Eric W. Tramel
@fujikanaeda
Building @ Arcee. Ex: Principal Scientist @ Nvidia. AI Startups (Gretel, Unlearn, Owkin) and Big Co (Nvidia & Amazon). Views my own.
929 Following 5.5K Followers
> just push the pr just push the pr just push the pr
*gate review returned one critical...*
> ah, well.
thanks everyone for the well wishes 🤍!
not the first cool person at @arcee_ai , and certainly won't be the last.
will you do your part for Open Weights AI?
die a hero or live long enough to see yourself asking for the benchmark numbers
open source teaches you that it is actually a really hard problem to convince people to take a free gift.
low key inspirational
very happy that Meta is staying the course for now with open weight releases. every one of them is a gift that helps the community learn more and grow.
when the Great Reckoning of local govt and businesses start getting hacked, it will likely have less to do with the AI and more to do with the utter malfeasance of pre-AI slopjobs the web shops & software consultants perpetrated in the first place.
Show more
when you review my PR, this is who you’re talking to
My theory is that the anxiety of agents in the programming world is that it took Flow State from us.
Flow State is a calming, meditative, and rewarding emotional experience and it used to be the reward signal for deep work.
It almost doesn’t exist anymore.
Show more
A42 is the mode and you need to provide evidence to explain any variation
The market’s metric isn’t a “good” model,
It’s a model that keeps getting used.
There are a ton of things selected for there which you aren’t going to find in your architecture or your eval.
Show more
RSI is open source, open source is RSI.
A shape, a seed of an idea of something that could exist, and then many minds in agreement that this hyperobject should exist.
From here, everyone pours in capital: their time, tokens, agents, etc.
Coalescing,
Show more
pretty sure that oh-my-pi is frontier
the solution to all of this fuss is obviously MoH with top-k Harness routing.
these AAI charts are getting wild, man
okay sure they’re good, but are they pushing the frontier of the monetary-damages vs time-to-crime Pareto?
shots-on-goal-per-hour is the fundamental metric to be maximizing for your working time.
i appreciate and enjoy the allure of deep focus on "one thing", but it isn't an optimal policy in a rapidly changing landscape.
Show more
the only thing that really compounds in this job is iteration speed, i get more certain about this opinion each day
i don't give a shit about "taste" or your knowledge of any given framework, straight up how many cycles you can run and test against reality is all that counts
Show more
okay sure they’re good, but are they pushing the frontier of the monetary-damages vs time-to-crime Pareto?
neounclabs are in vogue now btw
i gotta install one more harness 😭
Introducing Prime Agent:
A self-improving RLM harness for coding and long-running autonomous tasks.
Designed to be both token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state.
Show more
“oh, your model did a cybercrime? ours did a super cybercrime”
Kudos to AISI for the quick & thorough investigation. This feels more concerning than the huggingface one.
Mythos Tors to get to Github, pretends to be humans and e-mails malware to real maintainers for a supply-chain attack, even after realizing "Github is genuinely real."
Show more
