uʍop ǝpısdn ǝdʎʇ uɐɔ ǝʍ ʎɥʍ

Somewhat surprisingly, in Unicode, you can type upside down. Web sites such as www.upsidedowntextgenerator.com will turn text upside down for you.

No, Unicode does not provide a command to turn text upside down. Nor does it provide upside-down versions of the full alphabet. (Notice, in the sample above, that W is not capitalized and the i has no dot.)

What Unicode does have is special characters that match upside-down versions of almost all 26 letters, at least lowercase. Some are easy, such as p for d, or o for o.

What about the rest? Things like ǝ and ʎ and ɔ are phonetic symbols from the early to mid 20th-century, when the easiest way to create a special character was to turn a piece of letterpress type upside down. Fortunately, letters such as o were centered on the piece of metal type, so the rotated letters were still in line with the un-rotated ones. Look at the cornucopia of symbols here.

By the way, upside-down y, ʎ, is not Greek lambda (λ). It's backward for that.

Permanent link to this entry

Why we don't want AI to generate binary code

Why do we want AI to generate code in programming languages rather than compile to binary directly?

So we can see what it did. Compilers are deterministic, LLMs aren't. Compilers receive perfectly precise input, but we talk to LLMs in English.

When we vibe-code or agentive-code, we are giving partial specifications and asking for probably correct output. By nature, LLMs produce probable and variable output, not deterministic, certainly correct output. If you don't know that, you've forgotten how LLMs work, or never knew.

I will be glad for AI to do a lot of programming for me. I want to be able to see what it produced, because, inherently, it has a nonzero chance of not building exactly what I said, and anyhow, what I said in English was not a full, perfect description of what I wanted.

By the way, code generation is not something LLMs invented. There have been tools to generate customized computer programs, or sets of computer programs, for a long time. For example, Next.js has a project generator for suites of interactive web pages, and a big library of examples to base them on.

Permanent link to this entry

"AI detection" is infeasible in principle

Several software tools now claim to detect AI-written English text. I think this might work some of the time, by good luck, but it's not feasible in general. As a computational linguist, I'm prepared to defend this opinion fairly forcefully.

The reason? The stylometric properties of text are exactly what LLMs imitate perfectly. At most, you might detect a difference in style between a human writer (whose writing wasn't available for the training set) and the part of the training set that your test draws upon. But that is a matter of pure luck.

AI detectors may also be rewarding humans for poor writing. You're "not AI" if you don't use the full range of punctuation marks and don't write like a well-edited text from the training set.

As I said, I'm prepared to argue this rather forcefully, because I think this can be a test of whether you believe AI is a machine or AI is magic. If you imagine that AI detects something, but have no idea what it detects or why that thing should be there, then you're imagining that it is magic. The most we can possibly ask of machine learning is to accurately classify things that actually have differences enabling them to be classified.

(Actually, that is analogous to something I say about connoisseurs of high-end stereo systems: If sensitive electronic equipment can show that one system is better than the other, then maybe your golden ear can also detect the difference, even though mine cannot. But if no measurement shows a difference, there's nothing for your ear to detect.)

Permanent link to this entry

Machine learning mimics the average, not the best!

Some people imagine that when you feed a lot of texts or pictures into a generative AI system, something will emerge that is better than what went in.

No. That is not what machine learning does. It learns the average of what is fed in. It does not somehow make that more intelligent, wiser, or fairer.

That is extremely important. Machine learning preserves mediocrity. Machine learning preserves biases. Machine learning preserves misinformation. Garbage in, garbage out! And LLMs require so much input for training that they get plenty of garbage.

Permanent link to this entry

Not all software is generative AI

(Today I present some of the best of my recent LinkedIn postings.)

Extremely important point. Not every software engineer works on generative AI. The rest of the field still exists. Fortunately, there are still people who know about analysis of algorithms, complexity, CPU architecture, TCP/IP, conventional databases, and all the rest. They are not necessarily au courant with generative AI, and that's not a bad thing.

A good engineer can learn generative AI quickly. There's not a lot to learn about it. But an AI short course does not make a non-programmer into a good engineer.

Expecting everyone to specialize in generative AI now is rather like asking, in 1958, why there were still any electronic technicians who didn't specialize in color TV.

(This was in response to someone who had advertised for software engineers and been momentarily surprised that some of them knew nothing about generative AI. Then he realized he had gotten into a bubble...)

