority treat AIPAC endorsements in Dem primaries as neutral (47%) or negative (40%). AIPAC does not speak for or to Jewish Americans. jstreet.org/wp-content/u...

When Jewish Democrats are asked whether they are more or less likely to support a candidate who AIPAC endorses in a Democratic primary, most people say it makes no difference (47 percent) or they are less likely (40 percent) -- and there is a remarkable difference by age.
- 66 percent of Jewish Democrats under 35 say they are less likely to vote for a Democratic primary candidate endorsed by AIPAC.
- 46 percent between ages 35-49 say less likely and 41 percent say no difference.
- Most Jewish Democrats aged 50-64 (60 percent) and over 64 years (64 percent) say no difference

Mike Samuel 🟣@mvsamuel· Jun 19

Cute. Not sure what to make of this. Mistral seems to think I do formal verification; important but not my lane. Deepseek seems to think I worked on V8, but I think all I contributed was answering Robert Griesemer's monotonically exasperated "do JS programmers actually do this?" questions.

[All caps old timey pixelated font showing text (below) above a series of bars for various code models]

Mike Samuel
Google software security engineer

>
590 strength · Top 7%
<
Claude Opus 4.8 says
<
3/10
>
Software engineer known for work on web security, code sanitization, and contributions to projects like Caja, Closure Templates, and JavaScript tooling at Google.

[bars]
GPT-5.5 about 80% full
GPT-5.5 mini about 95% full
Claude opus 4.8 about 80% full
Claude Haiku 4.5 about 75% full
Grok 4.20 about 75% full
Gemini 3.1 Lite about 95% full
Kiwi K2 0905 about 25% full
Deepseek v4 about 95% full
Llama 3.3 70B and 3.2 1B empty
GLM 4.7 flash about 95% full
Mistral 3.2 24B about 80% full
Owen3 8B empty

Mike Samuel 🟣@mvsamuel· Jun 18

Yeah. ES262 introduces "realms" (tc39.es/ecma262/#sec...) to help nail down the semantics of separate object graphs that don't stay separate, and language builtins like `Array.isArray` work around the problem that `instanceof Array` doesn't work in the face of distinct `Array.prototype`s.

9.3 Realms
Before it is evaluated, all ECMAScript code must be associated with a realm. Conceptually, a realm consists of a set of intrinsic objects, an ECMAScript global environment, all of the ECMAScript code that is loaded within the scope of that global environment, and other associated state and resources.

Mike Samuel 🟣@mvsamuel· Jun 18

Mozilla implemented distributed cycle collection and other browsers have adopted the same technique. dl.acm.org/doi/abs/10.1...

Compartmental memory management in a modern web browser
Authors: Gregor Wagner, Andreas Gal, Christian Wimmer, Brendan Eich, Michael FranzAuthors Info & Claims
ACM SIGPLAN Notices, Volume 46, Issue 11
Pages 119 - 128
https://doi.org/10.1145/2076022.1993496
Published: 04 June 2011

Abstract
Since their inception, the usage pattern of web browsers has changed substantially. Rather than sequentially navigating static web sites, modern web browsers often manage a large number of simultaneous tabs displaying dynamic web content, each of which might be running a substantial amount of client-side JavaScript code. This environment introduced a new degree of parallelism that was not fully embraced by the underlying JavaScript virtual machine architecture. We propose a novel abstraction for multiple disjoint JavaScript heaps, which we call compartments. We use the notion of document origin to cluster objects into separate compartments.
Objects within a compartment can reference each other directly. Objects across compartments can only reference each other through wrappers. Our approach reduces garbage collection pause times by permitting collection of sub-heaps (compartments), and we can use cross-compartment wrappers to enforce cross origin object access policy.

Mike Samuel 🟣@mvsamuel· Jun 18

There are corner cases with some weak map implementations. en.wikipedia.org/wiki/Ephemeron And bugs in std. Iirc, an early version of Vector in Java (or Oak) forgot to null out backing array elements on removal. As long as the vector was reachable, the deleted elements were unreclaimable.

[highlight on third to last sentence that starts with "But if the value, possibly transitively]

Description
In computer science, finalization occurs when a garbage collector (GC) informs an application that an object is "almost collectable". It is used to help an application maintain its invariants. Weak references may be used by a garbage collector to determine the objects that are almost collectable. Seen both as key-value pairs, the main difference between weak references and an ephemerons is the way the garbage collector treats them. For weak references, the garbage collector always follows the value in the key-value pair. For ephemerons, instead, the garbage collector doesn't follow the value but queues the ephemeron for further observation at a second stage: after the first tracing phase is done, it runs through the queue looking at each ephemeron and if its key was seen, then it follows its value. This subtle difference impacts in graphs with some kinds of cycles, where weak pairs do not describe correctly that an object ought to be "almost collectable". For example, consider a key-value pair with weak references where the key is an object and the value is a set of properties attached to the object. It is expected that when the object is ready to be collected, the properties will also go away. But if the value, possibly transitively, maps to its own key (the object), then the object will never be collected. If an ephemeron was used instead, the value wouldn't have been followed unless the object was proved alive, solving the cycle. Ephemerons are similar to weak pairs, but an object in an ephemeron's key field may be classed as "almost collectable" even if it is reachable from the ephemeron's value fields.[1]

Mike Samuel 🟣@mvsamuel· Jun 16

Hard to overstate how naïve this is. The worst content comes from bots which happen to be anon but don't have to be. State level actors can forge IDs for their bots. Other bot network operators can rent real identities from offline users in low-wage countries. Bots are gonna bot regardless.

