advancing the computational foundations of science
Toronto, Ontario
Joined November 2014
- Tweets619
- Following536
- Followers2.4K
- Likes5.3K
My favourite game now is to create games with Opus 5.5. Here is a board game I made; can play against AI or p2p multiplayer (no server, pure static web, free hosting) zaffer.github.io/near-peer/
2
144
Recreate exact analog video effects with this open source tool that implements the actual NTSC hardware pipeline in code
Analog-Vision: synetic-labs.github.io/Analo…
1
108
Wanted a plain and simple app for my kid to play with on my phone, so built this little finger painting game. Got a little carried away and now it is legit my favourite drawing / note taking tool: zaffer.github.io/plaint/
~22 zettabytes of digital data in the world.
Training on all of it would not unlocked meaningfully more intelligence.
We need to harvest reality for the most useful information.
Analogising data as 'the new oil' works if you understand that only new information is more data.
122
# why clicking a button does what you think
Explicit code does exactly what you say deterministically.
Probabilistic machine learning algorithms do approximately what you want.
In reality, clicking a button on a website does not, in fact, always result in the outcome you anticipate. This is because the complexity of the code effectively makes it probabilistic too.
The difference with attempted determinism is when things go wrong they can go catastrophically wrong. The tradeoff you get with fuzzy logic is more predictable errors within a distribution, by sacrificing the fact that there will be some errors throughout your application.
Since we are faced with software today that is getting increasingly sloppy, why not just move to fully probabilistic code for non-critical software?
The paradigms of software engineering are appearing to become "what amount of robustness do you need":
- partially robust but mostly does exactly what you want
- very robust but always only probably correct
- fully robust and predictable, but requires 10x more time to craft
You could even do a combination.
When you click a button to make a payment the clicking of the button itself could work with a probability of 50/50, but as long as the feedback is good you can just keep clicking until you see that it worked. The backend that actually makes the payment after confirmation should be robust and deterministic, but the interface only needs to be deterministic in communicating the outcome, not necessarily the user experience of instantiating it.
Perhaps all front-end complexity can just accept probabilism, and in a way LLM based agents are doing that now already. Asking an agent to do something is a conversation that sometimes requires back and forth to get the end outcome you wanted.
The amount of time you spend training the AI is what increases the probability of correctness. So the time spent refining applications moves from explicitly coding, to explicitly crafting reward functions. You also trade in generalisation of code via abstractions, to generalisation through diversity of training data.
Nature works like this already, when you push a physical button in the real world, there are many ways that it could have failed to register. When it looks like the button press did not work, our natural instinct is to immediately try pressing it again, and again, until something happens. So as software starts using more biological inspiration, so too does it probably start resembling physical interfaces we deal with in real life.
115
Now that AGI has arrived, I see two adaptions all software needs to make:
1. Agent access - embrace the fact that an agent can do anything a human can on a computer now, stop banning bots
2. Agentic authentication - human based signup, but secure access without human in the loop
160
The LLM judge testing hallucinations in an AI agent hallucinated that it caught a hallucination. We are at false positive hallucinations now...
1
170
People finally starting to talk about Nick Land. I have a feeling he will be remembered as the most important philosopher of the 21st century. His ideas are shockingly deeply influential in Silicon Valley, and this is where effectively all global technology is developed.
1
267