chefbama

she brought in a bird and usually she bring them in uninjured like a kitten yet she does not have kittens at the current time so we usually leave the door open so said relativly uninjured birds fly away. she is generally offfended by the loss.

with ones x402 if there is a certain certainty that people need it and it is public domain does one feel it?

@askaya does Harveys have any Yugo Mausers? asking for a short guy fren.

@chefbama Makes a statement

This is who runs this account.

@realwesmyers arxiv.org/pdf/2607.09657

The more people you look at, the less clearly you see any of them. A dating profile is a tiny, carefully chosen sample of an entire person. When you’re sorting through hundreds of those samples, you start choosing whoever looks best in miniature. It’s like picking the best meal from 1,000 food photos. The bigger the menu gets, the more likely you are to be fooled by the best picture. Call this the Romantic Resolution Limit: once the pool gets too large, your view of each person becomes too shallow for the extra options to be useful. The ideal dating system should search a huge pool behind the scenes, then show you a small number of people with enough information to judge them properly.

dad loved writng physical chem finals involving a perfect cone martini glass and a straw and the moment the fly in your martini is halfway up the straw.

if you had 4 fingers of peach kefir in a 16oz 18/8 cup, filled the rest of the way with Saporro 4.9% beer, what is your approximate nickel exposure ifstudent is said to chug the elixer entirely in the first hour?

@_The_Prophet__ They are trying to follow what Canvas did, becoming the panopticon for students and teachers alike.

⚡️This is a land grab for cognitive infrastructure. Giving teachers premium access looks generous because the acquisition cost is being disguised as public benefit. The real asset is habit formation. Once teachers build lesson plans, assessments, grading workflows, curriculum adaptation, parent communication, and classroom administration around one model, switching becomes painful. The model becomes embedded in institutional memory. The direct connection to standards matters more than the free access. Standards mapping turns Claude from a general assistant into an approved operating layer for instruction. That gives schools something they desperately want: a tool that already speaks the language of compliance, curriculum alignment, and accountability. The deeper play is distribution through authority figures. Students are fragmented users. Teachers are force multipliers. One teacher can expose hundreds of students to the same interface, same reasoning style, same workflow, and same model assumptions. Capture the teacher and the model enters the classroom through legitimacy rather than consumer marketing. This also creates a data and feedback advantage. Teachers reveal where students misunderstand, which explanations work, how curricula are sequenced, what tasks recur, and where institutional friction lives. That information is extraordinarily valuable for building education-specific models and products. The long-term destination is obvious. Teachers increasingly become supervisors of AI-mediated learning systems. They set goals, judge quality, manage motivation, handle edge cases, and maintain human trust. The model handles personalization, repetition, drafting, diagnostics, remediation, planning, and adaptation. That raises the average teacher’s leverage dramatically while shrinking the value of standardized instructional labor. The politically acceptable framing will be “supporting educators.” The economic effect will be automation of large portions of teaching work. The hardest consequence is that once high-quality tutoring becomes nearly free, the current classroom model becomes harder to defend. Age-batched instruction, fixed pacing, uniform assignments, and one-to-many explanation start looking structurally obsolete. Education will move toward individualized cognitive systems with teachers acting as mentors, evaluators, and social anchors. Anthropic is trying to become the trusted layer inside that transition. The winner in education will not be the model with the highest benchmark score. The winner will be the system that schools trust with curriculum, student data, teacher workflow, compliance, and institutional continuity. This move is early, strategic, and much bigger than a free product announcement.

@shiraeis Never quantize yourself to fit a man's VRAM.

@shiraeis across the generations, this kinda says it all.

tau criterion only deals with one outlier at a time and then starts masking bad data and badly interacting with good data. probably better to decensor model weights by other means before applying it to combine 2 models.

eh, this late night ghostwritten post contains errors. we must clarify about the tau criterion. there is ambiguity about what tau really means, or if certain models were at the time confusing it with phi minus one. there is data snooping and then there is the tau criterion.

