leedrake5

Compacting ———-~~~~~~~ 40%

What RAM can be allocated to the NE? I am skeptical this is as easy as the GitHub repo suggests.

This distinction between ambiguous and incidental friction is brilliant. Was just talking about it with my wife because your specific example of a map has been a running debate between us. I just pick the least cost path via the algorithm to the point where I have poor spatial awareness. My father in law insists on using paper maps because he wants to know where everything is to be helpful. For me, the ease of the algorithm solves incidental friction. For him, it eliminates helpful ambiguous friction by denying him the opportunity to learn.

The wrinkle that needs to be added is that there are no truly universal rules as to what counts for incidental vs ambiguous friction - the definitions are relative to individual/project goals. I am working with some scientific instruments to map out chemical data, and 3d modeling is needed. I don’t particularly care about 3d modeling - it is incidental to me. The chemistry is the focal point. So the STL files are vibe coded so I can keep my focus on chemistry. But if I were working for the latest marvel movie, the reverse would be true. The actual chemistry would need to fit the script and the visual effect intended. To a scientist the visualization just needs to be good enough. To the film director, the world building physics and chemistry instead become the supporting actor. The challenge introduced by AI is that in ruthlessly eliminating incidental friction, you are being deliberate about what you choose not to learn. This is fine at a task level - but how many of us “found” our current expertise through incidental friction in the first place? I never wanted to do chemistry, I went to school for something else. But incidental friction led to discovery. That is my biggest worry, particularly for students and early career folks.

> as in your 1991 story “Division by Zero,” or a world where we raise robots as children

This is vastly more preferable than our current approach of raising children as robots.

“Chemical analysis was performed using a Thermo Scientific Niton XLt-900 XRF Analyzer equipped with a GOLDD X-ray fluorescence detector and real-time digital signal processing at a speed of 80 MHz. The XRF measurements were taken in the laboratory, using a Thermo Scientific SmartStand, ensuring consistent positioning. Since the Niton pXRF is programmed to utilize predefined matrices provided by the manufacturer, the “General Mining” matrix, which includes the relevant metallic and non-metallic elements, was selected for the current research. Calibration was performed using an internal calibration system defined by the manufacturer and validated using a known standard (SRM 2780).”

Translation: “We used a closed-source proprietary method by a manufacturer and validated it on a sample the manufacturer recommended.”

XRF generates spectra which represent atomic abundance (eg K-, L-, and M-line emissions). However, these are not international standard units. Laboratory units translate from spectra to mass fraction through meticulous preparation of standards. Handheld units often use a secret sauce algorithm to do this to keep analysis non destructive. I’m skeptical of the results - but what’s more important is that these results are not fully reproducible. If the company goes belly up, or less dramatically changes/updates the proprietary algorithms used to create mass fraction estimates, the reported results of this study won’t be replicable even with the same samples.

Not billions, but I was in a similar position. What saved me was the GPL license on the open source code and hiring a lawyer that kept my ownership of any software I wrote (though at a reduced hourly for them) and patenting new ideas connected to the project. When it came time for the company to scale up, I couldn’t have been in a better position.

The fact that so much of Çatalhöyük remains unexcavated is one of the reasons I am skeptical of so much speculation on its social structure. A palace could be under the vast majority that has yet to experience a trowel or brush. It, and sites like it, are so early that it is helpful - though extremely difficult - to forget the 9,000 years of future you know will happen to be able judge it on its own terms.

Possible. I pasted the output, but tried to get some paragraph spacing in the output, possible I accidentally deleted a character.

Thanks! First thank you for all the work to develop a math LLM. Those prompts are helpful. I shrank the weights down to f16 to fit on my system from f32 using the procedure at ggerganov/llama.cpp, then tried some of those prompts. From line 34:

> Problem: > If $\det \mathbf{A} = 2$ and $\det \mathbf{B} = 12,$ then find $\det (\mathbf{A \mathbf{B}).$

> Solution: > We know that for a matrix \mathbf{M}, the determinant of its inverse is given by $\frac{1}{\det \mathbf{M}}.$ We also know that the determinant of a product of matrices is equal to the product of their respective determinants. Therefore, we have: \begin{align} \det (\mathbf{A}^{-1}\mathbf{B}) &= \frac{\det(\mathbf{AB})}{\det\mathbf{A}}\\ &= \frac{\det\mathbf{A}}{\det\mathbf{B}} \end{align} > Therefore, we have: > $\det(\mathbf{AB}) = 24

The f32 answer: > We have that $\det (\mathbf{A} \mathbf{B}) = (\det \mathbf{A})(\det \mathbf{B}) = (2)(12) = \boxed{24}.$ > Final Answer: The final answer is $24$. I hope it is correct.

Final answer is the same, which is encouraging for quantization to expand hardware options.

Downloading this now to test it out - if anyone has prompts or tests they’d like to see let me know. I plan to run it with llama.cpp

I assume that this primarily benefits games and not any deep learning right? The most attractive aspect of Mac M1 is the huge memory boost. Might not be great for training due to the inability to distribute across multiple cards, but it makes for a great inference engine for stable diffusion, llama, and other large models.

> They found that the model relied on the style of language used in papers to judge if a study was replicable.

I think the failure here is to train the model on published results. I’ve worked with scientists who could write and infer a marvelous amount of information from shit data. And I’ve worked with scientists who poorly described ingenious methods with quality data. The current academic system incentivizes sensationalization of incremental advances which confirm previously published work. I’m not in the least surprised at the manuscript level that replication would fail.

The proper way to do this would be to log experimental parameters in a systematic reporting method specific to each field. With standardized presentation of parameters I suspect replicability would improve. But this would present an near impossible degree of coordination between different research groups. But it would be feasible for the NIH or NSF to demand such standardized logging as a condition of grant awards of a certain size.

Have you posted the drafts to Archiv? If so please share. If not, please do so. I don’t like the idea that a couple reviewers block knowledge transmission permanently.

Peer Review is the right solution to the wrong problem: https://open.substack.com/pub/experimentalhistory/p/science-...

On replication, it is a worthwhile goal but the career incentives need to be there. I think replicating studies should be a part of the curriculum in most programs - a step toward getting a PhD in lieu of one of the papers.