10 Breakthrough
Technologies 2023
LLMS
-
Claude Vulnerable to Limited Prompt Injection Attacks Affecting Safety
By
–
Update: Claude is *somewhat* vulnerable to prompt injection, but with limited harm. You can make Claude believe it's said things it hasn't, but seems to have no effect on its commitment to safety:
-
ChatGPT Limitations and Innovation Opportunities in AI
By
–
ChatGPT is incredible, but I think it's important to remember that it's not a silver bullet. As with any technology, it has limitations and it's important to approach it with realistic expectations. But this also means there are still plenty of opportunities for innovation.
-
Improving AI Model Behavior Through Specific Instructions
By
–
I get that — it has a sort of 'sanitized' feel. But you can get around that by giving it a very specific set of instructions re: how you want it to act. Future models should improve this.
-
GPT-3 limitations and expectations for GPT-4 improvements
By
–
Unfortunately the GPT-3 models were trained on data no more recent than Sep 2021. The world changed since then! Please keep kicking the tires and asking for intel assessments, will help us train the system, but big improvements not expected till GPT-4
-
Developer frustrated with ChatGPT rate limits on mega-prompts
By
–
I've got a couple more mega-prompts than I've been developing. Just need ChatGPT to stop booting me out for using it too much.
-
ChatGPT Capacity Issues and Research on Voice Tone Prompts
By
–
Tbh, it's at capacity a lot these days. And if you use it too much in an hour, it'll stop you from using more. So I'm mostly using it for research on improving its use right now. But I did use it somewhat to help create the voice/tone style prompts for testing.
-
Training Costs for Large Deep Learning Models Rising Significantly
By
–
I also made this prediction, which I still think will soon kick in, as the cost of training modern very large deep models is getting pretty high…
-
AI Models Approaching 100 Trillion Parameters Ahead of Schedule
By
–
In my 2018 keynote at ICML I showed this curve and predicted that in 2025 we would have models with 100 trillion parameters. We might get there sooner…
-
GPT-2 Development Finally Showing Results and Success
By
–
Finally all that tinkering since GPT-2 is paying off