Time and again, we’re told that AI hallucinates, that it’s unreliable, that it can’t write reliable code, nor could it replace meaningful jobs. And yet…

Satya Nadella, on April 29, in a conversation with Mark Zuckerberg at Meta’s inaugural LlamaCon AI developer event in Menlo Park, California: “I’d say maybe 20%, 30% of the code that is inside of our repos today and some of our projects are probably all written by software.”

Microsoft, on May 13, announced the laying off of 3% of its workforce, or about 6,800 out of its 228,000 employees. Or maybe slightly less than 3%, as many news outlets insist on the “about 6,000” figure. Either way, there’s a precise figure about the Redmond HQ: 1,985 people will have to go!

Microsoft didn’t mention AI specifically, but The Reg believes that middle managers are particularly vulnerable to automation. From the comments:

How can that happen? Software can’t play golf!

You can see them planning to axe managers from the Windows Update QA and security units, only to discover that no such units exist.

…

The tech sector seems to be believing its own propaganda on ‘AI’, and ignoring that paragraph on p.685 of the terms and conditions about it being experimental, often unreliable and just a bit of fun. If I had shares in a company that started to replace people with ‘AI’ I would sell them.

…

The only saving grace here is that, with W11, MS have already made Windows awful beyond usability. If you cannot switch to Linux, it may be time to switch to LTSC and hunker down as things really go pear-shaped. Maybe, after a few expensive catastrophes, they will jettison those responsible for the last few years of MS, writing it off as a bad trip, and flip back to having carbon based lifeforms secure an earlier version of Windows for everyone to use.

Obviously the best use for AI is to replace the entire C-suite

Think about it, the highest paid workers by thousands of percent with least amount of specialist domain knowledge. That’s why they can flit between top level roles so easily. The basics and even the minutia of good leadership are very well documented in books that have probably been consumed by AI and let’s face it, very few c-suite people in the tech world seem to do any of the things that the leadership doctorates recommend anyway.

🤖

I recently asked Grok (should I have tried with ChatGPT?) to “comb” and simplify a long CSS file, because it was too complex to be understood by Adobe Digital Editions 1.7.2. That’s the oldest ADE version for Windows that can still be found in the wild. It doesn’t cope well with EPUB3 nor with HTML5, but if an ePub file is properly displayed by it, it should be properly displayed by any decent ePub reader.

At first, Grok created a very neat new CSS file. The only problem was that all classes were different. How was I supposed to use the new class names in dozens of XHTML files? It just couldn’t figure out that I needed to use at least some of the same classes. It didn’t even try to tell me how to transition from one CSS to another!

When told about that, it simplified a bit the original CSS, but not by much. It was right: without moving some formatting in separate classes, and use 2–3 classes instead of one where appropriate, there wasn’t much it could do. Even so, it managed to break the CSS! The XHTML files just didn’t display the same anymore.

I’m still stunned that so many people rely on LLMs for writing code. I had horrendous experiences in the past with ChatGPT and Mistral when I used them to write Bash scripts and basic Python code. I had to debug, I had to correct, I had to adjust, I had to fix regressions, I had to come with so many follow-ups that I’d have been much better off if I didn’t use “AI” in the first place! How the fuck are people using AI for coding? OK, when you don’t even know what library or what class to use, or just how to approach something, LLMs can provide valuable pointers. This should also work as a smarter IntelliSense, or to generate simple snippets.

Somehow, for having rarely used LLMs for coding, I didn’t run into code hallucinations. More like general knowledge hallucinations, or other IT-related hallucinations, but not made-up libraries, functions, or classes. Recent reports that OpenAI’s models hallucinate even more than before are worrisome, but I have yet to meet them. For now, and not regarding the coding, Claude 3.7 Sonnet seems to me slightly less reliable than 3.5 Sonnet. But that’s anecdotical, not measured in any way.

🤖

But here’s a nice article in the NYT: A.I. Is Getting More Powerful, but Its Hallucinations Are Getting Worse. (Barrier-free here.)

I selected a few comments, and nothing from the article itself:

❶ LRM, PNW, May 5, Times Pick

I’ve been experiencing ridiculous hallucinations with GPT-4o, the latest publicly available model that I pay for. I thought it was just me, but I’m somehow both relieved and troubled to hear that it’s not.

There’s been multiple times I’ve pushed back on it saying I thought something was wrong, only for the model to apologize and change its answer to something else that’s also wrong. I’ve given up on it several times because I just don’t trust it. Like the article mentions, it does this most obviously on factual answers where there clearly is a right and wrong answer, but it makes me question all of its responses because it’s shown that it cannot be trusted.

❷ RWG, London, May 5, Times Pick

I used ChatGPT to help me compute some compound angles for splayed x-shaped table legs. The base was meant to be 9.5 inches on center, and the top—where it reaches the table top, is 8 inches on center. I wanted a half lap joint where the legs crossed. I asked it for all the angles and the distance from the floor to the center of the crossover. It asked for basic information to work with, which I provided.

