Eljan Mahammadli

Head of AI Provenance @ Polygraf AI | MSc CS @ George Washington University

[Eljan Mahammadli @eljanmahammadli]

@jerryjliu0 t3code

[Eljan Mahammadli @eljanmahammadli]

@theo Codex for GPT models, CLI for CC, since their desktop app is horrendous.

[Eljan Mahammadli @eljanmahammadli]

@Nerraograyghost Raycast

[Eljan Mahammadli @eljanmahammadli]

@rohanvarma For sure, rewind. Current forking is not enough.

[Eljan Mahammadli @eljanmahammadli]

@theo I believe you know how LLMs work under the hood, so I don’t think it’s insane. On the contrary, I’m amazed when it can format diagrams so well.
See this:
https://t.co/rrLGgPXn1g

[ahmet alp balkan @ahmetb]

How are LLMs are able to generate perfectly aligned ASCII diagrams/tables if they're only focused on the next token (and not seeing all lines at once)? What's the technical explanation of this?

[Eljan Mahammadli @eljanmahammadli]

@mntruell I would like a way for users to control the context window to some extent. Obviously, the agent manages it dynamically, but for advanced users, it would be helpful to have more control, especially when the agent might include information in the context that is not relevant.

[Eljan Mahammadli @eljanmahammadli]

@leerob @ericzakariasson Thanks for the clarification! I did notice a difference in token usage in the usage summary when files are open vs when I close them all, so that second point you mentioned likely explains it. Appreciate it 🙏

[Eljan Mahammadli @eljanmahammadli]

Is there any way to disable adding open files to the context in Cursor? Most of the time I have many files open, and I feel like this pollutes the context window. @ericzakariasson @leerob

[Eljan Mahammadli @eljanmahammadli]

@gsarayruhu1905_ Hocam bişeyler deniyo😂

[Eljan Mahammadli @eljanmahammadli]

@simonw Fine-tuning small models as “humanizers” can drop AI detection scores a lot. Prompting just simplifies, but fine-tuning can rewrite text without losing quality of text. Great for marketing or emails but arguably bad for academia since it can blur authorship and integrity.

[Eljan Mahammadli @eljanmahammadli]

@chribjel Hmm…

[Eljan Mahammadli @eljanmahammadli]

@ChujieZheng 4B & 8B

[Eljan Mahammadli @eljanmahammadli]

@ludwigABAP Same here. I have tried most browsers and still cannot find what I want :/

[Eljan Mahammadli @eljanmahammadli]

@AravSrinivas Not an issue but would be awesome if Perplexity streaming answers showed up right in Spotlight when you hit Cmd+Shift+P (similar to HuggingChat). No need to open the app, just stay in flow. I use the Raycast extension as a workaround for now.

[Eljan Mahammadli @eljanmahammadli]

@krishnanrohit I wonder if this is an emergent behavior due to the model’s size and complexity, or if it is explicitly a result of post-training.

[Eljan Mahammadli @eljanmahammadli]

@_chenglou Try sending random text. It tries to find some meaning/secret message behind it. So funny)

[Eljan Mahammadli @eljanmahammadli]

@teknium 🤣

[Eljan Mahammadli @eljanmahammadli]

@humanloop 👍🏻

[Eljan Mahammadli @eljanmahammadli]

@levelsio Which AI tool do you use to clone voice?

[Bojan Tunguz @tunguz]

When you’ve trained your LLM on 100,000 H100s for a couple of months and it returns all NaNs.