Guardrails don’t fail in code—they fail in adoption: if people don’t use the governed workflow, the guardrails never fire. Reliability collapses back to whatever the probabilistic generator does that day—which, by definition, won’t always be what you want. (3/3)
LLMs are probabilistic generative engines that need deterministic guardrails (automated constraints plus selective human oversight) to function reliably. If the flow isn’t rethought, the “AI gains” are mostly illusion: higher throughput of the same waste, plus new failure modes. (2/3)
Seen this movie before in ICT: automating the current mess just makes more mess faster. LLMs won’t improve processes by themselves —because they’re not digital employees; www.juicebox.com.au/insights/ai-... (1/3)
Picture of
Summer was fun and inspiring... let's get that new academic year started! 💪
“Bluesky may support a more interpretive, reflective mode of science communication.”
“Interactions on Bluesky were an order of magnitude higher than on X.”
“Bluesky may well become the next X for scientific discussion and will persist in the long term.”
Research confirms: Bluesky is a better home for academia! Posts about science on Bluesky are more original and get significantly more engagement than on X, according to a large-scale analysis reported by Nature.
A promising shift for science communication! www.nature.com/articles/d41...
Compliance is no longer optional… and change a certitude. Very nice summary of today’s cybersecurity reality.
I was putting together a PowerPoint slide with the headlines of major cyber incidents in our region from just the past two months, and I’ve already run out of space.
This has been a wild summer. Now a zero-day on Apple OS'es, CVSS score 7.8, "this issue may have been exploited in an extremely sophisticated attack against specific targeted individuals". Lovely.
thehackernews.com/2025/08/appl...
This is a much more objective, and sobering, assessment of the impact and risks...
www.theregister.com/2025/08/21/o...
The criticism of Orange for downplaying the risks of its incident, as noted in the article, is striking - especially since Orange also sells cybersecurity services to others. Hard to square that with the transparency, containment and mitigation NIS2 is supposed to drive.
Telecom operator Orange Belgium confirms a major data breach affecting 850,000 users. Personal data & PUK codes exposed. Subscribers warned to stay alert for scam messages further exploiting the breach.
www.vrt.be/vrtnws/nl/20...
It's a stochastic parrot. But this may challenge us to rethink what understanding really implies. If a purely statistical model consistently produces coherent, useful, context-aware answers… at what point do we just treat that (at least functionally) as understanding? And does it not have value?
Surprise surprise, commercial generic models aren’t the definitive solution to all problems. Specialized, open, and well-tuned beats big and generic. Curious to see how this shift will play out at the edge...
Fact is, LLMs learn only by predicting the next word—yet (some) reasoning and abstraction emerge. Maybe that’s a clue: language isn’t just a tool for thought, but a big part of how we perceive (and structure) the world. Not the whole story, but maybe more central than we guessed?
I had the Courier in those days... still remember flashing it to support X2 and getting the amazing 56k throughput (which was "almost as much as a digital 64k line!"). With the iconic "oink-oink" handshake noises at the start of each session...
That may sound blunt, but given the ongoing instability, it’s clear MITRE’s reliability as a cornerstone partner is compromised. That doesn’t mean they’ve lost all relevance, but sole dependence on MITRE for stewardship of key frameworks has now become risky.
"MITRE just terminated senior leadership at the Center for Threat-Informed Defense in a cost cutting exercise. It appears, at this early stage, MITRE can no longer be relied upon to be the steward of the back bone frameworks for global threat-informed cyber security."
We're closer to anthropomorphic misinterpretations than to Rise of the Machines.
Carnegie-Mellon research seems to confirm: LLMs often fail at long-horizon reasoning, social interactions, and tasks requiring genuine understanding. They perform noticeably better when tasks align with their training data, not when abstract or novel reasoning is required. arxiv.org/pdf/2412.14161
To me, this reads like we’re recreating Atlas Shrugged, but with venture capital and blockchains. Now it’s upgraded with crypto, AGI, network states, and Substack manifestos.
Exciting to see PHP adding one of my favourite Clojure features: function piping/threading. A big step forward for anyone with a soft spot for functional programming! Now let's get this finalized in JavaScript as well 👍
Just a gentle reminder that universities - even top-tier ones with significant operating budgets - are not safe from cyber threats and incidents. If it seems that "nothing ever happens" in your academic environment, that's because people are constantly hard at work to avoid it. bit.ly/3HhpibK
"Quite simply, 'LLMs are doing reasoning' is the 'look, my dog is smiling' of technology." Provocative article that certainly puts the AI/LLM hype into some sobering perspective. Though IMHO even then the Stochastic Parrot can have some value - when recognized as such.
It does have one great feature though - it'll still have security patches after October 14, 2025.
Discomforting conclusion: the global incident around shocking zero-day CVE-2025-53770 that took everyone by surprise… wasn’t surprising at all. Just a remix of a familiar pattern. And more like it? Not surprises, but inevitable.
Nothing new here. The real value of AI is in augmenting humans, not replacing them. LLMs require a human in the loop and verified tracebacks to be reliable — and that takes significant manual effort. No automated method today can guarantee accuracy in high-stakes use.
Het is niet zozeer een vlaag van zinsverbijstering maar eerder de toepassing van de wettelijke vereisten in de Codex Hoger Onderwijs, waarbij de instelling zich ook aan controle mag verwachten, en niemand gebaat is met complicaties. Zie o.a. www.ugent.be/nl/jobs/taal...
Edge computing on steroids: LLMs run locally on regular laptops, thanks to the rise of NPU-accelerated PCs. With AI hardware baked into new chips, running models like LLaMA or Mistral offline will become a private, practical way to handle your data.
Education tops the list with 31% of its public-facing assets vulnerable. This inherently open, dynamic environment demands understanding of asset ownership, purpose, context, and attacker perspective to manage exposure. That’s exactly what #NIS2 is pushing for. Security starts with context.
