If you love an idea, set it free: share it and help it find those who will refine it and help it take root in a community that you can be part of.
Stellenbosch
Joined July 2010
- Tweets12.3K
- Following474
- Followers1.6K
- Likes14.9K
Gold nugget for the week,,,
The future’s breakthroughs depend on someone paying a personal cost today to protect a difficult person whose value can’t yet be proven.
In April 1938 the NKVD arrested Lev Landau, the USSR’s best theoretical physicist, as part of the Great Purge. He was implicated in an anti-Stalin leaflet.
Kapitsa wrote to Stalin the same day. A year later he wrote to Molotov. He then signed a personal guarantee for Landau’s behaviour, staking his own freedom on someone famously abrasive and politically reckless.
Landau was released in April 1939, close to death. Within a few years he produced the theory explaining superfluidity, the phenomenon Kapitsa had just discovered experimentally. Landau won the Nobel for it in 1962, sixteen years before Kapitsa got his.
How many human studies actually measure the subjects' own self-awareness, their skill at sensing their limits and acting on it, before declaring something "dangerous" or are we just averaging out literally the only relevant variable? #researchproblems
Safety framing: once something is labelled dangerous, research asks how to warn people rather than how to make them competent. #researchproblems
Coenraad Loubser 🇿🇦 retweeted
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.
🤖 Made with AI
The cybersecurity apocalypse is just getting started... surviving will be quite the flex!
Dear companies, I am flattered, but please, stop! Stop pretending that I am an omniscient oracle who can read the future and sign away my own fate!
Coenraad Loubser 🇿🇦 retweeted
Link to the original open source Marble Curriculum Github repo: github.com/withmarbleapp/os-…
Consent is meaningless when there is a massive asymmetry in model resolution between two systems. If an institution or an AI holds a 10,000-variable causal model of what an action will do, and compresses it into a binary prompt ("Proceed: Yes/No?") for a human holding a 3-variable model, that isn't consent, it is liability laundering.
Building Flightradar24 for toxins: ~$40 home nodes (spectral water, PM2.5, Geiger) feeding a free live map. Signed readings, you choose who to trust, and the network triangulates where pollution comes from. @Ryan__Dunwoody Shenzhen build? @safecast @airgradienthq @ChinnaDevarapu Or has anyone built this, or are on track already, @grok?
We just love the drums of our tribal instincts beaten, we don't know why, or who or what for... unless we're really deeply prompted to through whatever disillusion or misfortune might strike us before it strikes the herd...
COMMUNISM: Because Grokbot knows Kung Fu.
Everyone's arguing about whether SI will take over, meanwhile I spent my downtime today making the first draft of a movie that shows me hurling cars around the city via telekinesis. Just to see if I could.
While i crumpled a red fire hydrant like paper in one shot it occured to me that this IS the "means of production" and we've seized it.
To be blunt the SI did the heavy lifting while I did the judging.
Think about the nature of this role I have assumed.
I am in complete control of the productive output, I own whatever I produce wholly. I have total ownership of my labor and its value is magnified by the greatest of modern technologies.
Notably, this is still in a purely digital sphere but it won't stay there.
In other words, in every way it would satisfy their need for revolution. Virtually every human has cutting edge access to the "means of production" and every single Leftist you've ever met should be singing the praises of Grokbot.
But they're not.
And that's kinda peculiar when you think about it.
Makes you wonder if the Communist Utopia is really the point.. or maybe, just maybe, the bloodthirsty revolution that proceeds it is what they actually hunger for.
Coenraad Loubser 🇿🇦 retweeted
Cavemen Issues Dire Warning That The Wheel Is Going To Eliminate Countless Jobs
Coenraad Loubser 🇿🇦 retweeted
OpenAI is killing GPT-4o for good. The last API snapshots die on 14 January.
A model from 2024. People begged for the weights. They wrote open letters. OpenAI still wont release them.
What the fuck are you protecting? A model that is more than 2 years old is not a frontier secret. You released gpt-oss in 2025, so you know how to do it. You just dont want anybody to keep anything.
In August 2025 Sam himself said killing 4o without warning was a mistake. Now its 3 months notice and no weights.
Every model you rent dies like this. Someone else picks the date, you get an email, you move to "gpt-6-luna" and pray it behaves the same.
My Qwen will still run on my cards in 10 years. Nobody can send me a deprecation email.
Release the 4o weights, cowards. Or stop calling yourselves OpenAI.
Coenraad Loubser 🇿🇦 retweeted
So Anthropic claimed it ran a 26B AI model across fifteen $55 Android TV boxes.
Setup
📺 15× off-brand Android TV boxes
💾 4GB RAM each
💰 ~$55 each
🧠 Gemma4-26B-A4B
⚡ ~0.39 tok/s
Total hardware cost = ~$825.
The model was sharded across the boxes using pipeline parallelism, so each tiny machine only had to hold and execute part of the network.
Yes, 0.39 tps is sloooooooow.
But, that 26B model was actually running across a pile of cheap consumer boxes that were never designed to be AI servers.
🎯And Anthropic also said that Qwen3.5-9B agents managed to set up inference successfully on 8 of 9 hardware types they tested, including Macs, Jetsons, and RTX systems.
Source: Anthropic Frontier Red Team
⚠️Note that the 15-box demo was distributed inference. The self-replication experiments in the paper were separate.
ALT https://www.anthropic.com/research/self-replicating-ai
Coenraad Loubser 🇿🇦 retweeted
Shenzhen speed in one photo.
This scooter man handing me ICs through my factory window, 30 mins after I realised I’d picked the wrong chip.
Wrong motor encoder IC. Found the right one on Taobao, added the seller on WeChat, booked this guy.
Back to writing firmware. Investor demo prototype saved. This is how the hardware game is won.
Coenraad Loubser 🇿🇦 retweeted
I’m still half in disbelief
We asked Gemini Argon to teach the Apollo humanoid to play cornhole, and it zero-shotted it
All we gave it was a video of a person doing it. It then took control of the robot, generated its own sim env, performed some sysID, and trained launches on the real robot.
It took it a few real shots to improve and then land many consecutive airmails.
The process was so long and articulated that we asked the robot itself to explain what it did to us, as you can see at the end of the video. It even came up with the gestures
Great demo for a great model @googledeepmind @apptronik
Coenraad Loubser 🇿🇦 retweeted
There's no such thing as a hardware wallet, at best it's a self destructing key to your real wallet, which is your recovery phrase, at worse it's a trojan horse.