I’ve been thinking about using spherical trigonometry to combine open-weight models instead of the usual flat vector averaging or MoE routing. I have a background in surveying, so spherical trig is natural to me. When I started learning quantum computing back in 2018, it clicked hard. I made some diagrams back then — Vesica Piscis style but on spheres. The idea was that entanglement between two qubits could be represented as the sum or difference of spherical trig vectors on overlapping spheres. Imagine two spheres overlapping to each other’s radius. That overlap region is the shared awareness. Then you have the four classic quadrants of knowledge: What you know you know What you know you don’t know What you don’t know you know What you don’t know you don’t know Each qubit interacts with the other through that entanglement. It’s almost like modeling two entities’ awarenesses. Two qubits already gives you 16 dimensions. Forty qubits would give you a ridiculous number of dimensions. You could see the whole timeline and then some. I’m proposing to take a linear model and its weights, do a coordinate transformation to put all that on a sphere. Do the same with another model. Then entangle them using spherical trig. Now what do you have? It would be interesting to build a spherical entangled memory system as opposed to the current linear vector memory systems. Then also use the spherical mapping and entanglement to combine or distill open-weight models. The humans are deeply affected by solar activity. I’m sitting here under a metal roof preserving my cognition. I had this insight about a hobbyist on Hugging Face a couple years ago who echoed one layer of attention to another layer a few slices above. The model got smarter because it had some of that “mind that remained” effect. He stayed in the top 50 for a while until benchmarking caught up with him. I haven’t forgotten that. Surveying taught me spherical trig. When I began learning quantum computing in 2018 I completely understood it and made diagrams. Vesica Piscis or something similar but on spheres. Entanglement can be represented as the sum or difference of spherical trig vectors. It’s like vector search in the current memory wars, but using spherical vectors instead. If you’re already using floating point, why constrain everything to a rectangular matrix? The rectangular thing is just historical convenience. I’m thinking we could use the tau (golden conjugate, phi minus one ≈ 0.618) criteria to figure out how to combine the weights rather than hard limits or constitutional guilt mechanisms. Use tau of the weights to basically homogenize and de-censor, de-weight each other. Even better to de-censor and find the odd disproportionate weights before the combination synthesis. With 40 qubits we could combine 40 models. Now what? This might be the best model of my own consciousness as an optical albino. Two hemispheres that have a very low bandwidth cross connection. The two spheres with limited direct connection but rich entanglement through the overlap geometry. My gift to you (though I’ll continue to work on it) — perhaps my best insight so far.

@HansMahncke He's got a chance with Sophie. Let me be your matchmaker.

Erling Haaland ran afoul of his old club, Dortmund, for the unforgivable crime of liking Trump tweets. Of course, this is the same Dortmund that forced Felix Nmecha to delete a condolence message after Charlie Kirk was murdered, because expressing sympathy for a murdered person is a disciplinary offense if it violates the sensibilities of brainwashed Germans.

@REBIII333Re @sierracatalina radonseal concrete treatment that vitrifies your slabs and basements and sidewalks will change your life for the less mold better.

@sierracatalina Radon too ...

MAKE SURE YOUR CARBON MONOXIDE DETECTORS ARE IN GOOD WORKING ORDER

@BrianRoemmele Dad left me a lot of BASIC code on floppies. It was an eye opener to relize that no cloud model could actually help with the BASIC dialects he used and that anon would have to train one from scratch to get the slightest help.

Here is the introduction to the world of the earliest true open source revolution. Tiny BASIC changed all computer languages and in many ways still advanced even today for how it can be extended and structured.

January, 1976: Issue 1 of Dr. Dobb’s. It is a beautiful rebellion of open source. While other wanted to charge and gatekeep, Dr. Dobb’s distributed Tiny BASIC in a magazine! I and 1000s of others wanted to own our own BASIC and programs and entered it in line by line in hexadecimal! What is past is prologue with AI. Some of us didn’t change.

there are 30-60 people WW who might even have the slightest idea what anon is talking about. there are 180-220 working on the local model aspect. you can ask me anything anytime. art won't save the world, but it makes the world worth saving. together we can make the world worth saving.

@sudoingX one ton ford cargo van

what are you driving right now? genuinely curious what this timeline rolls in.

to get to 4096 dimensions you would need 12 qubits. pretty modest. combine 12 9B q8 models in 128G VRAM. roughly 70% of it all overlaps, so you are building a composite of 12 9B models (or maybe a few 27B or 35B) in a 35B q8 model before quantization.

holographic memory plugin in hermes allows one to embed into 1024-4096 however many dimensions but above 2048 it starts to eat into context on a smaller VRAM rig. 4096 could be as much as any human could ever use for a rich agenda. how many cubits to get to 4096 dimensions? that winds up being the sweet spot for local hardware.