Mike Samuel 🟣
LightNews LightNews
banner
mvsamuel.bsky.social
Mike Samuel 🟣
@mvsamuel.bsky.social
660 followers 670 following 1.3K posts
I solve large software systems problems with programming language techniques.

Previously, I was the first frontend engineer on Google Calendar, and was a security engineer who worked on the industrial-strength Mad Libs undergirding Gmail.
Posts Replies Media Videos
mvsamuel.bsky.social
Sun Tzu wouldn't have had to write _the Art of War_ if he could've sat his generals down with _the Oregon Trail_.
Picture of an old style handheld game console on a wooden table. The small console screen has "The Oregon Trail" displayed at the top, with white text on a black background. Pixelated olde scrollwork separates the title from the body text, and the "press ENTER to continue" at the bottom. The body text has a question followed by 4 options. Text follows. Many kinds of people made the trip to Oregon. You may: 1. Be a banker from Boston 2. Be a carpenter from Ohio 3. Be a farmer from Illinois 4. FInd out the differences between those choices What is your choice? 1
November 6, 2025 at 4:14 PM
mvsamuel.bsky.social
Most of mmitchell's 13 mitigations don't suggest any override, and these 4 from the issue she linked are more about involving irl friends and blocklist user's understand who is on the blocklist.

Framing a request as a demand is a bit over the top too.
Screenshot of a items 6-9 of a numbered item list in a Github issue report. 6. Develop a mechanism for those who have been Blocked to be able to reach out to specific friends/colleagues on the app who have inappropriately blocked them in some way to alert them to the issue, for example, via a submission form moderated via Bluesky 7. Add a flag, warning, or some form of disclosure for users viewing a Block List: “People you follow are on this list” 8. Add a flag, warning, or some form of disclosure for users viewing a Block List: “People in your network follow people on this list” 9. When someone reposts content and another user sees it as Blocked, provide disclosure for that user on which Block List(s) are creating the block (related feature idea at "Blocked post" message should indicate the blocker #2127)
November 4, 2025 at 6:48 PM
mvsamuel.bsky.social
From x.com/JFREJNYC/sta...
Spotted in Borough Park
November 4, 2025 at 6:39 PM
mvsamuel.bsky.social
Ah, that nicely links voltage drops with apathy among policymakers.

(emphases added)
Screenshot of conclusion from https://behavioralscientist.org/the-five-vital-signs-of-a-scalable-idea-and-how-to-avoid-a-voltage-drop/. Emphases added to last sentence in first paragraph starting with "Policymakers, then" and the third in the second paragraph starting with "I advocate." researcher has conducted an efficacy test. Policymakers then, unwittingly, choose a certain number of published programs to scale, and the results are usually disappointing. The policymaker subsequently views the next generation of research with a more skeptical eye, and scientists lament how their work is being ignored. In my new book, given the ubiquity of such voltage drops and their five causes, I propose a different approach to generating evidence. Beyond exploring whether an idea works in the petri dish in the best case scenario with a select group of people, I advocate augmenting traditional experimental designs by bringing the warts and constraints that the idea faces at scale into the petri dish in the original testing. This is meant to move the question to: For whom, when, where, and why does my idea work? When such an approach is taken, we are better able to use science to advance truly scalable ideas, rather than use guesswork and gut feelings to choose which ideas should be scaled.
November 3, 2025 at 9:00 PM
mvsamuel.bsky.social
People who talk about AGI as human replacing hyper-intelligences are still going to be unhappy with this definition. I do like the way they drop shade on Microsoft's overly-investor-focused but still vague and useless definition.
Screenshot of a paragraph split across a page break with page number 15 just above the split. The phrase "human-level AI" is highlighted. Text follows: Our AGI definition is about human-level AI, not economically-valuable AI, nor economy-level AI. OpenAI and Microsoft have reportedly considered AGI to be an AI that can generate $100 billion in profit (TechCrunch, 2024). We do not conflate AGI with economically valuable AI because narrow technologies, such as the iPhone, can generate billions in economic value, despite not being generally intelligent. Meanwhile, Replacement AI is about economy-level AI, and it includes physical tasks, unlike AGI.
November 3, 2025 at 6:22 PM
mvsamuel.bsky.social
I don't place much stock in AGI discourse, but Dawn is not someone to dismiss lightly.

