I touched on the idea of sleeper agent LLMs at the end of my recent video, as a likely major security challenge for LLMs (perhaps more devious than prompt injection). The concern I described is that an attacker might be able to craft special kind of text (e.g. with a trigger
@karpathy
-
AI Systems Matching Human Performance in Five Years
By
–
Just to make things a bit more concrete here, you're saying there is a 10% chance that an AI system can match or outperform people at 95%+ of these in 5 years
-
AI as Imagination Amplification Beyond Intelligence Extension
By
–
(I’ll add thoughts to thread as they come up. )
Thinking of these tools as purely extending intelligence feels too constraining, could just as importantly be understood as imagination amplification; we’re seeing a lot of that too with generative IA. -
Intelligence Amplification: Empowering Humans Over AI Superintelligence
By
–
e/ia – Intelligence Amplification
– Does not seek to build superintelligent God entity that replaces humans.
– Builds “bicycle for the mind” tools that empower and extend the information processing capabilities of humans.
– Of all humans, not a top percentile.
– Faithful to -
LLM Customization Ecosystem: From Prompting to Finetuning
By
–
LLM customization ecosystem is heating up – Remarkable that prompt engineering works at all, but stagnates
– Retrieval can help few-shot prompts, but still…
– Finetuning (BC/RL) is the cannon. But is much more involved
Congrats @realSharonZhou & @GregoryDiamos on the launch! -
LLM Hallucinations: Origins, Mitigation Strategies, and Open Problems
By
–
Great tech talk on subtleties of LLM hallucinations by @johnschulman2 : where they come from, how to mitigate them, remaining open problems. https://
youtube.com/watch?v=hhiLw5
Q_UFg
… -
FaceTime with ChatGPT: Voice Interaction Experience Enhancement
By
–
FaceTime with ChatGPT https://t.co/52qfpzChrX
— Andrej Karpathy (@karpathy) 27 avril 2023
Fun, qualitatively different experience over "texting"
Ty @CallAnnieAI for shoutout to my tweet a ~year ago pitching this as an idea https://t.co/LE6dYcDsxZ
With improvements to personality, latency, ASR/TTS could be magical ✨ https://t.co/eMeYzyum5TFaceTime with ChatGPT https://
callannie.ai
Fun, qualitatively different experience over "texting" Ty @CallAnnieAI for shoutout to my tweet a ~year ago pitching this as an idea https://
x.com/CallAnnieAI/st
atus/1651379993413558272
… With improvements to personality, latency, ASR/TTS could be magical -
Gen2 Video Generation and ControlNet Advances Impress Karpathy
By
–
wow. coming from @runwayml #Gen2 https://t.co/P1xedllGbV
— Andrej Karpathy (@karpathy) 21 avril 2023
While on the topic of video generation I was also mildy mind-blown a few days ago by multiControlNet and friends: https://t.co/EJyjDi7wNS
And the earlier, bit more professional take, "anime rock paper scissors":… https://t.co/JJhIiucyDnwow. coming from @runwayml #Gen2 https://
research.runwayml.com/gen2 While on the topic of video generation I was also mildy mind-blown a few days ago by multiControlNet and friends: https://
reddit.com/r/StableDiffus
ion/comments/12i9qr7/i_transform_real_person_dancing_to_animation/
… And the earlier, bit more professional take, "anime rock paper scissors": -
LoRA Finetuning: Shifting the Balance in LLM Adaptation
By
–
There's a chance that LoRA finetunes work so well that it dramatically alters the finetuning vs. retrieval + few-shot prompting power dynamic in favor of the former for many applications.
— Andrej Karpathy (@karpathy) 20 avril 2023
PEFT (Parameter Efficient Finetuning, LoRA included) are emerging techniques that make it… https://t.co/u3Udqpk3dzThere's a chance that LoRA finetunes work so well that it dramatically alters the finetuning vs. retrieval + few-shot prompting power dynamic in favor of the former for many applications. PEFT (Parameter Efficient Finetuning, LoRA included) are emerging techniques that make it
-
Embedding and Ranker Choices: Ada and kNN Performance Comparison
By
–
For science I also added:
– Choice of Embedding: simple tfidf bigrams or the OpenAI API embeddings ada-002 (ada should work better (?), tfidf is much much simpler)
– Choice of Ranker: kNN (much faster/simpler) or SVM
Default that seems to be both good & fast is ada+knn