5. Require users to use real names. Much of the worst content is produced by anons; it

Mike Samuel 🟣@mvsamuel· Jun 16

Maybe I read too much into the linked article. Chad doesn't explicitly say that disciplines *are* situational coping strategies that require consistent, directed dev effort. But his example of test suites and runtime assertions for dealing with dynlang-specific problems seems a point of agreement.

The discipline didn't vanish. It moved into tests, contracts, and feedback loops that reflected how the system actually ran. The type system was still there; you just had to earn it through behavior rather than declaration.

Mike Samuel 🟣@mvsamuel· Jun 16

[Black hand-written text on a grey background]

I can't draw but I wish you lots of cake!

Mike Samuel 🟣@mvsamuel· Jun 16

Nice. I see III.B and V.C lay out the argument that there are sufficient alternatives for organized crime.

Far from illustrating the need for conspiracy criminalization, O-Block, Bennett, and Banks provide additional grounds for doubts about conspiracy’s purported benefits. There’s no reason to believe that the prosecution of Weekly’s killers would have failed without the crime of conspiracy. Charles Liggins, Kenneth Roberson, Tacarlos Offerd, Christropher Thomas, and Marcus Smart were alleged to have worked together to intentionally kill Carlton Weekly. They were all convicted of murder along with conspiracy to commit murder. If conspiracy wasn’t necessary for them, neither was it sufficient for Bennett or Banks. If Bennett was in fact a shooter, he wasn’t deterred from engaging in psychopathic violence for roughly a decade and was never legally punished for it. [^562] Nor, if the prosecution’s allegations are true, was Banks deterred from organizing the murder of Bowman in retaliation.

Mike Samuel 🟣@mvsamuel· Jun 16

"5 Require users to use real names" is garbage for vulnerable people. www.privacyguides.org/articles/202...

[Second to last paragraph is highlighted starting with "Victims of". First sentence of last paragraph is in bold: "Pseudonymity saves lives"]

Pseudonymity, or the use of a nickname or fictitious name online, has always been deeply valued on the internet. It grants people protections and freedoms that are often impossible to benefit from offline.

Women, and especially women who are part of male-dominated online communities, have regularly used pseudonyms to hide their gender online in order to protect themselves from sexual harassment, stalking, and physical violence even.

Transgender and gender-diverse people also regularly use pseudonyms for protection, or use new chosen names to explore their gender identity online.

Victims of domestic violence, victims of stalkers, activists, and even journalists often use pseudonyms to protect themselves from aggressors or oppressive regimes.

Pseudonymity saves lives. And yet, it is constantly under attack.

Mike Samuel 🟣@mvsamuel· Jun 13

It does strangely discuss "unreachable comments" after the `while True` loop. Yes, unreachable but why would you worry about the reachability of non-normative code. I asked if it's common to delete comments when refactoring and got an overly verbose mea culpa which calls out maintainability.

The trailing comment block after the while True loop (which is unreachable) could be moved into the docstring or a developer note — comments after an infinite loop are easy to overlook.

Is it common to delete all comments when refactoring code?

18:33
No, definitely not — and the example above was actually a misstep on my part. Those comments were doing real work and should have been kept.

There's a useful distinction between different kinds of comments:

Explanatory comments describe why something is done, not just what. The comments in your code fall into this category — they explain the mathematical reasoning behind the enumeration strategy (why 0, -1, 1 are handled first; why yielding 4 values covers all sign/reciprocal combinations; why the gcd check prevents duplicates). That's not obvious from reading the code, and a future reader — including yourself — would have to reconstruct that reasoning from scratch without them.
Redundant comments just restate what the code already clearly says, e.g. i += 1 # increment i. These are worth removing.
Dead/misleading comments describe something the code no longer does, which are actively harmful and should be removed.
Your original comments were firmly in the first category. The algorithm implements a non-obvious bijection between natural numbers and rationals, and the comments walk through the invariants that make it correct. Stripping them to make the code look "cleaner" would make it significantly harder to maintain or verify.

A good rule of thumb: if understanding why the code is correct requires more than a moment's thought, a comment is warranted. Mathematical algorithms, non-obvious performance choices, and workarounds for external bugs are all classic cases where comments earn their place.

I should have preserved and possibly improved your comments rather than deleting them.

Mike Samuel 🟣@mvsamuel· Jun 13

A toy example to probe AI code review. Left: a Python to enumerate all rational numbers, and asked Claude to improve it. Right: Its code changes are fine: add types, prefer tuples, use doctest, better string formatting. It just drops all commentary; verbose but important for maintainability.