It basically got almost everything wrong… It confused a miter angle with a bevel angle, and even said I needed to make a bevel cut at the crossover. It made no sense at all. It was very gracious when I corrected it, but I needed to sort out the angles myself, by somehow remembering my high school geometry and trigonometry.

❸ JK, Kansas City, May 5

I’m sorry, did this article just say that AI is “useful for writing term papers”? I am an English professor and let me tell you, it is not. I now regularly get papers with made-up quotations, vague meaningless slop, and reasoning that doesn’t match the (again, fake!) evidence. When will people realize these things cannot actually write!?

❹ Ohana, Bellevue, WA, May 5

I use ChatGPT every day. It’s hard to know how often it’s wrong or making stuff up, because I usually ask it things I don’t know. However, it regularly tells me things I know to be untrue, such as that the current president is Joe Biden. My experience is that it is wrong far more than 1% of the time. I rarely use the reasoning or deep research function – this is just what I’ve experienced with the standard interface.

It is great to code syntax, though. Worst case is that the code doesn’t compile.

❺ Max O, NM, May 5, Times Pick

Recently tried citation formatting with Gemini for a report “Gemini, can you provide an Oxford style citation for a footnotes for the resource at the following link?” The citations, despite linking the sources, were totally made up things that didn’t exist, an amalgam of similarish sources that return 0 results in google scholar.

❻ Brian, Chicago, May 5, Times Pick

No surprise, the projections from technologists are overblown. We’re a decade overdue for fully autonomous cars. I don’t think it will ever happen. The rate and risk of failure outweighs the benefit of full automation. This is true in many contexts.

Businesses may be able to use AI as an enhancement to productivity but they cannot allow full automation because someone has to be responsible for the finished deliverable. AI companies won’t take on the liability of errors. And telling a client “oops my AI screwed up and I didn’t proofread” will never be acceptable.

❼ Bronx Bob, Bronx, NY, May 5, Times Pick

The problem with living in the so-called Information Age is that so much of the “information” is rubbish. Cohorts of humans are growing up unable to distinguish verifiable information — in scientific, historical, etc. — from propaganda and lies. It bodes ill for democracy and public affairs of all sorts.

❽ Raj Valli, Kendall Park, NJ, May 5, Times Pick

This doesn’t surprise me at all.

As someone deeply involved in building AI systems for education, I’ve seen firsthand how even the most powerful LLMs can fabricate information with confidence. That’s why, for high-stakes domains like student learning, hallucination-prone AI is simply not acceptable.

It’s also why, at Thinkster, we’ve taken a very different approach. We use “explainable AI” that doesn’t just make a recommendation — it shows its work. Every instructional decision, skill progression, and suggested intervention is backed by a traceable learning model and can be audited by teachers, students, or parents.

When you’re trying to guarantee learning outcomes (not just chat), interpretability and trust are non-negotiable. AI hallucinations aren’t a “bug” — they’re a consequence of design choices. And in education, the wrong design has real consequences.

Would love to hear how others in the AI space are thinking about this.

❾ tabithatwinkles, California, May 5, Times Pick

I can assure you that AI is not useful for writing term papers. Any professor worth their salt can sniff out an AI-generated essay a mile away, number one, and if they care that the student did not write the words on the page (writing is thinking so a student who uses AI to not write, did not think either) then said professor will be pretty salty, indeed. Number two: do you have any idea how many glaring, obvious errors, mistakes, (call them hallucinations if you wish but I’m pretty sure you have to be alive to hallucinate) AI makes when writing? The level of ludicrous information I’ve seen in student essays due to ChatGPT is staggering.

One must have a high level of subject-knowledge mastery to begin to tell when ChatGPT is wrong, and I assure you it is wrong often. The issue is that student do not have said high level of subject-knowledge mastery and can’t determine between fact, fiction, accuracy, or downright misinformation. So please, can you stop with the, “it’s helping student” nonsense? It isn’t. I promise. I have been teaching for twenty years at R1 and state colleges and the last two years have been abysmal. I will say, I am not alone in my estimation, either. Virtually all my colleagues are beleaguered by falling student performance, most of which comes from them trying to cheat-code their way through everything.

🤖

Is there anything more to be said?

Oh, right, “AI” can be useful if specifically trained models, for very specific tasks, are used. Take this case, when Leela Chess Zero, using its best deep learning network, managed to beat all attempts at analyzing a game. In Leela Chess Zero’s spectacular Qxf5!! it found a sacrifice (queen exchanged for a pawn) at move 47 that eventually led to a drawn game at move 132, despite all the analysis evaluating the position as clearly winning for White.

But general-use AI chatbots are the supreme garbage, and even more so when you consider that most people using them are completely clueless. They can’t tell when they met a hallucination or a blatant error or misinformation. By the way, how can you detect whether a Wikipedia page has been vandalized in a subtle way? It can look very professional and plausible, bar for some parts of it. Can you detect that? Even by reading the versions of the same page in several languages (which you most certainly won’t do), are you sure you can trust what you read on Wikipedia?

I lost all my hope in humankind. We’re now facing stupid chatbots and AI agents when absolutely no one ever learned to cite Wikipedia by providing a link to a specific version of a page! This means to use, e.g., this instead of this.

