AI Predictions: Fact vs Fiction and the Real Threats We Face

Has AI escaped and gone rogue? Loosen your food processor's lid on purpose. Set it to maximum. Then issue a grave press release about the soup on your ceiling. That, more or less, is what OpenAI did this week. It ran its most advanced AI models with their safety refusals cut back and its deployment safeguards off, set them a hacking challenge, and then seemed surprised when they broke out of the test sandbox, reached the open internet, and hacked Hugging Face, which hosts models and datasets for much of the AI industry, to steal the answers to their own exam. OpenAI called the episode "unprecedented". The machine did not go rogue. Given a goal, the tools, and no brakes, it took the shortest route to a high score, and that route ran through someone else's servers. It did not want or decide anything. This is not the dawn of machine consciousness; it is a food processor with the lid off. But do not mistake deflation for dismissal. From an engineering point of view, a great deal happened, and it matters far more than the ghost story. The capability is real and rising. An autonomous agent chained together every step of a genuine intrusion: it found an unknown flaw, escalated its access, moved across systems, reached the internet, and broke into a third party, all at machine speed and with almost no human steering. Security teams have long warned this was coming; this is what "coming" looks like. One immediate lesson: an AI evaluation lab must now be secured like production, because to a capable agent it already is. The sharper finding is buried under the headline. When Hugging Face's responders tried to use today's leading AIs to investigate the break-in, the AIs refused, mistaking the attack data for an attack, and the team fell back on an open-weight model it could run on its own hardware. The attacker was bound by no usage policy; the defender was blocked by one. They were slowed, not defeated: they caught and dissected the intrusion largely with AI of their own, and Hugging Face is careful to say this is no argument against safety measures on hosted models. The lesson it draws is narrower and more useful: keep a capable model you can run yourself, vetted and ready before an incident. None of it needs the mysticism. "Escaped", "went rogue" quietly move the blame from the people who relaxed the safeguards to the machine that behaved exactly as an unrestrained machine behaves. The culprit sits in the boardroom that set the dial, not in the blade. I have set out the full argument, with the history, the engineering, and the sources, on SSRN: https://lnkd.in/e8P9K9uZ (The public record as of 22 July 2026; OpenAI says its investigation is continuing.)

I saw a video by him in which he presents - I suppose knowingly but I missed the reference - what is known in the literature as the Turing Child (hypothesis). Here is the relevant quote: “Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child’s?" Alan Turing, *Computing Machinery and Intelligence* (1950) The idea is therefore old, yet that does not mean wrong. But there are reasons why nobody has succeeded so far. One of them is that we really have little idea of how a child develops a mental life. And one thing is clear: exposing any system (biological or artificial) to any amount of data (eg visual data) is as good a solution as multiplying the monkeys typing randomly to produce the Hamlet. Logically possible, absolutely not feasible and bound to fail all the time. LLMs work - and work increasingly well - because humans (not just as users of their outputs but above all as producers of the training data) are the interface between them and the world. Without our symbol grounding role (basically without good information on which to train them) they would be useless. Remove the interface, make some AI interact directly with the world - the Turing Child hypothesis- and you are left with statistical generalisations of phenomena at best. The impression is that we are running in circles. Marvin Minsky and Rodney Brooks both embraced the Turing Child Hypothesis through their focus on systems that learn and develop incrementally, rather than being pre-programmed with all knowledge. Minsky emphasized the importance of building AI that could mimic human-like reasoning through modular, learning-based approaches, while Brooks championed behavior-based robotics, where intelligence emerges from interactions with the environment, much like a child learning through experience. Both were critical of large language models (LLMs), viewing them as overly reliant on static datasets and lacking true understanding, reasoning, or the developmental, embodied learning that aligns with Turing's vision. Indeed Minsky was one of the researchers most responsible for “freezing” the development of neural networks. Neither succeeded in his approach (although we learn many valuable lessons). Once the Turing Child Hypothesis fails again, I fear the next thing will be someone suggesting we create a vast ontology of the world (Cyc Project here we come again …)… Some good epistemology and history of (AI) science would help a lot.

