Taj You_Know

AI Engineer. Built and shipped products across startups and technologies before the hype. Now helping businesses build better software and put AI to work.

[Taj You_Know @Taj_youknow]

@nizamdesign This are looking too good impressive

[Taj You_Know @Taj_youknow]

@HeyAliux I would build a platform that optimizes the AEO and also pushes your content top to AI search results.

[Taj You_Know @Taj_youknow]

@HosseinGorji021 Sure lets build something meaningfull

[Taj You_Know @Taj_youknow]

@IanMcdevit20439 we are cooking we jev and jev is cooking frontier models

[Taj You_Know @Taj_youknow]

@0xMortyx the useful part is the split. Jev only does Choice / Score / Null. everything else stays in code and the LLM. I walked that generate vs decide piece here: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@ClementDelangue Jev is just the loud example: same softmax, tiny label space, probability out. that’s why a specialized model can be 100x cheaper here.
I sketched that generate vs decide split here: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@communicating agree on the system-one framing. hosted or local, the useful unit is “decision,” not “paragraph.” I tried to make that concrete without the cheaper-ChatGPT take: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@JamesWard building system two on top of system one is a good way to feel the constraint. Jev won’t write the next token, it will only pick. I used that generate vs decide split here: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@souorica this is the clean exec frame. the missing piece for a lot of teams is why it’s fast: it isn’t writing tokens, it’s scoring a few labels off shared state. I broke that part down here: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@saumya_writes not a chatbot. you give it state + a question with fixed options, it returns a probability. no sentence out. 2 min version: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@hbouammar @ilijabogunovic @typesafeai it doesn’t invent a new option, it just scores the ones you gave it. more than a toy classifier because the state can be messy, but it is still a constrained decision head. I put the softmax + parallel questions part on screen here: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@RobertSagurton @typesafeai some of it is launch noise, but the architecture split is real. LLM loops tokens. Jev answers a question you already defined. that’s why the TL feels sudden. short breakdown: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@OpenRouter This chart makes more sense once you stop treating Jev like a chat model. most of that spend is “pick / route / score,” not write. walked through the pipeline difference here: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@willdepue @natfriedman yeah that’s the part people keep missing. same softmax, just a tiny output space instead of 100k tokens. generate vs decide. I sketched that split here if it helps: https://t.co/fhOpMapbAu

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@itsalicesoul Overthinking that I can post tomorrow 😂

[Taj You_Know @Taj_youknow]

I made this video to walk through each point — not the “Jev is just a cheaper ChatGPT” version YouTube keeps posting.
LLM: predict the next token. Write. Loop.
Jev: look at the state.
Answer a question you already defined.
Probability out. No sentence.Same softmax. Different job.
One scores ~100k tokens.
The other scores 3 labels.That’s the fundamental split.

Generate vs decide.
I explain every step on screen.
#Jev #LLM #TypeSafe #SystemOne #AIAgents #Founders

[Taj You_Know @Taj_youknow]

@Tech_p001 Congrats bro

[Taj You_Know @Taj_youknow]

Are you still using RAG as the default for agent memory, or have you moved toward structured state + event-based memory for production systems?

#AI #RAG #software

[Taj You_Know @Taj_youknow]

@ardent__dev Fable