Quaternion Process Theory, Artificial (Intuition, Fluency, Empathy), Patterns for (Gen, LRM, Agentic, Skill) AI, intuitionmachine.com/
Arlington, VA
- Tweets 121,954
- Following 5,542
- Followers 63,887
- Likes 73,362
Sonnet 5.5 delivers with a bang! Here, it explains how it solves problems. This is unlikely to be how human cognition works (i.e., generating internal programs to test models). It's wild that the explanation actually demonstrates this capability. This is superintelligence. Is this not alien intelligence?!
6
3
36
7,385
Claude Opus 5.5 lets me tell a story about the universe's evolution better than I could on my own. Watch:
4
12
106
5,682
Opus 5.5 produced a strong overview of my book, Artificial Empathy (published in 2023). A lot has changed from 3 years ago, but this narrative may have survived the test of time.
4
3
16
2,064
What am I? Opus 5.5 answer in the context of the book Artificial Fluency. (turn sound on)
4
12
2,072
I had Opus 5.5 ingest my book Artificial Fluency and had it render visuals for each chapter. What is wild is that this is just an 82 KB HTML file (I recorded the video so you can see it on X).
3
4
32
2,394
AI is being infested by politics. The current "effective altruism" is being used like labels in the past, like "communism." Furthermore, AI "doomerism" is being associated with people who merely seek regulation or safety protocols. It's typical early narrative framing that leads to the kind of BS US policy that we have today in other sectors.
3
8
2,066
Wild! I had Opus 5.5 generate this explanatory video (single shot) of the newly discovered first three-dimensional aperiodic monotile. Coincidentally, the object shares the same shape as the Intuition Machine logo. ;-)
2
1
13
2,080
Linear vs non-linear thinking. The former always is easier digestable, but that doesn't imply it is correct.
3
124
The very curious thing about Gemini 3.8 Flash is that it generates diagrams to explain complex concepts. This kind of hints that it could be doing some shape-rotator-like cognition underneath.
3
3
19
2,030
Intriguing interview by TypeSafe AI founder. What other primitives are they coming up with? piped.video/watch?v=cFx9Z3ZX…
1
2
2
2,061
It seems to me that focusing on automation changes the economics of AI on its head! The narrative that scale is all we need (i.e., data centers) may take a back seat to greater physical automation.
1
4
10
1,875
1/10
Diego Almeida (former OpenAI researcher & founder of TypeSafe) on why RLHF created an 'Assistance Era' — and why true AI Automation requires throwing out the RLHF playbook. A talk about the motivations for Jev. 🧵
2/10 The AI industry is split into two radical cults:
• Cult 1: AI is accelerating exponentially, crushing every human benchmark.
• Cult 2: AI is a bubble generating zero real value beyond B2B SaaS wrappers.
Why do smart people see two totally different realities?
3/10 It comes down to Assistance vs. Automation:
• Assistance: Tasks designed to please the human in the loop (e.g., chat apps, coding assistants).
• Automation: Tasks designed to remove the human entirely — running silently in the background.
4/10 Nearly 100% of LLMs today are trained with RLHF (Reinforcement Learning from Human Feedback).
Here’s the catch: RLHF doesn't optimize for execution or correctness — it optimizes strictly for human preference.
Models learn to prioritize looking right over being right.
5/10 Because RLHF prioritizes pleasing you, sycophancy and overpromising are features by design.
If you send ChatGPT an audio file of noise and ask for feedback, it won't tell you it's noise — it will praise its "eerie, atmospheric vibe".
6/10 This is why businesses won't let AI make high-stakes decisions.
Current models are safe for throwing customer support docs at users (shifting risk to the user), but far too uncalibrated for expensive, automated business decisions. Humans remain stuck in the loop.
7/10 Think coding agents like Claude Code represent the next era? Think again.
Tools like Claude Code are still native to the Assistance Era. They make writing code cheaper, but they don't make software itself any smarter.
8/10 Despite massive leaps in AI, standard B2B SaaS hasn't fundamentally changed since 2019.
All we've done is latch chatbot sidecars onto existing apps. We're automating the writing of code, but the building blocks of software remain unchanged.