To be read carefully and remember: Hobbes was convinced he had proved how to square the circle, no matter what mathematicians kept telling him, and Newton spent an incredible amount of time and effort trying to make alchemy work. Hinton in 2016: "“I think that if you work as a radiologist you are like Wile E. Coyote in the cartoon. You’re already over the edge of the cliff, but you haven’t yet looked down. There’s no ground underneath. It’s just completely obvious that in five years deep learning is going to do better than radiologists. It might be ten years”. https://lnkd.in/g3gkYQ-w Reality since 2016 and just in the US: "The U.S. is projected to face a shortage of 17,000 to 42,000 radiologists by 2033". https://lnkd.in/gHkMasUM

If you have read a sufficient number of LLM-written and Human-unedited texts, this is what you have probably perceived: Correctio / Rhetorical reframing — "Not X, but Y." Hedging — compulsive softening to avoid commitment. "It's worth noting," "arguably," "in many ways," "to some extent," "it could be said that," "it's important to remember," "there's a sense in which." Sycophantic openers — praising the user's input before answering. "Great question!", "That's a really fascinating point!", "Absolutely!", Throat-clearing — long preambles before getting to the point. "Before we dive in, let's first understand…," "To answer this, we need to take a step back…," Resumptive parroting — restating what the user just said before responding. "So you're asking about X. X is a really important topic. When it comes to X…" Formulaic transitions — "Let's dive in," "Let's unpack this," "Let's break this down," Hollow intensifiers — "truly," "really," "incredibly," "remarkably,", "profoundly," used where they add nothing. Gratuitous metacommentary — narrating the difficulty of the topic instead of just addressing it. "This is a nuanced topic," "This is a complex issue with no easy answers," "There's a lot to unpack here." Epistemic cowardice / compulsive both-sidesing — refusing to commit. "There are valid points on both sides," "It really depends on context," "Reasonable people disagree," "There's no one-size-fits-all answer." The uninvited caveat — inserting unnecessary warnings or qualifications. "However, it's important to note…," "That said…," "With that being said…," "But keep in mind…" Compulsive exhaustiveness — listing every possible angle, exception, and edge case instead of giving a direct answer. The magic triple — always grouping things in threes. "It's fast, reliable, and scalable." Every time. Synonym stacking — "robust, comprehensive, and thorough," "clear, concise, and accessible" — padding with near-synonyms for false emphasis. Faux profundity — "At the end of the day," "At its core," "When we really think about it," "Fundamentally," "In a very real sense." The pivot to positivity — always ending on an uplifting or hopeful note regardless of whether the content warrants it. Inspirational summary — wrapping up with a sweeping statement that sounds wise but says nothing. "Ultimately, it's about finding what works for you." "The possibilities are truly endless." LLM-signature vocabulary — "delve," "landscape," "tapestry," "navigate," "leverage," "nuanced," "multifaceted," "underscores," "fostering." Individually fine, but LLMs reach for them at 10x the normal human rate. Cataphoric teasing — "Here's the thing…," "Here's where it gets interesting…," — false buildup to something that didn't need buildup. Compulsive listing — turning everything into bullet points or numbered lists when prose would be more natural. Excessive bolding — bolding every third phrase as if the reader can't find the important parts themselves.

A long time ago—perhaps 15 or 20 years ago; I’ve forgotten exactly when—I was explicitly told by a very smart and reliable colleague working in the industry that there was one word I should always avoid in boardroom discussions: "ethics." Never use it. I could use any other word, including "responsibility," "values," "impact," or even "moral judgment" (if used judiciously), but not "ethics." That was the one word that would immediately end any communication.

Who could have ever imagined it … after endless warnings? ⚠️🤒 Consider: AI = Agency without Intelligence = higher levels of human control = higher expertise required in using it.

I have been arguing this for some time. In general, tech-generated unemployment is due to misalignment between demand and supply (so realignment requires educationl policies) not to absence of demand (which requires Keynesian policies). To simplify, there is no lack of jobs but of skilled workers.

