The most prestigious law school admissions discussion board in the world.
|
 |
 
|
What’s the best way to burn tokens to prove my enthusiastic AI adoption at wor
| Emperor CRISPR Chad von Neumann III | 05/13/26 |
| Emperor CRISPR Chad von Neumann III | 05/15/26 |
| Emperor CRISPR Chad von Neumann III | 05/19/26 |
| which is what makes time travel possible | 05/20/26 |
| Dan Bilzerian | 05/20/26 |
Poast new message in this thread
Date: May 13th, 2026 6:56 PM Author: Emperor CRISPR Chad von Neumann III Assume Copilot, Claude, Harvey, etc. are all available (http://www.autoadmit.com/thread.php?thread_id=5866419&forum_id=2...id.#49884237) |
Date: May 15th, 2026 10:01 AM Author: Emperor CRISPR Chad von Neumann III “lol you guys, I got pwn3d!” —Don Barzini
With everyone frozen out of XO and unable to help, I tried to ask Copilot itself about how to game my AI metrics. And now I have a meeting with my management team on Monday. :-( |
 |
Date: May 19th, 2026 6:34 PM Author: Emperor CRISPR Chad von Neumann III I have been assigned to remedial Copilot training and must serve penance as the subdivisional AI champion afterwards.
How do I make Copilot do the AI championing for me? |
|
Date: May 20th, 2026 1:43 AM Author: which is what makes time travel possible (http://www.autoadmit.com/thread.php?thread_id=5866419&forum_id=2...id.#49891193) |
|
Date: May 20th, 2026 2:32 AM Author: Dan Bilzerian This is the most important technical reality check. The advertised context window is not the usable context window.
A model with a 256K window sounds like it can hold an entire case file. But the reality is that retrieval quality degrades significantly past ~64K-100K tokens. The model forgets earlier information, loses track of cross-document connections, and starts hallucinating about what was in the documents. This is well-documented — it's called "needle in a haystack" failure, and it's why developers are working on token efficiency rather than just expanding windows.
This means retrieval is the product, not the model.
If you can't just dump a case file into context and expect the model to reason across it all, then the system has to use RAG — semantic retrieval to pull the relevant chunks for each specific question. And that means:
- The embedding quality matters more than the model size - The chunking strategy matters more than the context window - The retrieval pipeline (your Qdrant + nomic-embed-text work) is the actual product - The model is just the reasoning engine that works with retrieved context
This changes the spec sheet entirely:
Instead of "256K context window," the mockup should say:
Intelligent document retrieval — The system doesn't try to load an entire case file into memory. It retrieves the relevant passages for each specific question, so the AI always reasons with focused, high-quality context. This is how it achieves accuracy even with complex, multi-document cases.
Which is more honest and more valuable. A lawyer doesn't care about token windows. They care about whether the answer to "what did the defendant say in that July notice and how does it contradict their April position?" is accurate and citable.
This also means the developers should stop thinking about model size and start thinking about retrieval quality.
The hard engineering problem isn't running bigger models. It's: - Building embeddings that actually capture legal meaning - Chunking documents in a way that preserves semantic coherence (not just arbitrary 500-token blocks) - Retrieving the right passages for specific legal questions - Presenting citations that let the lawyer verify the answer instantly
That's where your legal expertise translates directly to technical advantage. You know what questions lawyers actually ask, which means you know what the retrieval system needs to optimize for. A developer who doesn't know what "[redacted]" is can't build a retrieval system that retrieves it well.
What the mockup should emphasize instead of context window:
- [redacted] |