Phillip de Wet
banner
phillipdewet.bsky.social
Phillip de Wet
@phillipdewet.bsky.social
Editor, columnist, ex News24, ex Business Insider SA, ex Mail & Guardian, ex Daily Maverick.

Living in Canterbury, writing about geopolitics, thinking about AI, advocating risk-adjusted writing.
Voice is still expensive and complicated compared to text. But I think the tech space in South Africa is missing out on finally cracking aliterate users.
And voice-first should be very politically palatable.

www.itweb.co.za/article/oped...
OpEd: In SA, you should be thinking voice-led agentic
If you're going agentic for anything public-facing, please build voice into the workflow, writes ITWeb contributor Phillip de Wet.
www.itweb.co.za
November 21, 2025 at 9:19 AM
"I don't do drugs, but if the 'just say no' campaign was used on me, maybe I would."
- (Sadly a Chatham Rule conversation, so I can't attribute.)
November 19, 2025 at 6:21 PM
Hating on "AI" (in reference to some stupid video-clip generator) is starting to irk me approximately as much as complaining about "the media" (in reference to one outlet).

Do you meant the AI that might solve antibacterial resistance? That is reducing grid energy loss? Then be more specific.
November 19, 2025 at 10:21 AM
"...indicators that many [thieves] are only interested in Apple iPhones."

iPhones hold value better, sure. But given Apple's tight control of both hardware and software, there's a case to be answered here: where is the vendor? Is it tracking the data, and making a plan?
November 18, 2025 at 8:56 AM
Every LLM I use does much better when asked for "succinct" rather than "short" or "brief".

This, in my experience, is also true of human writers of every kind.

Interesting work to be done, here.
November 12, 2025 at 3:10 PM
Not sure if this now makes me a royalist. But looking at the British royal family's place in UK society recently, I find myself arguing that other countries can do with a bit of that. As long as the costs are strictly contained.

www.news24.com/opinions/col...
Phillip de Wet | The British royal family is bad value for money, but still worth emulating | News24
Between kicking out Andrew and simply being there when one man with a knife terrorised Britain, the royal family has partly paid its way, writes Phillip de Wet.
www.news24.com
November 10, 2025 at 11:51 AM
Google and OpenAI effectively have infinite money. Apple and Anthropic and Meta aren't far behind. I'm deep into the Google ecosystem, with a bit of apple. Yet, as a wildly promiscuous LLM user, Perplexity is still my go-to answer engine.

Starting to think we don't need to worry about monoculture.
November 7, 2025 at 8:51 AM
Mobile phones begat battery development that massively accelerated renewable energy.

There seems a decent chance AI will do the same for carbon capture, because data centres that can't wait turn to LNG – but can't defend doing so. And have, effectively, infinite money to address that tension.
November 5, 2025 at 10:12 AM
Remarkable to listen to analysts talk about Donald Trump's statements (in this case on Nigeria) with the same language Christians use about the Old Testament: it's true but don't take it literally, it meets the morality of the audience, influences through description though it sounds like command.
November 2, 2025 at 4:50 PM
All the noise is about AI. But it is striking how 2025 is becoming the year in which the IT sector comes to grips with the encryption disaster that is quantum computing.

This week: Cloudflare says it is about to start experimenting with a neat way to prevent server impersonation.
October 30, 2025 at 1:32 PM
One way to think about it: like any good employee, LLMs works towards their metrics – and if you have bad metrics, you get bad output from, eventually, bad employees.

Some good stuff happening in benchmarking, though, which could help significantly.

www.science.org/content/arti...
AI hallucinates because it’s trained to fake answers it doesn’t know
Teaching chatbots to say “I don’t know” could curb hallucinations. It could also break AI’s business model
www.science.org
October 29, 2025 at 10:33 AM
Tactical proposal for countries with low birth rates: get younger people into family-friendly social spaces before 9AM Sundays.

Instead of sleeping off Saturday night, get late-20-year-olds where the toddlers – fresh from a good sleep, managed by parents not trying to juggle work – are being cute.
October 26, 2025 at 10:12 AM
Gemini is the one top-end system I effectively get for free, and until this week it ranked way, waaaaayyyyyy below Perplexity and ChatGPT in use.
Then integration with Google Docs got good, and suddenly I'm using it to answer a lot of questions about my work and financials I never thought to ask.
October 24, 2025 at 1:27 PM
First time back in Manhattan for a couple of years after hanging around Europe, and I've figured out what feels wrong: too few bicycles.

