clbrmbr

my understanding is that a sufficiently large model will memorize the training data once enough representations are built up. Opus 4 scale seems to have been sufficient. cf NYT vs OAI.

(another guess) to convert to dollars, the gold would need to be delivered to the custody of the US Fed, which is easiest when the bars are physically in the Fed's vault. in a low-trust situation, you have all the gold in your own country, but then its so much leas liquid bc you need to physically deliver it to trading partners, or they must trust your accounting of their share in your vault.

ikr? ive been really hoping to be embedded already but Zephyr/C seema still the pragmatic choice for most embedded.

bump up to EmbeddedLinux and i think Rust wins ovr C

It's totally plausible portmanteau. I get the feeling thought that it sounds more marked in English than in German where these kind of mashups are formed more freely.

@labo333 do u think soon we may need a dictionary? ive been playing with something u may like, but my approach has been to ask for definitions in-session so hard to do from outside.

https://www.themachinevernacular.net/

whoah. with direct demo, whats going on over there?

Bau is awesome. He is often the adult in the room who asks "but why", and cares deeply about human understanding. I am very bullish on Northeastern.

is an xtensa lx7 (esp32-s3) target available that does not use llvm?

shudder. but, where else could we go? switzerland? my point being we need to try to get 1995 back. id gladly trade my iphone for a car phone.

Indeed I think much is shared across sessions and projects. I say we learn The Machine Vernacular [1].

[1] https://www.themachinevernacular.net/

> let a coding agent run loose with them without approving every tool use

You should try. There’s a difference in kind between approving each tool use and spinning up an all-night session with full permissions to run any tools, encouraged to run subagents to preserve context.

How are you using it in Claude Code? What is the native harness that GLM was post-trained in?

Creating on-demand automations for an ESP32-based product, that includes driving a display up to 1000px.

Oh just that it made me confront ABIs. I have a special purpose use case (home automation), so I wasn’t worried about general purpose compilation, just the subset of lua that fable tends to write for the prompt we use.

I have no idea how it’s implemented. Lua goes in, position-independent dependency-free xTensa LX7 code comes out the other. Maybe I should read the code?

It’s extremely fast. 3kB of lua becomes 1kB of machine code. I’ll have to get a blog post together on it.

Fable 5 is the model you work with, Opus 5 is the model Fable works with… it’s meant as a subagent.

I was battling GCC… until the new guy (a smart business major) pointed out I could just compile from Lua to ASM directly. Claude was happy to write a compiler over night. :facepalm:

The range on one RF system I worked on was 2mi with a wet noodle for an antenna.

I discovered last week that Fable 5 can write perfect xTensa LX7 assembler code without tools or references. Mind blown.

But, when working on a creative graphics task, the results were best in Lua, middling in integer-only C, and underwhelming in ASM in terms of creative depth.

You just need to be able to remotely take over a system with a compatible radio within range.

That’s why I’m so concerned about Bluetooth

Would it be practical to have the ESP P4 self-host?

Whoah, so cool how the map shows the shadow of mountains at that late time in the day. Geneva and the Jura look awesome.

There are many wireless pump-and-reservoir systems that while not internet connected, use insecure RF links. These local RF (and casting a wider net, Bluetooth) interfaces are also ripe for abuse.

Yikes. Man, there’s a market opening for someone to redistribute open source projects with supply chain assurances!

Mamba/SSMs could change this picture.

The policy I set at my org is to put an organically written message at the top of your Claude blob. Something like:

tl;dr: ___human-written____

____claude blob here_____

Where the idea is that humans only read human writing, and Claude is able to share context between sessions with the jargon-riddled blobs.

Trouble is, I catch people every day blatantly writing the Tl;dr with Claude or ChatGPT.

Many orgs (esp w ISO27000) have a vulnerability management policy that involves patching at least critical CVEs within a short timeline. Tools like trivvy make it possible to do the scans…

Are you thinking of fundamental architectural changes (like the Transformer)? Or incremental (like MoE or GQA)? Are there specific neolabs or techs you are following that lead you to this prediction?

Indeed it seems quite possible we are one architecture breakthrough away from existing chip stock driving us all the way to ASI.

Nobody has done AFAIK the information theory to prove it’s not possible.

There’s another way besides distillation that’s way cheaper: You can have the big model build prescriptive skills that the small model follows.

Take the “train” portion of tasks on some benchmark, have K3 complete it, and then output detailed descriptions of tools used and why, then run the validation tasks with some small model that has access to the skills.

Funny retcon, but come on… TIL huggingface started as a chatbot for teens who didn’t get enuf hugs.

I think so, if you can inject synthetic errors with the same distribution as the errors you care about.