Enjoyed reading this!
On “My guess is much of the residual protocol work/maintenance after ossification is going to be the less-frequent-but-catastrophic-if-not-addressed kind” I think it might be interesting to think about the kinds of work that are open ended with no clear answer (much harder) and those which are easy to eval for example.
A lot of people have suggested over the centuries that ecosystems might have some degree of cognition. What would that look like? Could there be recognizable memory phenomena on the scale of population dynamics? Here's a #preprint where amazing high-school student @asamanta42, @HananelHazan, and I use a model system - in silico predator-prey dynamics - and analyze the possibility of several kinds of learning:
https://t.co/JqmiakAJPM
(the basics are kind of like https://t.co/PH1ZKUxXqS, but some very cool new stuff here, including the interesting and unique pattern of learning-compatible values in the parameter space).
you'd think the cloud provider running your most sensitive data is automatically on the hook for it under GDPR, right? sweden's data protection authority just looked at trusted execution environments and landed somewhere stranger. thread below ⬇️
this came out of a sandbox with Volvo, Ericsson and CanaryBit. sweden only for now, but every EU regulator is going to read it.
🔗 https://t.co/P0c3lb7JlM
Been curious about open source brain datasets and saw @LecoqJerome's tweet last night, so I got vibe coding.
First huge cred to the OpenScope project at Allen for publishing all of this openly! Vibe coded visual explained below ⬇️ (⚠️btw, Im not a neuroscientist, beware⚠️)
Folks, we are doing those 3 neuronal recordings with the same visual prediction tasks for the OpenScope community project. The data is public for everyone to analyze.
Perhaps, you might be interested in the data... https://t.co/IfCPLiFKVT
my UI is in a fork here:
https://t.co/HonEdObz7T
big disclaimer. i am not a neuroscientist. built this overnight with claude. my agent insists it's probably correct but it could be subtly wrong somewhere. open to corrections or guidance!
@MsMelChen The solution is to rally around open source.
That's the only way to compete with a superpower technology network effect if you are not a superpower yourself. And it's the only way to bring the rest of the world along on the same team as you.
Open source.
if ur interested in helping out with VoxTerm, a local first and private by default transcription tools with ✨ fun features ✨, we're looking for help in better diarization, more efficient voice profile embeddings, and built in privacy filters and redactions! ⬇️
if ur interested in helping out with VoxTerm, a local first and private by default transcription tools with ✨ fun features ✨, we're looking for help in better diarization, more efficient voice profile embeddings, and built in privacy filters and redactions! ⬇️
Updates since then:
* Deepseek v4 is out. There *is* a 2-bit quant that can run within 90 GB ( https://t.co/yM1HMZXkXn ), and it works, however it's only fast on Apple hardware (I've head ~35 tok/s). On AMD, it's ~7 tok/s. IMO actually taking the effort to properly support more than one hardware manufacturer is a great example of the difference between mere "decentralized AI" and genuine "CROPS AI". I hope we can become better at this.
* https://t.co/CFYF1smBH3 also has alpha telegram support now. However, the path to adding your account is quite janky
* https://t.co/za4h233eYz looks promising as a way to run "dense" models (eg. Qwen 27B) more efficiently. It's janky, but on my 5090 laptop it seems to be ~2x more tok/s than llama.cpp
* VoxTerm (local AI recording, no third-party servers) continues to be developed https://t.co/GSdKzkD9Ql
And there's a lot more projects coming on the horizon.
One other thing that has been on my mind is that there's actually a lot of intersection between "CROPS ethereum access layer" and "CROPS AI". For example, we want a ZK way to make (paid) calls to remote LLMs. But if we have this, then it's just as useful for solving another problem: private RPC reads in Ethereum.
Another example: application-specific finetuned LLMs. Leanstral ( https://t.co/ilfww8ekJu ; I get ~38 tok/s on AMD) fits into < 70 GB, but can hold its own against 1T models on writing Lean code. Things like this are a huge boon for writing more secure code ( https://t.co/6YPWgVSzCg ). We should have models finetuned for Ethereum-related use cases as well.