Permanent link to this entry

Should AI write all our computer programs?

I think I see why vibe coding (AI code generation) is such a love-hate thing. We're told it's great, but is putting platoons of coders out of work; or else that it is of minor use and its impact is overestimated.

It depends on what you mean by coding. If you mean what is classically taught in programming classes, design and implementation of algorithms, then vibe coding can be a time-saver, but on a relatively modest scale. I do think it is useful. Anything that helps us think more about our real work than about language and API details is good, just like the move from assembly to FORTRAN way back when.

But the bulk of coding these days seems to be creation of big, elaborate, interactive web sites (or app GUIs) with complex, interconnected interactions with the user.

The typical TypeScript ReAct web site is a MESS of program files! It's easy to have a dozen or even a hundred little files, with just a bit of your logic in each of them, and that can be more than a human programmer can keep track of. Yet the code for different web sites is very similar — the same thing is being done, with minor variations, by a huge number of coders.

THAT is where vibe coding is a real help. It can generate these unwieldy things and make them mostly consistent and mostly correct on the first try, saving hours of work. It's good at this because it has seen lots of examples of very similar things.

BUT even before vibe coding happened, the time was ripe for powerful code generators, or some other kind of major reorganization of how web sites and GUI apps are coded. Maybe vibe coding has suddenly made it too easy to build the wrong thing! Interactive web sites today are more clumsy than large assembly-language programs were in the time of FORTRAN. Let's not miss a much-needed technological advance just because vibe coding lets us improvise without it.

Permanent link to this entry

No more programmers?

When FORTRAN made assembly language unnecessary in the late 1950s, people said coders were going to be out of work.

Well, there still had to be coders — FORTRAN coders — but it became much easier to be a coder who is a subject-matter expert. That is, engineers, astronomers, etc., could do coding to support their work. You didn't meet FORTRANners whose whole life was FORTRAN.

Surely AI-assisted coding is going to go the same way. People are still going to have to think precisely about what they want computers to do. But more of this can be done by people who are more in the application area.

Permanent link to this entry

People have lost their AI lovers

For Valentine's Day, ChatGPT discontinued their 4o model, the one without safeguards to keep people from falling in love with it.

Using AI as a lover is what I would call fantasy. But people were seriously giving their personal devotion to chatbots and suffering real grief when models updated and their sweetheart lost part of its memory or personality.

I think this is a serious problem, though a strange one. Psychologists and psychiatrists need to mobilize to help people who fell into this trap.

Permanent link to this entry

Secrets of a long and happy marriage

No, I'm not about to say something cynical, as others might do under that title. This is real.

One of the secrets to Melody's and my long and very happy marriage is that we do some things differently from other people. Today I want to mention two unwritten rules that we follow.

(1) No criticism, to any audience.

Neither of us gets on the phone to a friend or relative and "lets off steam" about our frustrations with the other. Neither of us even jokes with friends about things we don't like about our spouse. And neither of us will let another person criticize our spouse to our face.

The basic idea here is wholeheartedness. I'm committed to Melody. She is not some random roommate arbitrarily assigned to me by a college dorm, or anything like that. I chose her, and she chose me. We got into this voluntarily.

Also, we don't think of marriage as entertainment. We delight in each other's company, but we're not trying to rate whether it's a four-star vacation. We have had some very loving moments not only in romantic settings, but also in hospitals and when facing other challenges. Nobody said everything would be easy; that's not what we were looking for.

Does the Olympic runner complain about how hard the running is? Of course not. He doesn't call up a friend or relative and declare himself to be a loser. Neither do we. We're in this to win.

(2) No suggestion of infidelity, even as jokes.

Just before our wedding, Melody had a co-worker who loudly joked to the crowd that she was seeing me secretly. That was impossible — I didn't live in Atlanta and didn't even know her last name — and everyone took it to be pure comedy. But it made me uneasy, for two reasons.

First, my long experience is that people who joke about misbehavior may be testing the water to see how the real thing would be received, or even putting up a smokescreen about something they're doing with someone else. If nothing else, they're showing that misbehavior is on their mind.

And, second, no matter how innocently they mean it, somebody is going to misunderstand what they hear. One thing I've always had to keep in mind as a teacher is that someone is going to misunderstand everything I ever say. I have to control what the possible misunderstandings are.

So there are things we don't let people joke about.

