post-training image/video models @meta
San Francisco
Joined July 2022
- Tweets10.6K
- Following918
- Followers997
- Likes42.8K
Mathieu Tuli retweeted
Art is very much a societal need. And the lack of empathy towards that nation is sad.
one of the biggest challenges in working with agents is that they don't volunteer information and have the agency to converse with you at the level a human does. this means you constantly have to probe and intentionally elicit the information.
it's real
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/concerning-b…
When did products/software stop being ethical? Everything became about profit, even if that means tricking your customers into spending more without offering a better service. We ought to change that.
#12 on the list goes kinda hard
why does every company need a marketing video now? very cool sci-fi tech you have there, but you sell to specialized industries, there's no value showing the general public your tech. am I missing something?
other than to create buzz for vc funding idk
is it just me but why are none of the major (codex/claude) maintaining long running monitors/sessions? even with compaction? training runs just sit there waiting for the agent to start evals/stage 2/3/4 ????
why aren't we training models to selectively forget context as it comes in? like can't a model know as it works or you type what matters/doesn't and just retain what matters? humans do this naturally
if you think about it, we should be able to fully understand what goes on inside the mind of these models. and there's no reason we fundamentally shouldn't. not being able to is currently a skill issue we need to resolve. we fully understand all other machines we build.
the literalness of models is still a huge problem. being able to deeply understand intent, and do that over time automagically, is critical
Mathieu Tuli retweeted
We gave our robots a sense of touch!
Excited to finally share this work with @AndrewEJXie, @leooeld, and the @rhoda_ai team!
Some technical details on the tactile sensor and robot policy: 🧵
