Yosef Shammout

cs student.

[Yosef Shammout @Yosef_Shammout]

@sunghopark20124 hey sungho nice to meet you, i’m yosef, undergrad at wayne state. having a poster at a NeurIPS workshop this year on knowledge graphs, and lately i’ve been getting into memory for LLMs and agents.

[Yosef Shammout @Yosef_Shammout]

@W20000Anonymous 125

[Yosef Shammout @Yosef_Shammout]

@harjjotsinghh @gdb I’m rolling

[Yosef Shammout @Yosef_Shammout]

If you’re interested, check it out: https://t.co/nAwZ4N1U3y

[Yosef Shammout @Yosef_Shammout]

We know that most language models write one word at a time, from left to right, and once a word is generated it’s set in place. I find it pretty interesting that it doesn’t actually have to work that way. I recently came across a paper from Stanford called Next-Concept Modeling that generates text in layers instead. It starts with a rough version of the whole sentence and sharpens it into exact words over a series of passes. For example, when it translates a sentence, the early passes get the rough shape down and the later passes fill in the details and clean up the wording. It also came out a lot faster than similar models at about the same quality, though it’s only been tried on small models so far, so I’m curious whether it holds up at scale.

[Yosef Shammout @Yosef_Shammout]

@mattbergland grinding sir, grinding

[Yosef Shammout @Yosef_Shammout]

@lukecodez git commit -m “….”

[Yosef Shammout @Yosef_Shammout]

@astrodanish This is hard

[Yosef Shammout @Yosef_Shammout]

@kanis_patel Could you share?

[Yosef Shammout @Yosef_Shammout]

On long tasks, an agent's history can outgrow its context window. Context compaction handles this by summarizing earlier messages, but a summary can keep the conclusion and drop the reason behind it. For example, an agent might remember that an approach failed and forget that it only failed because a tool was misconfigured, so it never goes back to an approach that would have worked. I think graphs are worth exploring more in agent memory, especially as a way to keep each decision linked to the evidence behind it and to any corrections that come later.

[Yosef Shammout @Yosef_Shammout]

Pop smoke was peak, we need another

[Yosef Shammout @Yosef_Shammout]

I get that SI companies are dealing with expensive GPUs, power costs, and the need to keep improving models. At the same time, usage limits on the cheaper plans ($20 range) still feel too restrictive for regular use and spinning up ideas. A $20 plan should get people meaningfully closer to the capacity of today’s $100 tiers.

[Yosef Shammout @Yosef_Shammout]

yes bro how'd you know i use grok cli

[Yosef Shammout @Yosef_Shammout]

Remote control has to be top 3 features ai companies have released

[Yosef Shammout @Yosef_Shammout]

this multi-agent RL paper is one i constantly refer back to whenever i think about where multi-agent systems are headed. the second multiple agents actually learn or adapt their policies together in a shared environment, standard single-agent optimization breaks down from non-stationarity, moving targets, and credit assignment traps. a lot of work is being done in this space and i think it should continue to be a major priority.

if you're interested in this space, ive linked a foundational paper below
https://t.co/b3b4J2B75M

[Yosef Shammout @Yosef_Shammout]

some of the highest quality of work from ai models comes from multi-shot examples. dont deploy your agents for real work until you've evaluated the key principles and approaches for your specific task. just clicking plan is not enough. this mainly applies when you're creating a product from scratch.

[Yosef Shammout @Yosef_Shammout]

Has anyone used gemini 4 argon? What do you guys think, debating whether to upgrade my gemini account.

[Yosef Shammout @Yosef_Shammout]

LLMs need to stop assuming background context, a constant issue I've seen that needs to be addressed.

[Yosef Shammout @Yosef_Shammout]

This is peak