_A Definition of AGI_ by D.Hendrycks, D.Song, et al: arxiv.org/abs/2510.18212
Artificial General Intelligence (AGI) may become the most significant technological development in human history, yet the term itself remains frustratingly nebulous, acting as a constantly moving goalpost. As specialized AI systems master tasks once thought to require human intellect—from mathematics to art—the criteria for “AGI” continually shift. This ambiguity fuels unproductive debates, hinders discussions about how far AGI is, and ultimately obscures the gap between today’s AI and AGI.
November 3, 2025 at 6:18 PM
mvsamuel.bsky.social
@leaflet.pub Do you need all these permissions?

Your service looks cool so I'll give it a try with an unlinked account, but I'm not going to connect accounts if it requires broadening who can manage my profile and posts and if it allows for unlimited delegation.
Screenshot of oauth login with leaflet.pub wants to access your mvsamuel.bsky.social account at the top and "Deny Access"/"Authorize" at the bottom. The middle text, quoted below, includes the requested authority: https://leaflet.pub/api/oauth/metadata wants to access your mvsamuel.bsky.social account Email Read your account's email address Bluesky Manage your profile, posts, likes and follows as well as read your private preferences Repository Create, update, and delete any public record Authenticate Perform authenticated actions towards any service on your behalf By clicking Authorize, you will grant this application access to your account in accordance with its terms of service and privacy policy.
October 31, 2025 at 3:56 PM
mvsamuel.bsky.social
Good times.

www.theatlantic.com/technology/a...
is still being cleaned up. For a few weeks at the end of 2012 and into 2013, toxic vapors got into two campus buildings, possible exposing the office workers there to levels of chemicals above the legal limit set by the EPA. A
October 28, 2025 at 5:24 PM
mvsamuel.bsky.social
In case you didn't already know. SV is a bit larger than Santa Clara county, but it's not wrong.

cleanair.camfil.us/2017/11/21/t...
Silicon Valley is located in California’s Santa Clara County. Santa Clara is home to more Superfund sites than any other county in the United States. Superfund sites are areas that the EPA has designated as being contaminated by hazardous waste and require special attention to clean up the area. Why is Silicon Valley home to more Superfund sites than anywhere else in the country? What can be done to clean up Silicon Valley Superfund sites?
October 28, 2025 at 5:19 PM
mvsamuel.bsky.social
I get that we Australians have to own Rupert Murdoch, but Gibson is one of your lot.
Screenshot of Personal details from https://en.wikipedia.org/wiki/Mel_Gibson Personal details Born Mel Columcille Gerard Gibson January 3, 1956 (age 69) Peekskill, New York, U.S. Citizenship United StatesIreland
October 28, 2025 at 3:15 PM
mvsamuel.bsky.social
Cast operators in modern PLs have two distinct senses making them effectively bivariant.

The `as` operator requires there be some relationship between the left operand's type and the type specified in the right operand. But it could sensibly be subtype or supertype.

This is a PiTA to solve.
Sceenshot of syntax-highlighted code in a language that looks superficially like TypeScript. let a = myCat as Animal; // widening type(left) <: right let b = anAnimal as Cat; // narrowing right <: type(left)
October 24, 2025 at 6:45 PM
mvsamuel.bsky.social
Bradbury's *The Martian Chronicles* cover up to 2026.

