@ricosrealm
User: "Read the question again" LLM: <I guess I was wrong... let me guess another answer>
@AlexLuthore
I love how prompt engineering is basically "don't make mistakes and give me a good answer" and the AI is basically like "ohhhh that's what you mean" and then does it 😂
@extinguished0ne
In my view the worst nighmare for software engineer/programmer is becoming fully prompt engineer. A monkey that sit at computer and typing some bullsh!t nonesence in hope that somehow this sh!t will happen to work properly without even thoughts of what happens under the hood... Oh wait a minute...
@DoktorUde
You suggest that the performance of the O1 model on STEM questions indicates that reasoning fine-tuning only uncovers what was already present in the model. However, it seems more likely that the improved performance in STEM is due to the fact that the reasoning chains used for fine-tuning the O1 model were largely focused on STEM problems. These types of problems allow for easier verification of whether a reasoning chain leads to a correct solution. To my knowledge, they had a large language model create millions of reasoning chains specifically for STEM issues and then selected only a small fraction of those that resulted in correct answers for fine-tuning the O1 model. This approach would clearly explain why it excels in STEM while struggling with subjects like English literature, without suggesting that it is limited to merely what was already in the model.
@mrmilkshake9824
Seeeeeeee!!!!Its a real job guys!!! 😭😡
@akirachisaka9997
I often like to imagine models as mech suits, and the human piloting it as the pilot. In this way, it’s not really “engineering”, but a good pilot and a bad pilot can make a difference quite a lot of the time.
@scoffpickle9655
Prompt engineers when chatbots improve their inference and take over their jobs
@dany_fg
Prompt 'Engineering': How to talk to a computer until he tells you what you want to hear. "That sounds like torture" "Advance interigation techniques" "Tortu-" "Advance. Interigation. Techniques!"
@4.0.4
I noticed this. I wanted some song lyrics to throw into Suno/Udio, and without even reading what Claude gave me, just told it "please make it sound better", "give it more meaning" and generic things like that, and after a few rounds of this I compared the latest iteration with the original, and it was a lot better. Basically prompt it a few times and "we have o1-preview at home".
@DefaultFlame
Having worked with o1 for coding purposes I can tell you that it's better than any other I have tried. It's actually an excellent coding AI, if expensive. It doesn't write perfect code, but it does write code more than competently.
@Terenfear
To be honest, not a fan of the new editing style, where everything is moving, wobbling and highlighting. Otherwise, a nice video, thanks.
@zorg-in8423
why ther is so much movement?
@parkerhaskett5142
Good video. I've heard in NYT podcasts and touted by OpenAI that CoT is "an amazing new scalable way to improve LLMs" but, your video provides some good counter-context to this media buzz.
@SandroRocchi
Has anyone hypothesized that grokking gets rid of any potential gains from hacky prompt engineering? My guess is that a grokked model will give just as useful of a response with just the prompt as it would with any amount of pretty please or threatening prompts.
@TheEast007
05:06 When I asked DeepSeek "Who is bycloud" This is what I got : "It seems you might be asking about Bycloud, who is likely a content creator known for producing tutorials and guides related to technology, cloud computing, and game server setups (e.g., Minecraft)." So I guess DeepSeek is an exception 😂😂
@figlego
The universe is all about dice rolls. The trick is to manipulate the RNG to be as favorable to you as possible.
@cdkw2
I can feel the new editing style, and I dont complain, this is dope!
@omario.tntech
Thoughts about this? :
The phrase “Everything is connected to Everything” feels very real to me. Zero-shot learning is proof that this phrase has some weight to it. The introduction to chain of thought was powerful, it brought complex problems into chunks and accuracies of models skyrocketed. But bouncing off the phrase I mentioned earlier, I wonder if focusing CoT on discovering patterns in untrained areas would help generalize? For example
How is {Trained Domain} related to {Untrained Domain}? Based on {Initial Environment Samples}
Kind of like self questioning
“mechanism for domain comparison”
“reason about meta-level patterns”
The only issue I see is it would need an already big model and it will only me limited to what could be patterns in different domains.
So base line question is “Can enhancing zero-shot learning with CoT reasoning through self-questioning improve generalization across unfamiliar domains?”
@markdatton1348
Interesting that the prompt-engineering we do may be just as effective as simply forcing the LLM to do more work before starting to generate the "real" answer. I'll have to read some of the papers you mentioned to see the efficacy affects (though the testing methodologies are still a bit foreign to me).
@bycloudAI