How DeepSeek Built The Current "Best" Math Prover AI
Related videos
How DeepSeek V4 Broke AI’s Cost Curse
bycloud
101.8k views
The Most Clever Trick To Speedup LLMs
bycloud
17.7k views
The Celtics PROVED They're The BEST Team In The East
The Arena
92.6k views
Is This The Dumbest Thing California Has Ever Built?
Matt Walsh
61.8k views
The Death of RAG?
bycloud
15.0k views
DeepSeek Just Added Parameters Where There Were None
bycloud
34.5k views
The Only Reason Why The INSANE AI Datacenter Build Out Would Make Sense
bycloud
20.7k views
Invisible Ghost Galaxy Proves Dark Matter Built the Universe, Cloud 9 Explained
Anton Petrov
42.4k views
How to Solve the Biggest Problem with AI
Futurepedia
25.9k views
DeepSeek V3.2 Just Broke SoTA Again… But How?
bycloud
168.2k views
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Related videos
How DeepSeek V4 Broke AI’s Cost Curse
bycloud
101.8k views
The Most Clever Trick To Speedup LLMs
bycloud
17.7k views
The Celtics PROVED They're The BEST Team In The East
The Arena
92.6k views
Is This The Dumbest Thing California Has Ever Built?
Matt Walsh
61.8k views
The Death of RAG?
bycloud
15.0k views
DeepSeek Just Added Parameters Where There Were None
bycloud
34.5k views
The Only Reason Why The INSANE AI Datacenter Build Out Would Make Sense
bycloud
20.7k views
Invisible Ghost Galaxy Proves Dark Matter Built the Universe, Cloud 9 Explained
Anton Petrov
42.4k views
How to Solve the Biggest Problem with AI
Futurepedia
25.9k views
DeepSeek V3.2 Just Broke SoTA Again… But How?
bycloud
168.2k views
Top Comments (10)
Seems like this method of recursively breaking down problems into smaller sub-problems until triviality is achieved could be applied to many problem domains, not just math.
hey cloud can you cover every deepseek paper because these people are brilliant
I love seeing open source winning.
3:58 the “there exists delta” and “for all epsilon” need to switch places
There's a significant mistake in the video (though admittedly not just in the video, which stems from the original paper). As discussed in the Lean Zulip channel, the 7b model simply picked up on an issue in the Lean proof assistant v4.9 when it solved these 13 Putnam problems the bigger model could not solve. Basically, some 'apply?' leads to a proof silently being sorried which is not shown by Lean in the final result. This was fixed in later Lean versions. So the 7b model is simply reward hacking these 13 problems. 49 is the final result, NOT 62.
Now we need the same thing with assumed facts in generated text responses. How do you verify a statement? How do you train it to say “I’m not sure, let me check” instead of stating everything full chested as fact
Rare sponsorship W
I fucking love that small clip you always use at the beginning on "before we dive into it", no sarcasm.
I love the ending Patreon names
I think it is safe to say that the DeepSeek team consists of the smartest researches in the world. These guys just hit it out of the park every single time.
Unlock the Data Inside
Turn Videos into Knowledge
- Get FREE 10/day: transcripts, summaries, chats
- Chat with videos, export text & PDF
- $1 free API credit for RAG, chatbots & research
Free forever plan • All features unlocked
Top Comments (10)
Seems like this method of recursively breaking down problems into smaller sub-problems until triviality is achieved could be applied to many problem domains, not just math.
hey cloud can you cover every deepseek paper because these people are brilliant
I love seeing open source winning.
3:58 the “there exists delta” and “for all epsilon” need to switch places
There's a significant mistake in the video (though admittedly not just in the video, which stems from the original paper). As discussed in the Lean Zulip channel, the 7b model simply picked up on an issue in the Lean proof assistant v4.9 when it solved these 13 Putnam problems the bigger model could not solve. Basically, some 'apply?' leads to a proof silently being sorried which is not shown by Lean in the final result. This was fixed in later Lean versions. So the 7b model is simply reward hacking these 13 problems. 49 is the final result, NOT 62.
Now we need the same thing with assumed facts in generated text responses. How do you verify a statement? How do you train it to say “I’m not sure, let me check” instead of stating everything full chested as fact
Rare sponsorship W
I fucking love that small clip you always use at the beginning on "before we dive into it", no sarcasm.
I love the ending Patreon names
I think it is safe to say that the DeepSeek team consists of the smartest researches in the world. These guys just hit it out of the park every single time.