Carlos Granier

25 years building at the edge of emerging tech. Today that edge is AI. Founder. CTO. Builder. Miami 🌴

[Carlos Granier @cgranier]

What happens when the agent picks the #aivideo model? Strong analysis of the orchestration landscape.

[RCTV @rctvdotcom]

The split that decides who wins: Adobe and Runway orchestrate only their own stack; Pika, Luma and Higgsfield route across everyone's. We make the call on which bet wins — and what "the agent picks the model" does to model pricing power: https://t.co/Si4kCGommB

[Carlos Granier @cgranier]

4/ A pattern named once costs nothing to refuse the next fifty times. That's the move.

Full piece: https://t.co/QHXlZ8bcH2

[Carlos Granier @cgranier]

The most useful page in my LLM Wiki isn't a record of what I've kept. It's a record of what I've decided to refuse. Same pattern applied to refusal. The refusal compounds the same way the memory does.

[Andrej Karpathy @karpathy]

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

[Carlos Granier @cgranier]

3/ The naive fix is "save less." That's wrong. Real signal sometimes hides inside the noisy genre. The actual problem is re-evaluation cost. Every templated post costs processing budget every single time, with zero compounding.

[Carlos Granier @cgranier]

3/ Generalizes: for any slash command with model-facing output, the diagnostic is the side effect, never the output appearance. The model can describe the operation; only the system can perform it.

Most operator-tier content focuses on content design. The plumbing is the next failure surface.

[Carlos Granier @cgranier]

1/ Just ran /compact on a Claude Code session at 60% context. Result: 609,664 → 8,963 tokens in 110 seconds. 98.5% reduction.
@realamrutpatil was right: "You are not running out. You are leaking." The leak is fully recoverable.

[Carlos Granier @cgranier]

2/ But the first attempt failed silently. /compact followed by a blank line breaks the slash-command parser. The model wrote a beautifully formatted summary as its reply (looked identical to what /compact would produce) but no compaction fired. Context didn't shrink.

The fix: /compact (literal space) + argument continuous, same line.

[Mark Kretschmann @mark_k]

It bears repeating: you cannot seriously be an AI and technology enjoyer while fully aligning with today’s left. The modern left has become the political home of anti-technology, deceleration, overregulation, and reflexive AI panic. Every major leap forward is treated as a threat first and an opportunity second.

AI should be seen as one of the greatest tools for abundance, medicine, education, productivity, and human flourishing. Instead, much of the left talks about it like it’s mainly a labor threat, a misinformation machine, or something that needs to be slowed down before it can change too much.

At some point, you have to choose: progress or stagnation. Acceleration or fear. Building the future or endlessly trying to regulate it out of existence.

[Brivael Le Pogam @brivael]

Sanders et AOC veulent geler la construction de tous les data centers IA aux États-Unis.

Il faut comprendre ce qui se passe vraiment. Ce n'est pas une bataille politique parmi d'autres. C'est la dernière convulsion d'une vision du monde qui a compris, inconsciemment, qu'elle est condamnée.

Le socialisme n'est pas une théorie économique. C'est une structure morale qui a besoin de trois choses pour exister :

1. De la rareté à redistribuer
2. Des victimes à défendre
3. Une classe d'intermédiaires pour orchestrer le tout

Retirez un seul de ces trois piliers et l'édifice s'effondre. L'IA est en train de retirer les trois en même temps.

La rareté d'abord. Pendant 200 ans, l'économie politique a tourné autour d'une question : comment répartir une production limitée ? Marx, Keynes, Piketty — tous bâtissent sur ce postulat. Mais l'IA inverse l'équation. Le coût marginal de l'intelligence tend vers zéro. La production de logiciel, de design, d'analyse, de code, bientôt de matière manufacturée par robotique avancée — tout cela devient quasi-gratuit. Dans un monde d'abondance, la question "qui mérite quoi" perd son sens. Il n'y a plus rien à arbitrer.

Les victimes ensuite. L'IA est le plus grand égalisateur d'accès au savoir et aux compétences de l'histoire humaine. Un gamin au fin fond du Bangladesh a aujourd'hui accès au même tuteur que l'héritier d'une famille new-yorkaise. Un développeur solo produit ce qu'une équipe de 20 produisait il y a trois ans. Les barrières s'effondrent. Or sans victimes structurelles, plus de cause à défendre, plus de mandat moral à exercer.