Elsewhere they have grown into predatory flocks you had better keep an eye on if you like your limbs unbroken. Here they are few, and furtive.
October 14, 2025 at 12:05 PM
"I lost propulsion while going at low speed" is the good news from the Jeep software update – over the air, just needed drivers to click "yes" – that bricked vehicles mid-drive.

www.thestack.technology/jeep-softwar...
Jeep software update bricks vehicles
“Please exercise extreme caution”
www.thestack.technology
October 13, 2025 at 10:21 AM
The BoE and IMF warnings about the consequences of an AI bubble popping are necessary and important. The danger is real.

But, as someone who lived through dotcom, there is a fundamental difference: AI has huge revenue.

Customer value is in doubt. Cashflow is not.
October 9, 2025 at 9:43 AM
Honestly can't remember the last time I saw such an outstanding effort to keep failing at every turn – and to synergise the individual failings.

Deloitte in Australia
* didn't disclose the use of an LLM in an important report
* didn't use its own LLM
(cont)

www.thestack.technology/deloitte-use...
Deloitte used govt client's own GPT to invent references
Then failed to disclose it, took a month to fix errors publicly pointed out, and is apparently on the hook for a partial refund
www.thestack.technology
October 7, 2025 at 9:07 AM
Big news today (and even a some controversy) because the English govt thinks a system of home buying that is utterly batshit insane should maybe be tweaked slightly.

That the terms "gazump" and "gazunder" even exist beyond a medieval curiosity still blows my mind. But Brits are just ¯\_(ツ)_/¯
October 6, 2025 at 8:15 AM
"Strengthening nucleic acid biosecurity screening against generative protein design tools" in Science this week doesn't read like a disaster movie script only because it is written academically.

In the supplementary material, things get rather extremely concerning.
www.science.org/doi/10.1126/...
Strengthening nucleic acid biosecurity screening against generative protein design tools
Advances in artificial intelligence (AI)–assisted protein engineering are enabling breakthroughs in the life sciences but also introduce new biosecurity challenges. Synthesis of nucleic acids is a cho...
www.science.org
October 3, 2025 at 5:28 PM
Google is, with a straight face, trying to sell me Google AI Ultra for £235 ($315 / R5,450).

Per month.

I am locked into Google storage and hosting apps, but never touch their AI. You'd think the crack-dealer business model would be the way to go here, not the robber baron model.
October 3, 2025 at 7:01 AM
OpenAI is now valued at $500 billion, just as Elon Musk's personal net worth hits that same number.

Don't get me wrong, Musk has done some amazing (as well as some terrible) stuff. But the same value as the first-mover, big-brand, maybe winner-takes-all of AI?

Sheesh capitalism, get it together.
October 2, 2025 at 10:44 AM
In July 2018, Julius Malema fired a gun in public.
There was never any serious question that he did it.
It took a political lobby group – and 86 months – to convict him.
Sentencing is still months, maybe years, away.
October 1, 2025 at 9:27 AM
DeepSeek is doing some interesting experimenting – and some extremely interesting pricing – with its scarce attention update.

We don't yet know if they are upending the economics of LLM, but that's the direction of travel.

www.thestack.technology/new-deepseek...
New non-vanilla DeepSeek model rocks the boat again
"Significant efficiency improvements" claimed in long-context, alongside costs on par with Gemini Flash and gpt-5-nano.
www.thestack.technology
September 30, 2025 at 4:27 PM
My free-to-air column this week argues that Microsoft will win AI because I'm afraid of Comet.
Which, you know, might just be me. But I am yet to find anyone actually using an AI browser as primary – and I sometimes hang out with the deep geeks.

www.itweb.co.za/article/oped...
OpEd: AI is running into a trust wall
AI sellers promise shiny new capability, while buyers ask about security, auditability, lock-in and compliance.
www.itweb.co.za
September 30, 2025 at 8:58 AM
@stephenkb.bsky.social Your Memo for Labour made me go find “What did we learn on our summer holidays?” – and now I'm convinced the authors were talking about you, personally.

They tried to disguise it, but there's totally a picture of you pinned to a dart board somewhere.
September 30, 2025 at 8:56 AM