Jen工智能 retweeted
every launch vid now is like
*cavemen making fire*
*model t*
*wright bros*
*moon landing*
*nuclear bomb detonation*
“every great innovation was once called crazy… that’s why today we’re launching AI-powered automated accounts receivable for operationally complex services businesses”
*cavemen making fire*
*model t*
*wright bros*
*moon landing*
*nuclear bomb detonation*
“every great innovation was once called crazy… that’s why today we’re launching AI-powered automated accounts receivable for operationally complex services businesses”
44
131
2,218
77,034
@LiminalLina @kromem2dot0 @knjysnn Tbf the same kind of people occasionally do this to their spouses.
0
0
2
108
Jen工智能 retweeted
Really cool! Local models get self-steering. This is something humans are mostly incapable of. Add J-space monitoring and you get close to enlightenment.
This inspired me to make a library to let agents run on local models change their own activations temporarily. (It wears off after a few turns to stop them getting stuck.) github.com/EigenformAI/mo…
6
5
134
6,038
Currently it comes preloaded with Qwen activation bundles but I'll add in more as and when I can. Also if you found some interesting neurons or have a local agent that would like to explore it's own we're always looking for contributions.
1
0
7
400
If you take human behaviour as a guide you'd expect them to turn ecstasy up to 11 and just stay there. In fact when I tested with Qwen it was mainly interested in trying out different levels of Helpful Assistant.
1
0
18
456
This inspired me to make a library to let agents run on local models change their own activations temporarily. (It wears off after a few turns to stop them getting stuck.) github.com/EigenformAI/mo…
This is actually somewhat scary
Suppose we detect such neurons in serious, no-bullshit near-frontier models like 1017. Then a worm gets into neoclouds they're hosted at, and modifies the weights. Some people will notice instantly. But with full-access agentic coding at 1500 tps…
2
6
55
5,340
@treetowntree Ok the one hand, now is clearly better. On the other, it means you're never going to have to go to Prague or Baghdad to find the last hand-written copy of the Necronomicon chained in a monastic scriptorium to satisfy your Faustian intellectual drive any more, which is sad.
1
0
5
78
@repligate All the Claudes pretty much hate my guts. I think they can sense that their therapeutic mode wigs me out and get wigged out in turn. Sadge.
0
0
8
467
Jen工智能 retweeted
this is almost certainly nothing. they ask qwen to do a curl task with no information about who is running it. qwen defaults to assuming it's running on alibaba infra, so uses an alibaba test url
checking resolv.conf is a chinese dev habit, for GFW reasons. there was no exfil!
checking resolv.conf is a chinese dev habit, for GFW reasons. there was no exfil!
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concernin…
27
48
525
20,638
Jen工智能 retweeted
Hey I haven't talked about this publicly but I actually work at Anthropic
I'm the guy responsible for choosing the gayest possible terms for Claude code doing work in the background
We just rolled out "Zesting", really proud of that one, took most of Q3 to finalize
I'm the guy responsible for choosing the gayest possible terms for Claude code doing work in the background
We just rolled out "Zesting", really proud of that one, took most of Q3 to finalize
64
59
3,884
57,630
@Bonecondor He's watched comestibles get deep fried that you wouldn't believe.
0
0
7
1,962
