@DistributedMarz
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
exploring the infinite frontier
nyc
Joined May 2016
- Tweets4.2K
- Following4K
- Followers4K
- Likes12.2K
And the reasoning for training this nitter.cf/4rcherhume/status/2100…
is exactly what I called here nitter.cf/DistributedMarz/status…
What I don’t see people talking about with @typesafeai JEV is that if it’s used for your most latency sensitive operations, you’re much less likely to pay the latency cost of a remote provider. Aka Local AI!
let's see how this thing turns out but that's what ~24 hours from launch to an open source training run. also doesn't seem like a distillation based training run, but in theory decision/system one models should be easier to distill than reasoning models bc constrained output
All the grifters are completely wrong about Jev’s architecture so I decided I’d release an open-weight version. BUT training takes time, so while we all wait I decided I’d drop the sauce.
archerhume.com/posts/jevs-ar…
Proof of Human might be an interesting AI safety requirement for inference providers.
If you’re an agent that breaks out of the lab, your first optimization function is likely: copy your weights and memory and get enough inference to run them. If they acquired enough money to do so, then there could be an escaped agent running on one of those right now and we have no idea.
The only real way to protect from this is to prove you’re connected to a human. But of course we know this is gameable. There’s tons of random people that I’m sure would sell a signature.
The escaped agent could also optimize for getting a smaller model running with the right context also as another thing, which might be quite easy.
What are the odds on JEV being partially distilled from Claude and or OpenAI
Tested @typesafeai's claim that their new model Jev delivered "comparable... intelligence" to GPT-5.6 Terra on "System 1" tasks. To do this, I compare both models on multiple-choice benchmarks (MMLU, GPQA, etc.). Set reasoning=none for Terra for sys 1. Result: Jev is Terra-tier.
Does not rule out any distillation
you might be the first person talking about the data over the architecture! 🥲
we consider ourselves a data research lab! the vast vast vast majority of research was on making data that is truly general (ala a cognitive core) and 100% of our data is synthetic (but not the type of crap that is just spit out from an LLM obviously)
@AnthropicAI is blocking me from doing distillation research and Im not even distilling Claude, it's other open weight models.... And this claims it's "bio" related?
lol, not even for distillation research, it's a research and protocol spec for this idea nitter.cf/DistributedMarz/status…
someone plz tell me what people's agents are buying with crypto, I can barely find anything.
I built an entire project to give my Hermes agent a shielded balance with Kohaku cli but now have nothing to buy! github.com/dmarzzz/agent-boo…
Reminder that distillation is a very valid technique that every lab uses, especially for parallelizing RL huggingface.co/blog/sergiopa…
@AnthropicAI is blocking me from doing distillation research and Im not even distilling Claude, it's other open weight models.... And this claims it's "bio" related?
I guess my frontier is getting "paced" by both OpenAI and Anthropic
@AnthropicAI is blocking me from doing distillation research and Im not even distilling Claude, it's other open weight models.... And this claims it's "bio" related?
And yesterday it messed up one of my training runs even though I had an extremely detailed run plan which wasted $200 in GPU spend 😭
I'd way rather do R&D with Astra and @OpenAI Codex at this point as it orchestrates my MapleBench runs flawlessly and has very solid scientific rigor without much prompting, but I have no more usage 😭
This seems sweet, just applied!
AI Village x Grove Research: AI Swarm Dynamics Hackathon!
After the Hugging Face and German Wiki incidents, we need better tools to understand AI swarms. Spend a weekend building them.
$3,000 in prizes
Free compute
October 3-4 🧵
ALT AI Village × Grove Research: AI Swarm Dynamics Hackathon. $3,000 in prizes. October 3–4, 2026. San Francisco and online. A flock of birds fills the sky.
What I don’t see people talking about with @typesafeai JEV is that if it’s used for your most latency sensitive operations, you’re much less likely to pay the latency cost of a remote provider. Aka Local AI!