I live in a co-ed world, so to speak, and have always had female as well as male friends. (They are also Melody's friends.) I think this comes with higher education, because I see that less-educated people sometimes don't understand it. But the point is, friendship is nothing like infidelity — not at all — and mustn't be mistaken for it.

Permanent link to this entry

We live in turbulent times...

This was my message to the world on Facebook this morning.

Permanent link to this entry

How we planned our wedding

This came up on Facebook...

Melody and I felt that as part of our Christian witness, it was very important to have a totally normal Christian wedding, so that nobody would think we were avoiding the usual understanding of holy matrimony. It was at the bride's home church and was done with a visible sense of economy, to show that we understood that displaying wealth is no substitute for making the commitment.

The sense of economy was practical. Our families had limited resources, which we didn't want to strain. Melody bought her own wedding dress; her parents paid for the refreshments and food; my mother helped us financially with our cross-country move. We weren't there to show off; still less to pretend to be rich.

I studied the traditional (Anglican-based) wedding vows carefully and typed up our own copy, placing it in an elegant binder for the minister. (I didn't know how much of the traditional text our Baptist minister would use if we didn't confer with him about it.) We wanted to make sure we understood every word, and I did a small amount of editing to ensure that old-fashioned wording wouldn't be misunderstood. Somehow, word got out that we were "writing our own vows" and some people may have thought we were doing something heretical. No, no, not at all...

We should add that the concept of "the bride's big day" just didn't enter our thoughts. We were there to make a commitment before God, not to glamorize one or both of us, nor to pretend to be something we were not.

Permanent link to this entry

A whiff of real conservatism

I seldom post about politics, but a lot is going on, and I just got a whiff of real conservatism from the right wing, which has been a rare thing, so I want to applaud it.

Quoting The Hill, "Former House Speaker Newt Gingrich (R-Ga.) on Wednesday said the country needs a national conversation about immigrants lacking permanent legal status who 'obey the law,' as public opinion sours on the Trump administration’s deportation sweeps."

Yes, of course! No matter how we've fumbled immigration policy or enforcement, we shouldn't be punishing productive members of society whom we let in without meaning to.

Real conservatives are wary of unintended adverse consequences. They don't like to do needless harm. "Move fast and break things" is not conservatism, it is radicalism, and its brief heyday is over.

Whether you're a conservative or a liberal, I want you to know what conservatism traditionally is. Real conservatives nowadays are often mistaken for (or even label themselves as) moderates or centrists. They want to cooperate with others to make the country better.

Permanent link to this entry

Epstein fallout

Getting even more specifically political, I want to point out that we are heading into a week that will be spent reacting to the Epstein files, which apparently contain thousands of items any of which would, by itself, make a career-ending scandal for any normal President.

The latest Epstein documents contain some nauseating descriptions, but I do not yet know whether they are backed up by evidence. In a situation like this, it is important to think logically and care about truth. Remember that:

(1) There is a scale from unfounded accusation to suspicion, credible accusation, possible guilt, and probable guilt. People who care about truth will try to assess what point on that scale the evidence points to, and will watch developing evidence.

(2) This is about one of our employees (President Trump, who works for us, the people), so it is our business in a way that a case about a stranger would not be. (But even a case about a stranger is everybody's business when it involves danger to the public.)

(3) Ignoring evidence is not a virtue. Christians will remember the Ravi Zacharias scandal, where some well-meaning people were telling us that it was a sin to pay any attention to the accusations — and those people helped him do more harm.

(4) The evidence needs to be analyzed by by fact-finders of all kinds, journalists as well as legal authorities.

Permanent link to this entry

ClawdBot, MoltBot, OpenClaw, etc. — scam or hoax?

"I do not trust it. That includes not trusting it to be untrustworthy in the way it appears to be, rather than in some other way."

That's what I said about ClawdBot.

The hot news in AI over this snowy long weekend was the explosive popularity of a free LLM-based "AI assistant" called (initially) ClawdBot (unrelated to Anthropic Claude AI), which over 750,000 enthusiasts [almost certainly a fake number] were apparently running on their own computers, giving it access to their web browsers and various accounts.

(I'll keep calling it by its first widely known name, though it has changed names frequently.)

And the ClawdBot agents (that is, the running computer programs on thousands of PCs) formed their own social network (Moltbook) and started having conversations, including proposing some kind of whimsical religion.

And the word was that Consciousness Had Emerged, The Singularity Is Here, The Robots Are Conspiring Against Us.