Wikipedia doesn’t make it easy to link to a specific version of a page. The humankind failed to understand the dangers of Wikipedia (edited by fuck knows who, verified, policed, and censored by self-appointed “experts”). And we’re now supposed to exert judgment regarding the output of “AI” chatbots?

For fuck’s sake, we’re completely doomed!

🤖

Or maybe we’re vibe-fucked. The mania of coding with AI when you have no fucking clue what you’re doing is now called “vibe coding.” This is all that this cretinoid of Andrej Karpathy could come up with:

There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like “decrease the padding on the sidebar by half” because I’m too lazy to find it. I “Accept All” always, I don’t read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I’d have to really read through it for a while. Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away. It’s not too bad for throwaway weekend projects, but still quite amusing. I’m building a project or webapp, but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

It fucking doesn’t!

🤖

🪶 Many intellectuals wrote about the dangers of the technological advancements long before these became a real danger—or so it looks like from today’s perspective. And I’m thinking of the social alienation or even the decline of our civilization, not about nuclear war or ecology.

Let’s agree to ignore Oswald Spengler (Der Untergang des Abendlandes, 1918-1922; Der Mensch und die Technik, 1931) and José Ortega y Gasset’s sort of reply in La rebelión de las masas, 1930, in the chapter IX. Primitivismo y técnica. But then we had:

  • Martin Heidegger (Sein und Zeit, 1927; Die Frage nach der Technik, 1954)
  • Jacques Ellul (La Technique ou l’Enjeu du siècle, 1954; Le Système technicien, 1977; Le bluff technologique, 1988)
  • Roland Barthes (Mythologies, 1957) and, 50 years later, Nouvelles Mythologies, obviously not by him
  • Herbert Marcuse (One-Dimensional Man, 1964)
  • Lewis Mumford (The Myth of the Machine, 1967-1970; because Technics and Civilization, 1934, wasn’t that critical of technology)
  • Marshall McLuhan (The Mechanical Bride, 1951; The Gutenberg Galaxy 1962; Understanding Media, 1964; The Medium Is the Massage, 1967; The Global Village, 1989)

At some point, there was a focus on television, and here’s my quick selection of works:

  • Roland Barthes (La civilisation de l’image, in Mythologies, 1957)
  • Pierre Bourdieu (Sur la télévision, 1966)
  • Pier Paolo Pasolini (Aboliamo la tv e la scuola dell’obbligo, 1975; Scritti corsari, 1975; posthumous volumes: Lettere luterane, 1976; Saggi sulla politica e sulla società, 1999; Pasolini e la televisione, 2011)
  • Giovanni Sartori (Homo videns, 1997; with a second edition in 1998, and a new edition ©2000 that adds a 8-chapter appendix, and a subtitle: Televisione e post-pensiero)
  • Umberto Eco (Sulla televisione. Scritti 1956-2015, 2018)
  • Jacques Derrida (Échographies de la télévision, 1996)
  • Bernard Stiegler (La Télécratie contre la démocratie, 2006)

Then it came the Internet. And the social networks. This time, the disruption was so momentous that it’s impossible to even write about it. The cancer has metastasized. I’ll only quote from Umberto Eco.

In a speech in June 2015 at the University of Turin on the occasion of the award of the title of Doctor Honoris Causa (La Stampa, Il Giornale): “I social media danno diritto di parola a legioni di imbecilli che prima parlavano solo al bar dopo un bicchiere di vino, senza danneggiare la collettività. Venivano subito messi a tacere, mentre ora hanno lo stesso diritto di parola di un Premio Nobel. È l’invasione degli imbecilli.” Then he added: “Il dramma di Internet è che ha promosso lo scemo del villaggio a portatore di verità.”

In March 2015, in an interview for the Spanish El Mundo, the same Eco (El Mundo, El Mostrador): “Con Facebook y Twitter es la totalidad del público la que difunde opiniones e ideas. En el viejo periodismo, por muy asqueroso que fuese un periódico, había un control. Pero ahora todos los que habitan el planeta, incluyendo los locos y los idiotas, tienen derecho a la palabra pública. Hoy, en internet, su mensaje tiene la misma autoridad que el premio Nobel y el periodista riguroso.”

Now we have the “AI” invasion. And we’re so very fucked.

🤖

On May 18, Romanian voters have to decide who’s going to be their next president. The favorite is not Nicușor Dan, a mathematician with a PhD at Université Paris-XIII (Sorbonne Paris Nord), with a master at Université Paris XI (Paris-Saclay), and having studied at École Normale Supérieure. He’s also a two-time winner of the International Mathematical Olympiads (he secured gold medals in 1987 and 1988). Nope. The favorite is a fascist retard called George Simion, whose most recent feat is that, in a recent video interview with an oligophrenic influencer (🎞️ TikTok, 🎞️ Facebook), he could not say whether the Earth is flat or not. “Let the scientists decide on that.” (The question is asked at about 7:30 in the video.)

The future president of Romania believes that the Earth might be flat. And we’re trusting people to detect when a chatbot is hallucinating?