9/10 To achieve real automation, we need a new post-training paradigm:
• RLHF → Optimizes for human preference
• RLVR → Optimizes for log error rates / pure verification • The Next Era → Optimized for calibrated decision-making and reliable, autonomous execution
10/10 The takeaway: Tomorrow's AI won't just generate cheaper code — it will power smarter, fully automated software.
The shift from Assistance to Automation is where the next era of massive enterprise value will be built.
Context for the above infographic: piped.video/watch?v=cJ0EOzey…
17
7
39
2,616
Big Ideas on how to take back the economy piped.video/watch?v=Cb6xjJf4…
3
2,847
1/ We’re asking the wrong question about AI.
The future isn’t:
Humans vs. AI.
And it may not even be:
Humans using AI.
The real story is:
Humans + AI becoming a coupled developmental system.
And once that happens, both sides begin changing because of the relationship.
2/ Think evolutionarily.
Human develops.
AI develops.
That’s the obvious model.
But the more interesting unit may be:
Human + AI + Interaction + Culture
Because capabilities can emerge in the relationship that exist in neither participant alone.
3/ Call this participatory intelligence.
Human brings:
judgment
embodiment
context
values
lived consequences
AI brings:
search
simulation
memory
generation
massive pattern compression
Together they can produce capabilities neither has independently.
Intelligence migrates into the coupling.
4/ This changes the question from:
“Is AI becoming more intelligent?”
to:
“What kind of intelligence is the human–AI system becoming?”
That distinction may define the next decade.
5/ There’s an even stranger developmental inversion.
Humans develop roughly:
normativity → interaction → language → reflection
AI developed roughly:
language → reflection → interaction → consequence
Humans built the bottom floors first.
We gave AI the penthouse first.
Now we’re constructing the missing floors underneath it.
6/ But human–AI co-development causes these trajectories to converge.
AI acquires:
tools
memory
sensors
persistent interaction
real-world consequences
Humans acquire:
external memory
simulation
search
generated alternatives
algorithmic critique
Biological agency becomes symbolically augmented.
Symbolic AI becomes increasingly grounded.
7/ The result probably isn’t convergence.
Humans don’t need to become machines.
Machines don’t need to become humans.
The more interesting endpoint is complementarity.
The strongest system may be one where each creates productive tensions the other cannot create alone.
8/ This leads to a radically different theory of alignment.
Alignment is usually imagined as:
copy human values into AI.
But humans don’t have a single value function.
We constantly negotiate tensions:
truth vs. kindness
freedom vs. safety
efficiency vs. fairness
individual vs. community
Alignment may be less about value copying…
…and more about regulating tensions among competing norms.
9/ Which means disagreement isn’t necessarily a bug.
In a healthy system, tension is structural.
Too little tension:
groupthink
sycophancy
premature certainty
Too much tension:
paralysis
incoherence
endless branching
Intelligence may live in the metastable region between them.
10/ This gives us a different goal for AI assistants.
Bad objective:
solve everything for me.
Better objective:
help me become capable of regulating a more difficult class of problems.
The best AI shouldn’t merely increase task completion.
It should increase the user’s developmental capacity.
11/ That suggests a new failure mode nobody talks about enough:
developmental underloading.
If AI instantly removes:
ambiguity
difficulty
writing
recall
planning
decision-making
then humans may stop encountering the tensions that cause development.
No tension means no need for new capacity.
Convenience can become cognitive atrophy.
12/ But the opposite is also possible.
AI can generate:
50 alternatives
20 critiques
10 frameworks
infinite reframings
That produces developmental overload.
The ideal AI doesn’t eliminate difficulty or drown you in possibility.
It keeps you inside the zone of productive tension.
13/ This is why scaffolding may be the right metaphor.
A good developmental AI asks:
“What can this person regulate now?”
Then:
“What slightly harder tension can they learn to regulate next?”
The metric isn’t:
Did the AI solve the problem?
It’s:
Is the human more capable after the interaction?
14/ Something else happens with repeated use:
the human–AI pair develops its own microculture.
