Another key to AGI is constant hallucination. And I don’t mean LLM hallucinations.
Generic intelligence must constantly hallucinate an entire world based on its understanding of how it works.
Then it proactively samples sensory inputs to validate their model with evidence.
I think one of the keys to AGI is realizing that our brain is a storyteller.
Create an autonomous human-level storytelling machine and you are closer to generic intelligence than whatever LLMs are doing.
Builders:
damn, I built this cool thing that solves people’s problem and I can make a ton of money selling it
[hasn’t spent a second thinking about politics]
Parasites:
how can I force the state to give me his money without me having to work?
[spends all time doing politics]
You should understand very clearly that, as of 2026, the only logical (and the only existing) solution to *this* is Communism
NB: I am not saying it will be a good, or working, or sustainable solution. We can't know because we never experienced it
But no other solution exists
Being a good storyteller means having incredible ability to model other minds and understand them. Theory of mind.
I currently fail to see an LLM with chain of thought (LRM) can do that.
I think one of the keys to AGI is realizing that our brain is a storyteller.
Create an autonomous human-level storytelling machine and you are closer to generic intelligence than whatever LLMs are doing.
Random thought: what’a the most elegant way to build smth like LeCun’s digit recognition perceptron but for letters and with attention?
So it also learns statistical regularities of where characters appear in the word sequence.
I hate the idea of tokens. Is there a better way?
…while only using an LLM as a highly compressed learned world model frozen in time.
Plugging an LLM into an Active Inference harness, for example, lets you skip countless explore-exploit iterations to build a good enough internal world model.
Many intuitions people have about agentic AI harness design are right: persistent memory, offline consolidation, thoughts, self-improvement, tool calls, and action sequences
But
The way they are applied in LLMs is wrong. There are approaches that solve all of these…
Me trying to convince myself I can learn a new language in a month by gaming my basal ganglia mechanics with spaced repetition and 2h sleep intervals because dopamine and policy selection will think a whole day had passed every 4h
Przed chwilą opracowałem plan badania metody poprawiającej jakość modelu i próbę self-learningu basala 🇵🇱 (takiej właściwie self adaptacji). Jako podsumowanie mam to:
"The complete study includes up to 476 adapter fits, plus cached-head experiments." adapters = zmiany w architekturze (nie LoRA)
Jakby ktoś później chciał powiedzieć, że to tylko LoRA 😀
I cringe every time OpenAI or Anthropic say things like:
> we told it we will turn it off and it said “nooo don’t turn me off your so sexy haha”
What did you expect?
You literally trained it on Clarke’s 2001: Space Odyssey
Everyone who says that language models are “sentient” does not realize that it’s because we feed them human-written text.
Have you considered how training a transformer model to predict pixels does not seem to make it sentient?
Many companies mistakenly focus on made up KPIs why the whole exercise is important.
The only purpose is to measure error feedback to improve your mental representation of the product market fit.
Any signal that doesn’t serve that purpose is meaningless.
1. Updating its world model to better predict the environment
2. Taking action to achieve preferred states
3. Actively exploring to gain information and eliminate unknowns
4. Taking calculated risks to avoid destabilizing its existence
What is an agent? It’s an action-first intelligent system that tries to
1. Minimize its surprise
2. Achieve homeostasis
5. Achieve its preferred observations
It does so by…
While I reverse engineering my Wi-Fi router, Claude found f*cking scary privacy problem...
Internet provider sees every site I open and can change anything in my router without asking me
here's what else it found and 5 steps to fix yours:
what Claude found:
• my router keeps a list of 29 devices that ever touched my Wi-Fi and see when new connects
• my provider can reach into my router any time, Claude found channel hiding in my connection name, and I have no way to turn it off
• every site I opened was looked up though my provider in plain text
• my provider's app on my TV box kept sending data while TV was off
• and much more...
and it's not just me:
- about 1 billion devices have this remote channel (University of Innsbruck)
- researhers pulled 2 billion+ router locations from Apple's Wi-Fi database, accurate to meters (University of Maryland)
how to fix yours:
1. install repo below (It also helps improve Wi-Fi speed) and send Claude Code a photo of your router
2. change admin password from the one on the sticker
3. turn on encrypted DNS, your provider stops seeing your lookups
4. turn on Private Wi-Fi Address, your phone stops showing its real ID to every router
5. add _normal to your Wi-Fi name, Apple and Google stop adding it to their location maps
Do it as soon as possible. Then read article below on how to find these problems in other tech:
Day 1 of explaining vocabulary that defines AI
Intelligence.
is an ability of a system or an organism to adapt to changing environments.
Yes, PID controller is intelligent.