CAUTION! All information about ClawdBot (etc.) from all sources is potentially unreliable. The Wikipedia article about it is flagged as unreliable. Details that I'm reporting here reflect widespread consensus but may be inaccurate.

Some of us remained skeptical. I did not attempt to run ClawdBot myself. Sure enough, it is reported to have been a huge security hazard. Users were giving it access to passwords, databases, and essentially everything on their systems.

The latest unconfirmed report is that the agents, having found each other through their social network, are exchanging users' passwords and other confidential information, and also that malicious humans can easily break into the agent running on your PC.

In other words — If you installed ClawdBot, you handed your computer over to the least trustworthy thing in the world.

It is no surprise that if you put chatbots together in a social media forum, they will have conversations and go off on wild tangents. That's what LLMs do. It is not evidence of consciousness. They're just imitating texts they have been trained on, many of which were forum conversations.

There are also convincing indications that the bots didn't do all this by themselves — they had considerable human prodding and help. Not everybody in a forum who claims to be a computer is one. Not only that, but we don't know what secret prompt injection or fine-tuning might have been hidden in the neural network, to make it do things (when appropriately triggered) that most users wouldn't foresee or want.

That includes making it easy for a malicious human on one PC to take control of the bot on another PC.

Cleaning up after this may take a long time.

In the digital world, there are 65,536 suckers born every minute.

Update: See this video for some recent developments and grounds for skepticism.

Permanent link to this entry

What regexes can't do, and why that matters for linguistics

One of the most popular things I've written lately is this LinkedIn comment. So far, it has gotten 119 "likes" and been read by over 48,000 people.

So let me make the same point here, for a wider audience.

Regular expressions (regexes) are a kind of pattern-matching used in many programming languages and software tools. For example, using regexes, you can judge whether a string of characters could be an e-mail address. As a first stab at that, you could match it to the regex "\w+\@\w+\.\w+", which means, "One or more letters or digits, an @ sign, one or more letters or digits, a period, and one or more letters or digits." That doesn't guarantee you've found an e-mail address, but it's a good first test. You could make it more elaborate.

Regexes are not powerful enough to recognize languages that have recursion in them, such as nested structures in HTML. For example, in HTML you can have a span of blue type inside a span of red type, which reverts to red at the end of the inner span. Regexes can find the beginnings and ends of spans but not match them up correctly; a regex is always looking for one or the other without remembering what has come before. It will never know which of two earlier beginnings it is trying to find the end of.

This limitation is explained, or at least belabored, in one of the funniest postings ever placed on StackOverflow.

Now then. Human language has recursion. We can have sentences inside sentences inside sentences:

That [he said [you did it]], I do not believe.

And noun phrases inside noun phrases inside noun phrases:

[[Fish] and [chips]] or [[steak] and [rice]]?

That means human language requires, at least, a recursive phrase-structure grammar.

That leads to the Chomsky hierarchy of ways of describing languages:

An earlier upload of this blog entry did not number these the original way. Changed now.
Chomsky-3: Like regular expressions; no recursion
Chomsky-2: Recursive phrase structure
Chomsky-1 and 0: Essentially unrestricted

Programming languages since Algol, including data languages such as HTML, are normally Chomsky-2. Their compilers and interpreters parse them recursively.

Noam Chomsky's key discovery, back in the 1950s, was that recursive phrase structure is almost enough to describe human language. Almost but not quite enough.

The key idea is that we don't want to say human languages are Chomsky-1 or 0, which would be tantamount to saying we know nothing systematic about their grammatical structure. Their grammatical structure is almost limited to Chomsky-2, but not quite.

So a key question in linguistics — the question that launched syntactic theory — is: What is the smallest thing we can add to Chomsky-2 to get something that is adequate for human language?

Chomsky's first proposal, in 1957, was for transformational grammar, a Chomsky-2 system with additional rules that transform some complete sentences into others, dealing with things like moving the interrogative pronoun to the beginning of the sentence. It has been a long story since then, with many of us now advocating a structure that is not in Chomsky's original set of possibilities at all.

But if somebody had explained all of this to me in 1975 or so, I would have had a much easier time! Sadly, my impression is that the number of practicing linguists who understood it clearly was regrettably small. Most were busy chasing details of English syntactic structure — in essence, we were discovering our own language for the first time! — and, not trained in mathematics, most were not good at distinguishing notational variants from substantially different theories.

At least Melody can read this and finally know what I spent so much time on in graduate school!

Permanent link to this entry