Maybe scifi.fandom.com/wiki/Timelin... has more fodder. At a glance, 2021 and 2063 seem busy. 2552 if you're into Halo and palindromery.
Screenshot of excerpt from https://en.wikipedia.org/wiki/The_Martian_Chronicles Text "ranged from 1999 to 2026" is highlighted. ---- Chronicle structure The Martian Chronicles is written as a chronicle, each story presented as a chapter within an overall chronological ordering of the plot. Overall, it can be viewed as three extended episodes or parts, punctuated by two apocalyptic events. Events in the book's original edition ranged from 1999 to 2026. As 1999 approached in real life, the dates were advanced by 31 years in the 1997 edition. The summary that follows includes the dates of both editions.
October 24, 2025 at 3:33 PM
mvsamuel.bsky.social
Might you have meant Collins?
Screenshot of the last two lines from the OP's article. Last word of first line, "Mills", is highlighted. state. His allies are not doing him any favors if they allow him to believe otherwise, and if you don’t like Mills you need to find someone else to challenge her.
October 22, 2025 at 5:57 AM
mvsamuel.bsky.social
Do R users really kvetch about its 0-basedness enough that this is worth calling out?

I suppose Fortran users get to feel smug either way.
https://juliahub.com/products/julia Screenshot of a website. "JuliaHub" logo at the top left, discreet tabs top middle, and search box and "Login" at the top right. Under a white-on-blue boxed header "Julia Comparisons >" are six rectangles of text in 3 columns x 2 rows. Each box has a header of the form "Julia vs " some other programming language. The other languages are: MATLAB, Python, R, C/C++, Fortran, Mathematica. Some text in the "Julia vs R" box is highlighted: "1-based indexing" Full text of that box follows: Julia provides R-like statistical capabilities with dramatically faster execution. 1-based indexing, better memory management, and native support for parallel computing make Julia ideal for large-scale data analysis.
October 17, 2025 at 5:45 PM
mvsamuel.bsky.social
Yeah. Racket's syntax desugaring breaks emoticons.
Screenshot of racket interactive shell session showing a backtick quoted emoticon entered at the command line and a different one printed as output. 🐚$ racket Welcome to Racket v8.17 [cs]. > '(:- {}) '(:- ()) >
October 7, 2025 at 10:33 PM
mvsamuel.bsky.social
sh and bash are similarly eager & streaming; a shell script can rewrite its tail. Also, there're multiple stages of lexical processing and substitution.
dl.acm.org/doi/abs/10.1... is an attempt at semantics there. iirc, it gets the line orientation but makes simplifying assumptions about the FS.
Filesystem. Our model of the filesystem is quite simplistic: we track a hierarchy of files and directories, but not file contents. Such a simple model suffices to simulate pathname expansion and file descriptor redirections (dup) and heredoc redirections (here), but not file redirections (file). It is a matter of engineering effort to produce a better symbolic filesystem. It would be interesting to link our symbolic system up with SibylFS (also written in Lem) [Ridge et al. 2015] or Forest [Fisher et al. 2011], or to use Ntzik et al.’s reasoning [2017; 2018]. We can also imagine implementing a read-only filesystem that allows access to the real, underlying filesystem, but treats writes (and other dangerous operations, like execve) as noops. We consider platform-specific filesystems, like /proc, as out of scope; we could in theory apply platform-dependent reasoning [Nita et al. 2008].
October 6, 2025 at 4:53 PM
mvsamuel.bsky.social
Btw, for the exhaustive answer to this, the CLDR data tables are surprisingly accessible.

cldr.unicode.org/translation/...

Those show a timezone after the am/pm marker.
Screenshot from github. URL bar reads https://github.com/unicode-org/cldr-json/blob/main/cldr-json/cldr-dates-full/main/en-001/ca-generic.json The code browser shows dateSkeletons and timeFormats. Text below. "dateSkeletons": { "full": "GyMMMMEEEEd", "long": "GyMMMMd", "medium": "GyMMMd", "short": "GGGGGyMMdd" }, "timeFormats": { "full": "h:mm:ss a zzzz", "full-alt-ascii": "h:mm:ss a zzzz", "long": "h:mm:ss a z", "long-alt-ascii": "h:mm:ss a z", "medium": "h:mm:ss a", "medium-alt-ascii": "h:mm:ss a", "short": "h:mm a", "short-alt-ascii": "h:mm a" },
October 1, 2025 at 2:22 PM
mvsamuel.bsky.social
At the risk of over-defending a fellow Australian, I don't think Tao has been as silent on politics as is being reported, but agree with you in the large.