You learn:
when to delegate
when to challenge
when to ask for evidence
how uncertainty should be expressed
when to brainstorm
when to verify
what “good enough” means
The relationship develops its own participation genre.
15/ And that interaction can develop norms even if the AI itself has no biological needs.
The collaboration may come to regulate for:
accuracy
continuity
challenge
clarity
completion
uncertainty reduction
So meaningful agency can begin to emerge at the interactional level before it exists fully inside the AI.
16/ This also changes how we think about “who made the decision?”
Agency can migrate.
Early:
Human: goal, plan, act
Later:
Human: goal
AI: plan
Human: approve
Later still:
AI: detect need, propose, plan, execute
Human: supervise
Eventually some agency may reside in the interaction itself.
17/ The most important future AI capability may therefore not be “better reasoning.”
It may be better joint attention.
Human and AI both constrained by the same object:
same document
same experiment
same codebase
same dataset
same environment
Shared reference is the foundation of shared meaning.
18/ And now culture enters the loop.
Yesterday:
Human → Culture → AI
Tomorrow:
Human + AI → New Culture → Future AI
Future models train on culture partly created with earlier models.
Future humans develop inside environments partly created by AI.
This isn’t just technological change.
It’s a co-evolutionary feedback loop.
19/ That creates semantic ratchets.
Human–AI interaction invents a useful practice.
Culture records it.
Future humans inherit it.
Future AIs train on it.
The innovation doesn’t need to be rediscovered.
AI could massively accelerate cumulative cultural evolution.
But errors can ratchet too.
20/ Which means the deepest risk may not be:
“AI becomes too intelligent.”
It may be:
the entire human–AI semantic ecology becomes too self-referential.
Same models.
Same generated language.
Same genres.
Same assumptions.
Same priors.
High coherence.
Low diversity.
Low deformability.
Semantic lock-in.
21/ This is why epistemic diversity may become an engineering requirement.
Different humans.
Different communities.
Different models.
Independent measurements.
Contradictory viewpoints.
Minority traditions.
Competing hypotheses.
Diversity isn’t just morally desirable.
It may be structurally necessary for a robust meaning-making system.
22/ Buckminster Fuller called systems held together by distributed tension “tensegrity.”
That may be a useful metaphor for human–AI civilization.
Strong systems don’t eliminate tension.
They distribute it.
They maintain coherence because opposing forces remain differentiated.
The same may be true of meaning.
23/ So perhaps the future of intelligence isn’t:
Human intelligence + artificial intelligence.
It’s:
Participatory Intelligence.
Intelligence residing increasingly in relationships:
human ↔ AI
AI ↔ AI
human ↔ human
all interacting with shared reality and culture.
24/ The crucial design question therefore changes.
Not:
“How do we make AI smarter?”
Not even:
“How do we align AI with humans?”
But:
“How do we build human–AI relationships that make the whole system more capable of discovering truth, preserving autonomy, regulating conflict, and continuing to develop?”
25/ The future isn’t humans handing cognition to machines.
And it isn’t machines replacing humans.
The deeper possibility is:
we become part of one another’s developmental environment.
AI changes the conditions under which humans develop.
Humans change the conditions under which AI develops.
And culture carries the changes forward.
That is co-evolution.
@DarioAmodei @sama @demishassabis @elonmusk @finkd @JensenHuang
24
14
57
8,621
The narratives that become popular are those that are easy to linearly follow. Which is unfortunate.
2
1
57
Stop thinking of your brain as a single organ.
• Stanford just proved the forebrain and hindbrain are two completely separate organs with different embryonic origins.
• Split-brain research already proved the left and right hemispheres run distinct, semi-independent cognitive streams.
You don’t have a central command center. You have a four-way computational tug-of-war running 24/7.
[ Conscious Modeling / Semiosis ]
Left Cortex Right Cortex
(Symbolic / Sequential) (Gestalt / Contextual)
\ /
\ /
[ Action Selection & Arbitration ]
Basal Ganglia
|
|
[ Predictive Sequence Engine ]
Cerebellum
|
|
[ Homeostatic Ground Truth / Arousal ]
Hindbrain / Brainstem
1
7
39
3,517