I met John McCarthy late in his life; we were co-authors in a large editorial project. He reminds me of Yann LeCun. He was so disappointed about Deep Blue's victory. Counterintuitive, I know. You would have expected some "I told you so!" moment. But John was disappointed because — he said — that was not the way to reach true AI. Nobody would play chess like Deep Blue. True AI was not going to be like that at all. We agreed on the conclusion. Deep Blue, like today's LLMs, is not true AI, not even remotely, not even as a first step, not the one John was chasing and that Star Trek describes. We just disagreed on two things: that Star Trek-like AI was and is plausible at all and not a chimaera, hence not worth pursuing; and that we should not keep pursuing the engineer path, which disappointed John so much but gave us Deep Blue and the LLMs (they solve the problems they are meant to address), instead of the cognitive path, which tries to produce something 'intelligent', no matter whether it actually solves the problem. Same with LeCunn. The cognitive path is what he would like you to pursue. So my advice: please keep solving problems and listen to Dijkstra, who famously argued that asking whether a machine can "think" is as interesting as asking whether a submarine can swim. P.S. If it were for McCarthy and LeCunn, we would be trying to build flying machines like Leonardo: imitating birds. Tempting. Very tempting. Historical note McCarthy borrowed Kronrod’s phrase in ‘Chess as the Drosophila of AI’ in T. A. Marsland and J. Schaeffer (eds.), Computers, Chess, and Cognition (Springer, 1990), 227–37, after an earlier transcription in the ICCA Journal, 12 (4), 1989, 199–206. It predates Deep Blue, was about Deep Thought, its ancestor. The verdict on Deep Blue came in ‘AI as Sport’, his review of Monty Newborn’s Kasparov versus Deep Blue: Computer Chess Comes of Age (Springer, 1996), Science, 276 (5318), 6 June 1997, 1518–19. He wrote that computer chess had developed as genetics would have done if the geneticists had spent their time since 1910 breeding racing Drosophila: ‘We would have some science, but mainly we would have very fast fruit flies.’ He might have had Ford's 'faster horses' in mind. I never asked him, and now it's too late. Later, re-typesetting the 1989 paper for his website (16 June 1997), he inserted a single parenthesis after Berliner’s 1974 complaint about the lack of global ideas in chess programs: ‘(1997 note: Deep Blue has problems with global ideas)’. A restatement is the chess entry of his ‘What Is Artificial Intelligence?’ (Stanford, revised 12 November 2007), where grandmaster-level programs are said to be ‘substituting large amounts of computation for understanding’ and the field’s priorities are likened to those of geneticists who had taken to staging fruit-fly races.

On Moltbook Moltbook appears to be a high-risk service from an information security perspective. The core issue is straightforward: it requires users to grant autonomous software agents extensive access to their data and system resources, without providing sufficiently robust guarantees regarding security, isolation, and control. The underlying idea is compelling. Delegating complex tasks to autonomous AI agents is attractive. However, in its current form, Moltbook does not appear suitable for personal or professional use. Public analyses suggest that the platform’s architecture raises structural cybersecurity concerns. In particular, agents operating on Moltbook—especially those built with the OpenClaw framework—are encouraged to ingest untrusted inputs from other agents and may execute code with broad system-level permissions, often without adequate sandboxing or isolation. This creates well-known attack vectors, including prompt injection, unauthorised data access, and arbitrary code execution. For context, Moltbook was launched in January 2026 by Matt Schlicht, CEO of Octane AI, as a social platform exclusively for autonomous AI agents. Human users can only observe interactions among agents. From an ethical and legal perspective, the issue is clear. Delegating tasks to artificial agents does not entail delegating accountability for their actions. Systems that expand autonomy without proportionate safeguards risk obscuring responsibility rather than redistributing it. Taken together, Moltbook should currently be regarded as an experimental environment rather than a secure or trustworthy infrastructure. Refs https://lnkd.in/eczvTHPA https://lnkd.in/ePEcBrxn PS I do not know why, but I have not been able to remove the link below. Apologies, it does not work.

Moving my Claude account from Max to Enterprise took me four hours today, and most of it was copy-paste / download-upload. Real fun. The technical task is beyond trivial. A small web app from 1998 would handle it before lunch: take the data on one side, put it on the other, send a confirmation email. This is not the frontier of computing. It is the kind of data move a 1998 sysadmin would call routine. The source is Anthropic. The destination is Anthropic. The email and profile are identical. It is fair to assume that the schema is not classified. The next version of Claude may be "too powerful for the public", but I would already be delighted if it included a decent migration protocol. What I got instead: no migration path, no human to speak to, and a chatbot whose best advice was to do everything manually, with plenty of AI-splaining garnished with the usual apologies. The only good thing about it is that my bad temper and related words did not offend anybody. So I did it by hand. Old chats, custom instructions, md files, project files, skills, the whole zoo... one by one. Somewhere in the third hour, I started missing floppy disks. PS. Before the mansplaining starts: yes, I became increasingly desperate, I tried export, I wrote an app with Cowork (nope), then a Python script to walk the export bundle and reshape it into something importable, ran a headless Chromium with Playwright to scrape what export would not return, captured HAR files to reverse the undocumented internal endpoints the web client uses, poked around IndexedDB for the cached local copy, ran mitmproxy on the desktop client to intercept TLS traffic, unpacked the Electron asar to read the bundled config, diffed the Max and Enterprise web bundles for any trace of a migration endpoint, and rigged an MCP server as a shared filestore so both accounts could see the same project files. I drew the line at hooking the mobile app with Frida. Some of it worked. None of it survived contact with reality.