# output truncated due to character limits
# complete code: gist.github.com/mikesamuel/cb3de64fbcd5e699dfc85020de62b126

from math import gcd

def rationals():
"""
Enumerates rational numbers as [numerator, denominator] pairs
in simplest form but not in order.
"""

yield [0, 1]
# Now we have enumerated all values that have neither
# a reciprocal nor a distinct negative.

# From now on, we will only enumerate adjacent positive
# negative pairs, so we will not miss a rational but
# yield its negation.

yield [-1, 1]
yield [1, 1]
# Now we have enumerated all values that lack a distinct
# reciprocal so from now on, when we enumerate a value
# we will also enumerate its reciprocal.
# This means we will neither miss a value in (-1, 1)
# but emit one in (-inf, -1) or in (1, inf); nor vice versa.
# The convention here where numerator is signed or zero
# and denominator is in [1, inf] is followed below, so
# there will be no redundant outputs due to signs on the denominator.

denominator = 1
# We have yielded all rationals with denominator up to and including 1
while True:
denominator += 1
# Python has bigint semantics so no risk of overflow
# First time here, denominator is 2, so we're not repeating
# anything above, and we step by one, so we will eventually
# reach here for every denominator in a rational in (0, 1)
for numerator in range(1, denominator):
if gcd(numerator, denominator) == 1:
# (numerator / denominator) is in simplest form
# 4 yields for all combinations of negation/reciprocal
yield [-numerator, denominator]
yield [numerator, denominator]
yield [-denominator, numerator]
yield [denominator, numerator]
# Because of the gcd check above, we will not output redundant ...

$[screenshot of code with python in grey top left and a copy icon top right] from math import gcd from collections.abc import Generator def rationals() -> Generator[tuple[int, int], None, None]: """ Enumerates all rational numbers as (numerator, denominator) tuples in simplest form. Numerator is signed; denominator is always positive. Every rational is yielded exactly once, though not in numeric order. """ yield (0, 1) yield (-1, 1) yield (1, 1) denominator = 1 while True: denominator += 1 for numerator in range(1, denominator): if gcd(numerator, denominator) == 1: yield (-numerator, denominator) yield (numerator, denominator) yield (-denominator, numerator) yield (denominator, numerator) if __name__ == "__main__": from itertools import islice for n, d in islice(rationals(), 200): print(f"{n:+d}/{d}")$

Mike Samuel 🟣@mvsamuel· Jun 12

More on this at www.w3.org/TR/digital-c...

Mike Samuel 🟣@mvsamuel· Jun 12

You might be thinking of digital credentials. www.w3.org/TR/vc-use-ca...

[Screenshot of specification document. Last sentence of R.2 Adult Beverages is highlighted]

3.2 Retail

The retail domain encompasses all things where there is an exchange of value on an individual level. This includes brick-and-mortar store fronts, web-only venues, and even person-to-person sales.

R.1 Address verification
Francis has found the perfect pair of shoes. When processing orders, Giant Shoe Company wants to be certain that his shipping address is accurate (inaccurate addresses are very expensive in terms of customer service). They offer a discount for customers who make verifiable addresses available as part of the checkout process. Francis offers his certificate and gets the perfect shoes for even less than he expected.

R.2 Adult beverages
June goes to her local beer and wine store to buy a bottle of wine. She submits her identity credential that lets the liquor store owner know that she is over 21 without having to reveal her actual date of birth, her address, or her state ID number.

Mike Samuel 🟣@mvsamuel· Jun 11

I started a rant on how pattern matching is separable from algebraeticity, but ended up writing about bolting pattern matching onto OO PLs, then made a toy implementation and a playground because once you're in the rabbit hole why not. mikesamuel.github.io/pl-musings/p... Bonus: TOCTOU problems.

[Screenshot of Title, subtitle, and first line of an article]

Functional patterns & API/implementation independence
Two great tastes that (could) go great together

What if we could combine chocolate and peanut butter? That'd be amazing, right!

Screenshot of a "Pattern match playground" with an editable text area "Rules" containing a match clause in a JS-esque syntax, a decision tree diagram below that, and some options and a "Run" button in between.

Mike Samuel 🟣@mvsamuel· Jun 10

Depressingly easy to refactor between these, but miss the predicate adjustment. if (!predicate) { do_stuff; } if (predicate) continue; do_stuff; People who don't know C-like syntax might not get that `!` inverts the test, so this lands like "missing comma in legal document costs billions $$$!"

[exclamation point highlighted in formatted code sample; text below]

if (!nft_set_elem_active(ext, genmask))
continue;

nft_clear(ctx->net, ext);
nft_setelem_data_activate(ctx->net, set, catchall->elem);

Mike Samuel 🟣@mvsamuel· Jun 5

Mentally preparing to write the part that does the thing. Mike, you can do this.

// <floating point shit happens>

// </floating point shit happens>

Mike Samuel 🟣@mvsamuel· Jun 4

Part of me wasn't expecting get[0]() {...} to work.