terrytao.wordpress.com/2024/11/08/t...
Screenshot of a comment on a blog post with "8 November 2024, Terence Tao" at the top left, a small headshot of Terence on a beige background at the top right. The last few sentences, starting with "Perhaps my most notorious example ..." at the bottom are highlighted. Text follows: I actually have made efforts to do that, both here (e.g., in my recent post on the Venezuelan election), and on my Mastodon account (I have collected some of those posts here). I am definitely planning on posting quite frequently on my Mastodon on many topics going forward, including on the intersection of math with the real world. Previously, when I ventured too far from the abstract world of pure mathematics, I got pushback of various types, including ones along the lines of “You should stay in your lane”. Perhaps my most notorious example of this was my surprisingly topical 2016 post “It ought to be common knowledge that Donald Trump is not fit for the presidency of the United States of America“. Staying in one’s lane can be considered reasonable advice under normal traffic conditions. But the traffic ahead will be the complete opposite of normal, and so we have to change our driving habits now.
September 8, 2025 at 5:54 PM
mvsamuel.bsky.social
Confirmed.
Screenshot of interactive python3 session. >>> import dateutil.rrule >>> dateutil.rrule <module 'dateutil.rrule' from '...'> >>> del dateutil.rrule >>> dateutil.rrule Traceback (most recent call last): File "<stdin>", line 1, in <module> AttributeError: module 'dateutil' has no attribute 'rrule'
September 3, 2025 at 4:05 AM
mvsamuel.bsky.social
∃
Book title page with "SURREAL NUMBERS" in a Serif font at the bottom, widely spaces. In the middle, a stone wall with oddly shaped stones in a line-art style. The wall has 6 stones in five columns and hints at more stones to the side. Perhaps they're meant to spell "(oil)" Above the stone wall: how two ex-students turned on to pure mathematics and found total happiness Between the wall above and the title below: a mathematical novelette by D. E. Knuth
September 2, 2025 at 5:06 PM
mvsamuel.bsky.social
He is Australian, not American and might fear that an admin that deports people on work visas for speaking out about the admin would target him.

But he's said this: terrytao.wordpress.com/2024/11/08/t...
Screenshot of a comment on a blog post with "8 November 2024, Terence Tao" at the top left, a small headshot of Terence on a beige background at the top right. The last few sentences, starting with "Perhaps my most notorious example ..." at the bottom are highlighted. Text follows: I actually have made efforts to do that, both here (e.g., in my recent post on the Venezuelan election), and on my Mastodon account (I have collected some of those posts here). I am definitely planning on posting quite frequently on my Mastodon on many topics going forward, including on the intersection of math with the real world. Previously, when I ventured too far from the abstract world of pure mathematics, I got pushback of various types, including ones along the lines of “You should stay in your lane”. Perhaps my most notorious example of this was my surprisingly topical 2016 post “It ought to be common knowledge that Donald Trump is not fit for the presidency of the United States of America“. Staying in one’s lane can be considered reasonable advice under normal traffic conditions. But the traffic ahead will be the complete opposite of normal, and so we have to change our driving habits now.
August 27, 2025 at 3:16 PM
mvsamuel.bsky.social
I suspect I can't give you useful guidance as to users' preferences re simpler types / simpler semantics tradeoffs.

