Kayode Samuel

Software Developer | DevOps engineer

[Kayode Samuel @samkaypro]

We need to stop judging AI agents by the model alone. The harness matters just as much, if not more.

[Kayode Samuel @samkaypro]

Or San Francisco you are going to get an offer the same day

[shafu @shafu0x]

finding a job in tech is easy just move to new york

[Tesla @Tesla]

Real life, baby

[Kayode Samuel @samkaypro]

They call it iPhone Duo

[OpenAI @OpenAI]

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

[{Bhide}.Svelte @Sikandar_Bhide]

wtf this is Crazyyy

https://t.co/hXg4nAE33r
https://t.co/GOpadcz5Dg
https://t.co/3U0lBx4S4K
https://t.co/tQJwaFr8UO
https://t.co/SF7PKf53Wy
https://t.co/q3vjOc578t
https://t.co/42ag2nFcFL

[Kayode Samuel @samkaypro]

@Papacryptoo Your dm is closed how can I reach you?

[SpaceX @SpaceX]

Watch Falcon 9 launch 24 @Starlink satellites to orbit from California https://t.co/Lu37B7JGCB

[Kayode Samuel @samkaypro]

One day I’ll get a bill like this

[jay @jayvraavi]

ouch

[Kayode Samuel @samkaypro]

Check out my latest article: Change Is Noiseless. By The Time You Hear It, It Has Already Happened. https://t.co/H7otsnYi95 via @LinkedIn

[Kayode Samuel @samkaypro]

If you don’t go with the change, change will abandon you.

[Kayode Samuel @samkaypro]

Humans are going to replace Humans with AI, adapt now before it’s too late.

[Kayode Samuel @samkaypro]

Looped language models do something subtle:
Each hidden state gets decoded for a prediction AND fed back into future computation. The model carries forward its internal reasoning, not just the token it emitted. It’s the difference between thinking out loud and actually remembering what you thought.

[Kayode Samuel @samkaypro]

We’ve been wasting the most valuable thing a transformer produces.
Every step, the model builds a rich hidden state, decodes one token from it, then throws the state away. A new paper on looped language models feeds that state back in as runtime memory instead.

[Kayode Samuel @samkaypro]

For coding agents, context rot is the #1 failure mode. Not reasoning. Not capability.
The model is smart enough to solve the task. The problem is the context doesn't stay clean. Search, exploration, backtracking, all of it piles up noise that poisons every later step.
Clean context beats big context.

[Kayode Samuel @samkaypro]

I personally think any model can be good at reasoning, it just needs to be trained to ask better question and verify its answers.

[Kayode Samuel @samkaypro]

Speculative decoding is finally boring enough to be production-ready.
Pair a heavy target model with a lightweight drafter, let idle compute predict ahead, verify the batch in one pass.
If you’re running local/on-device LLMs and not doing this yet, you’re leaving 2-3x on the table.

[Kayode Samuel @samkaypro]

Everyone’s still optimizing prompt tokens while the actual latency win is at the decode layer.

Gemma 4’s MTP drafters get 3x faster generation with zero quality loss because verification happens in parallel.

Inference is memorybandwidth bound, not compute bound. Build accordingly.

[Unsloth AI @UnslothAI]

Gemma 4 now runs 2x faster with MTP GGUFs! Run locally on just 6GB RAM. ⚡️

MTP enables Google Gemma 4 run ~1.4–2.2× faster with no accuracy loss.

Gemma 4 12B MTP can run at 162 t/s vs. 52 t/s without MTP. 31B reaches 101 t/s.

GGUFs + Guide: https://t.co/c4gAUlb6YE

[Kayode Samuel @samkaypro]

@zamoose Check your message request

[Kayode Samuel @samkaypro]

@ForexTijoriHQ Your dm is closed can you please send me a message.