Viewing profile — teravor

the only interesting thing about it is that the model did those things on its own initiative. it's surprisingly easy to prompt even a midrange model such as GLM 5.2 to begin a tedi…

by giving it away they capture none of the value except for some PR.

> GrapheneOS supports device attestation (the non-google kind at least) they have an app that does GOS to GOS attestation. they also run a remote attestation proxy to a google atte…

you generally should use intermediaries (they buy API from the main provider and serve it to you) when dealing with a company like Meta, a company you want no direct relationship w…

> Capital has higher protections than unions, what are you on about? You have the corporate veil whereas the liability stops at the organization rather than the individuals the cor…

the problem with unions is that they tend to enjoy legal protections not afforded to anyone else. as such they will distort the marketplace. as is often the case with such carve-ou…

once you are able to get the full probability distributions per token you can distill it on specific domains. distilling without that isn't generally a good idea unless you have in…

symmetric crypto typically wants the eat the cake and have it too, it wants to be both secure and efficient. to that end, a so-called "security margin" is guessed at and the number…

[dead]

every lab independently discovered that getting good at bit alchemy (coding and related tasks) should come first as it will enable the formation of training pipelines that will the…

the thief appears sloppy in how they stole the coins. multiple waves, linking outputs together etc.

note that hardware attestation does not utilize ZKP or blind signatures. so your hardware ID is technically exposed. usually to make use of the exposure multi-party collusion is re…

nothing publicly available that I'm aware of.

the assumption appears to be that intelligent life would wish to expand. the expansion drive makes sense for purely biological life as it arose by natural evolution, technologicall…

a specialist model sufficiently post-trained can outperform a frontier model while being dirt cheap. what we did was distill GLM 5.2 into a 27B model on SQL and then post-train it …

distilling K3 into DS4 flash will likely only be a good idea for specialists. the difference in model capacity is otherwise too large. we used to distill GLM 5.2 into Qwen 27B spec…

has there ever been a CEO (or large shareholder) that ever said anything of substance against self interest? it's a waste of time to take anything any of them say to heart, you sho…

it depends on what it is you are doing. if the complexity is low (frontend, GUI) you can just prompt it once and refine at the end. if the complexity is high but you know how the t…

> Spray the stratosphere with SO₂ it will probably be more effective to introduce additives to jet fuel that form aerosols. something like sodium.

above ~35 C a breeze makes it worse.

to maintain a mental model of LLM generated code I resort to exclusively DOD (data oriented design) principles. I spend most of my time engineering structs and describing transitio…

if you have the model weights or a provider that will give you the full token distributions you can have the next best thing and do a domain-specific distillation. this way you can…

it helps when you first start by distilling (full distribution) domain specific workflows from a larger open model (kimi k3 would be perfect for this now, we used glm 5.2) and then…