One datapoint tho: here're two industry languages, one typed and one dynlang, that have both chosen A style for null-safe calling: arguments are not evaluated when there is no callee.
A screenshot of a shell session that invokes the Kotlin REPL, runs some code, and does the same for node. Text below elides deprecation warnings and boilerplate: 🐚$ kotlin -Xrepl Welcome to Kotlin version 2.2.0 > var callee: ((Unit) -> Unit)? = null > fun prints() { println("Hello, World!") } > callee?.invoke(prints()) res2: kotlin.Unit = null > ^D 🐚$ node Welcome to Node.js v24.5.0. > function prints() { console.log("Hello, World!"); } undefined > null?.(prints()) undefined >
August 25, 2025 at 7:10 PM
mvsamuel.bsky.social
tbf, the 2007 Stupid Backoff paper lists web as its fourth of four corpora and by far the largest.

aclanthology.org/D07-1090.pdf
Screenshot of a few column inches from a paper. The top has a list of four items, each starting with a bolded term (target, ldcnews, webnews, web) and a colon. It's followed by a regular indented paragraph. Text follows: target: The English side of Arabic-English parallel data provided by LDC5 (237 million tokens). ldcnews: This is a concatenation of several English news data sets provided by LDC6 (5 billion tokens). webnews: Data collected over several years, up to December 2005, from web pages containing predominantly English news articles (31 billion tokens). web: General web data, which was collected in January 2006 (2 trillion tokens). For testing we use the “NIST” part of the 2006 Arabic-English NIST MT evaluation set, which is not included in the training data listed above7. It consists of 1797 sentences of newswire, broadcast news and newsgroup texts with 4 reference translations each. The test set is used to calculate translation BLEU scores. The English side of the set is also used to calculate perplexities and n-gram coverage.
August 21, 2025 at 8:04 PM
mvsamuel.bsky.social
You're right about mgmt pushing this (super frustrating given how hard it is to get them to pay for developer tools devs want) but it's rolling downhill to them.

The industry pushes itself via pieces like the below which give board members questions to grill C-suiters with.

jbai.ai/index.php/jb...
Screenshot under headline "Technology (Chief Technology Officer - CTO)" with a short paragraph followed by a bullet list of questions. Text follows: Since AI is a novel technology, it is not surprising that your CTO will be very involved with developing and implementing your AI strategy. Most CTOs are comfortable learning about new technologies, and despite its potentially world-changing impact, the fundamentals of AI are not rocket science. Expect your CTO to be able to make intelligent decisions and advise your CEO. What she or he might not be doing is taking AI seriously enough. Here are some probing questions to consider asking your CTO: - What steps are we taking to protect our AI systems from being manipulated or compromised through cyberattacks? For example, there are techniques for corrupting AI models by feeding them compromised data. - What AI technology should we / must we create ourselves, and which AI applications should we undertake using outsourced tools? - Which capabilities should we secure through strategic partnerships? - Can AI be used to write more of our code? What are the risks? - Should we pursue an open-source approach to AI, such as Facebook, or a proprietary approach like OpenAI? - Can you rank the top ten AI technologies and applications that will have the largest impact on our company? - What is our AI technology stack currently? What should it be? - What are the risks and benefits of relying on nimble AI startups who may not yet have proven their business model, and might disappear?
August 21, 2025 at 3:53 PM
mvsamuel.bsky.social
A bit more progress on type solving. Call constraints have a lot of moving parts: multiple overloads to choose among, explicit type actuals and vars to expose computed type actuals.
They're like mini state-machines: some linkages only make sense once excess overloads have been eliminated.
A dot graph with tables for nodes. The diagram has some simple nodes("call", or "uses") but most nodes have a bold name at the top like 'gPass and edges coming out labeled with operators (<: or ==). The nodes tend to have table rows under their label with operators at the left and types at the right. Some nodes have a type like List<Int> right under their label with no operator. At the right is a cluster of nodes most of which have labels like ValueBound([]) or List<('listLiteral0)> feeding down into nodes with simple ticked names like 'h0. There is a similar stack on the far left that feeds into a List<('P0)> that connects in the middle to a "put" node that feeds into 'gActuals. The bottom middle is full of already-solved nodes: nodes whose name is a backticked label and which has a row under it with a type with no further backticked names.
August 20, 2025 at 3:48 PM