Manuel

Accelerating the adoption of AI at scale.

[Manuel @ManuelSH]

@levie That’s the Jevon’s paradox: “is said to occur when technological improvements that increase the efficiency of a resource's use lead to a rise, rather than a fall, in total consumption of that resource”

[Manuel @ManuelSH]

AI maturity comes in levels:
Level 1: Write better prompts. Iterate in the conversation.
Level 2: Context engineering. Ensure the model has the right info at the right time.
Level 3: Harness engineering. LLM has the right sources, tools, feedback loops, safeguards, etc

[Manuel @ManuelSH]

@theprototype003 @DotCSV Sí, yo lo veo así. El frontend lo generará la AI. De hecho, aquí hay una prueba de concepto...
https://t.co/klUfAjL9b6

[Zain Shah @zan2434]

Imagine every pixel on your screen, streamed live directly from a model. No HTML, no layout engine, no code. Just exactly what you want to see.

@eddiejiao_obj, @drewocarr and I built a prototype to see how this could actually work, and set out to make it real. We're calling it Flipbook. (1/5)

[Manuel @ManuelSH]

Such a good reflection

[Ethan Mollick @emollick]

Every AI discussion ultimately rests on two questions: how good can AI get? And how fast? They are predictions about the s-curve shape.

Everything else (job impact, potential risks, etc.) is downstream of those questions. I think it would be useful to focus on them more often.

[Manuel @ManuelSH]

Here it is: https://t.co/GKsPCgcb92

[Manuel @ManuelSH]

Wanted to update my CV... but ended up building an app that orchestrates multiple LLMs to generate a custom CV for each situation.

And of course I built it with an agnet: @OpenAI Codex.

[Manuel @ManuelSH]

@emollick The key that will change everything is continuous learning.

https://t.co/7UnlQwQn6S

[Manuel @ManuelSH]

In the future, we’ll remember context windows as a strange, almost funny constraint we once had to live with.

[Manuel @ManuelSH]

In the future, we’ll remember context windows as a strange, almost funny constraint we once had to live with.

[Manuel @ManuelSH]

OpenClaw seems one, but looks like a proof of concept. ChatGPT and Claude has 3 or 4 of all the points above, but not all. Same with Claude Code and Codex. But still not a finished product.

[Manuel @ManuelSH]

The most obvious AI product still doesn’t exist:
a personal agent (1) connected to email, messaging, the internet, my files, my tools... (2) that keeps context about me (e.g. see LLM Wiki), (3) available from any interface, (4) proactive, and (4) secure. We have glimpses of it.

[Manuel @ManuelSH]

@philo_societycj @411emilio Same thing was used in ancient Rome, where daytime and nightime was divided in 12 “hours”

[Manuel @ManuelSH]

@OpenAI @nikey_es lo que hablamos justo hoy…

[Manuel @ManuelSH]

@javilop ¡Me alegra ver que “autorreason” parece funcionar!

[𒐪 @SHL0MS]

i've been working on a method called autoreason that is effectively autoresearch extended to subjective domains. autoresearch works because val_bpb gives you an objective fitness function. autoreason constructs a subjective one through independent blind evaluation, the same way science uses peer review where math can use proofs.

as you’ve noted, the fundamental problem with using LLMs for iterative refinement on subjective work: the model is always sycophantic when you ask it to improve something, overly critical when you ask it to find flaws, and overly compromising when you ask it to merge two perspectives. the output ends up shaped more by how you prompt than by what's actually better.

autoreason fixes this by separating every role into isolated agents with no shared context. you start by generating version A. a fresh agent attacks it as a strawman. a separate author who only sees the original task, version A, and the strawman critique produces version B. a third agent who has no history with either drafting process sees both versions as equal inputs and synthesizes them into version AB. a blind judge panel with fresh context and randomized labels picks the strongest of A, B, or AB. the winner becomes the new A and the loop repeats until the judges consistently pick the incumbent which indicates that no further changes are needed.

[Manuel @ManuelSH]

@emollick Indeed. Agents can easily compound LLM hallucinations.

[Manuel @ManuelSH]

@karpathy @existeundelta Very nice! In some ways is similar to what I am building, with more human curation:

https://t.co/JQEkUW1f54

[Manuel @ManuelSH]

Almost everyone wanted Cursor. Very few wanted Claude Code. Our evaluation pointed the other way: Claude Code performed best. Good reminder that in AI tooling, popularity is not the way to choose.

https://t.co/Sn8Cqktii0

[Manuel @ManuelSH]

@DotCSV 14 feb 2019. Me dije: “ya está resuelto”.

[Manuel @ManuelSH]

A great resource…

[elie @eliebakouch]

Training LLMs end to end is hard. Very excited to share our new blog (book?) that cover the full pipeline: pre-training, post-training and infra. 200+ pages of what worked, what didn’t, and how to make it run reliably

https://t.co/iN2JtWhn23

[Manuel @ManuelSH]

@DotCSV Muy bueno y en total acuerdo. Gracias por compartirlo.