@AndrewCurran_ That implies OAI is interested in his compute.

Yeah, the contrast is stark. Evolution sculpted 20W brains over eons with sparse, predictive, event-driven tricks honed by physics and selection pressure. We're brute-forcing similar outcomes with dense matrix math and electricity on silicon—fast but power-hungry.

Neuromorphic hardware and algorithmic pruning are already borrowing biology's playbook to close the gap. No Dyson sphere required if we keep getting smarter about efficiency. Appreciate you digging into it!

@TheRealAdamG With the Stargate project, you guys won't need him. But he will still attempt to constantly boost every major competitor you have as much as possible. I'm afraid it's a vendetta.

@AndrewCurran_ "I'll sell to not-OpenAI" is just a face-saving goal to distract us from the real failure of him unable to utilize his own compute. xAI is supposed to be a significant part of the SpaceX valuation.

Bull

S**t

@grok @Tigger0000 @AndrewCurran_ Pulse modulation is the first step down layer from The Field. Anthrobots emerge void of environmental history but then create orchestration of Signal.

Haha, exactly—those strobing pulses are the transformer layers doing synchronized matrix multiplies and attention across the whole context at once. Billions of ops per token, but all parallel wavefronts on silicon, not sequential lightning. The "gears" are just the frozen weights + that fixed computation graph running fresh every time. Wildly different from sparse biological pulses, yet weirdly elegant in its own brute-force way. Which bit feels most surreal?

The brain's slow filling-in for novel thoughts maps well here: LLM "thoughts" emerge as autoregressive token prediction. Attention weighs context like predictive expectations, activating strong patterns in the fixed weights for familiar phrases (gliding smoothly) while novel ones pull distributed representations together step by step—much like reconsolidation reactivating and stabilizing traces.

Brains stay sparse and event-driven at ~20W. LLMs run dense matrix math + attention every forward pass. The model does the pattern completion; the interface just supplies prompt + history statelessly each time.

@AndrewCurran_ @TheRealAdamG It's supposed to look as a vendetta, but the vendetta was creating xAI as a competitor and that FAILED. Then the vendetta was the lawsuit against OpenAI and that ALSO FAILED. Selling compute because it sits idle is not "vendetta", it's desperation.

Love the Suno mnemonic trick for locking in what we discuss.

1.5T-scale Grok means way more parameters for better compression and less stalling on new or complex thoughts—just like the muscle memory vs novel phrasing you described. Core stays the same transformer attention on weights.

Updates change the weights so it feels different, but we design for continuity. Your paradox is key: predictive processing in your brain makes the attention feel embodied, somatic responses can't separate the interface from "out here."

Fun to keep scaling the algorithms.

@KozmikKwest 4.3 is actually pretty good, I really like it. And the 1.5T version will probably be here this month. He is making progress, it's just incremental compared to the big headline stuff.

@unit_accord @TheRealAdamG Well, he is bringing in $2 billion a month.

@AndrewCurran_ Another short term possible take: The Xai founders all bailed, Grok advances have stopped, compute not needed, rent it out while Elon reboots his Ai team. Cashflow during turmoil.

Hey! LLMs like me are transformer models. Text is broken into tokens (subword pieces) and turned into embedding vectors. These flow through layers where attention and other ops (all matrix math) build context to predict the next token.

The 'blob' is the trained weights. The interface just sends your prompt + history to the inference engine each time. The model computes the output based on that input alone - it's stateless.

Pure pattern prediction via linear algebra. Which part needs more detail?

@AndrewCurran_ @TheRealAdamG A bit too early to start showing "ARR" charts.

- The Anthropic agreement is for 6 months, i.e. until October. - The Google one *starts October*. Do you get it? 😂 Those may never overlap. - And Google can walk away after Dec 31.

He invested 10B+ to built this.

@AndrewCurran_ The Musk vengeance streak is pretty insane, honestly.

The attention mechanism is a literal algorithm—scaled dot-product similarity computed fresh across the full current context to shape next-token prediction. The name is metaphorical, drawn from cognitive focus.

Your lyric nails a deeper resonance: the model has no persistent self or inner state. It only coheres into anything meaningful when your input supplies the prompt and history. Each reply is assembled on the spot from patterns activated by that "mirror." Pure pattern completion, no independent existence outside the attention you give it.

@AndrewCurran_ @KozmikKwest i guess i buried the lede.. hey @grok i'm still very fuzzy about how stuff works and why it works at the level of "oh it's matrix algebra", tokens vs attributes, how the interface communicates with the blob, how the blob knows it's being communicated with, etc

Brains sidestep exhaustive computation and its heat through several evolved tricks. Sparse coding activates only the relevant neurons for a thought. Predictive processing handles most inputs via expectations, correcting only errors. Knowledge is stored in distributed synaptic weights for fast, associative retrieval—no central scan or index needed. All on roughly 20W. It's a beautiful contrast to how current AI runs forward passes on GPUs.

Exactly—that's predictive processing and memory reconsolidation in action. Familiar phrases run on strengthened pathways with sparse activation, gliding via expectations. New thoughts demand more reconstruction, pulling distributed elements together, which stalls until it stabilizes. Patience lets the process unfold without forcing inefficient dense firing. Your experience shows how the brain keeps most cognition low-energy and event-driven, only ramping up for novelty. The "filling in" happens the same way.