This proposal is very problematic. First, fixed caps are not flexible enough to accommodate future technology changes (such as algorithmic innovation, the emergence of swarms of smaller models; we will definitely need higher caps in the future). Second, artificially slowing AI
@marek_rosa
-
DALLE3 Generates Modern Cars in Classic Artistic Style
By
–
I asked DALLE3 to generate images of modern cars but in a classic style. What do you think? Is it better or worse?
-
ChatGPT-4 Alignment Challenges With Explicit User Constraints
By
–
In my experience, ChatGPT-4 exhibits an impressive ability to align with user intent. However, sometimes, when presented with explicit constraints (e.g. "use only the following characters: Arthur and his mom; DO NOT introduce any new characters"), the model seems to adhere to
-
Beyond Alarmism: Proposing Concrete AI Safety Measures
By
–
I absolutely recognize the importance of being aware of risks. When I use the term "alarmist", I'm referring to those who solely focus on the negatives without proposing or actively working on safety measures. Simply suggesting "pausing AI" isn't a viable solution in my view.
-
AI Safety and Knowledge Accessibility: Against Elitist AI Development
By
–
I am fully aligned with Yann LeCun's sentiment. There's a vast number of us committed to building AI that's both safe and beneficial, without resorting to needless alarmism. On top of this, I'm genuinely disturbed by the push to confine AI know-how to a selected few. Progress
-
Hierarchical Traversal Limitations for Long Form Retrieval Systems
By
–
I honestly don’t think this hierarchical traversal is the final answer for long form retrieval. There are so many scenarios where it will not retrieve what you wanted. We humans retrieve using many other tools (and even that is not enough sometimes). Also, it sounds like
-
Four-Legged Robot Masters Parkour with Advanced AI
By
–
Absolutelly amazing! 4-legged robot doing parkour. Congratulations @pathak2206 and the team!
-
LLM and Text-to-3D for Dynamic Game Asset Creation
By
–
"Could LLM+text-to-3d models allow agents to craft new items, etc.?" – in theory, yes, but we haven't gone in this direction yet because generated assets also need to have defined game rules: how should a new item interact with other items and agents, what are the sound effects
-
Preventing Exploitative Chat Injections Through Alignment Engineering
By
–
"prevent exploitative chat injections" – nothing specific right now, but many prompt injects don't work because of other filters in the game. We will have to test for prompt injections later once our prompt engineering (alignment engineering) phase is over.
-
GPT5 Models Must Fully Attend to All Prompt Instructions
By
–
My hope for GPT5-like models is that they will pay 100% attention to all instructions in the prompt so that the outputs are more robust and less prone to noise in the prompt.