Les intermédiaires enfin. C'est le point le plus douloureux pour eux. Le socialisme a toujours eu besoin d'une caste : journalistes-militants, fonctionnaires-experts, ONG-prescriptrices, politiques-redistributeurs. Cette caste vit du fait qu'elle prétend traduire la réalité aux masses. L'IA rend cette traduction obsolète. Tout le monde peut interroger directement la source, vérifier un chiffre, comparer des modèles, simuler une politique publique. Le monopole de l'interprétation est mort.

Voilà pourquoi je dis que l'IA est un catalyseur de vérité. Elle ne crée pas la vérité — elle la rend ininterprétable. Les systèmes qui produisent de la valeur deviennent visibles. Ceux qui en captent sans en produire deviennent visibles aussi. Le voile tombe.

Et c'est ça qui est insupportable. Pas la perte de pouvoir — la perte de sens. Réaliser que ta vision du monde, ton militantisme, ta carrière entière reposaient sur un édifice qui ne tenait que par la rareté et l'opacité. C'est une blessure narcissique d'une profondeur abyssale.

La réaction est mécanique : il faut bloquer le catalyseur. Pas pour des raisons rationnelles (l'argument "énergie" est risible quand on voit leurs positions sur le nucléaire). Pour des raisons existentielles. Il faut empêcher l'avenir d'advenir, parce que l'avenir les efface.

300 lois locales. Un moratoire fédéral. Des moratoires européens (AI Act). Tout le pattern est le même partout : freiner, ralentir, encadrer, taxer. Pas réguler intelligemment — paralyser.

Mais ils ont déjà perdu. Et au fond d'eux, ils le savent. La Chine ne s'arrêtera pas. Les Émirats ne s'arrêteront pas. L'Inde, Singapour, l'Argentine de Milei, certains États américains — personne ne s'arrêtera. Bloquer la construction de data centers à San Francisco ne fait que déplacer le centre de gravité. Le seul effet net est d'appauvrir ceux qu'ils prétendent défendre.

C'est le rebond du chat mort. Un dernier sursaut avant l'immobilité définitive.

PS : tout n'est pas perdu pour eux. La porte est ouverte. Il suffit de comprendre que créer de la valeur est plus gratifiant que la redistribuer, que construire est plus puissant que dénoncer, et que l'entrepreneuriat est la seule forme contemporaine d'action politique qui change réellement le monde. La reconversion est possible. Elle commence par accepter une chose simple : personne n'a besoin de toi pour être sauvé. Mais beaucoup de gens ont besoin de toi pour construire.

[Carlos Granier @cgranier]

Worth reading Tan's thread alongside Brian Schimpf's Anduril investor letter. Same structural argument across two verticals:

@SchimpfBrian: "The defense industrial base was decimated by the consolidation and offshoring of the 1990s and 2000s… production capacity itself is a strategic asset."

@garrytan: "The politicians ... are destroying jobs and American competitiveness at the global level."

Defense infrastructure. AI infrastructure. The lesson runs through both; the side that translates technological advancement into deployable capacity at scale wins. The bottleneck is the same.

[Garry Tan @garrytan]

Sanders and AOC introduced a bill to pause ALL AI data center construction. 300+ local bills filed. Half of planned 2026 data centers facing delays or cancellation. Each one brings billions to local economies.

The people who say they want American jobs are trying to block the biggest job creation engine since the interstate highway system.

[Carlos Granier @cgranier]

Communism at work.

[Garry Tan @garrytan]

Sanders and AOC introduced a bill to pause ALL AI data center construction. 300+ local bills filed. Half of planned 2026 data centers facing delays or cancellation. Each one brings billions to local economies.

The people who say they want American jobs are trying to block the biggest job creation engine since the interstate highway system.

[Carlos Granier @cgranier]

🚨 Tired of agents torching your X API budget?

Built xapi — a local cost-aware HTTP proxy that actually measures the burn:

• $0.005/post (reads)
• $0.01/user lookup (2× the post rate)
• $0.001 for your own bookmarks
• OAuth 1.0a silently rejected by v2 endpoints

Cache + ledger + budget gate, runs on localhost or your tailnet.
https://t.co/axvRgnrgEg

[Carlos Granier @cgranier]

@johnennis You're welcome. Wonderful piece and insights. I've noticed the same... AI has allowed me to work on so many more problems, each of which opens up many more.

[Carlos Granier @cgranier]

"The machine handles the production. The human handles the judgment."

Ennis works from sensory science and embodied cognition to a thesis the AI industry mostly hasn't internalized. The cleanest articulation I've read of what work actually looks like now. Worth your time.

[John Ennis @johnennis]

https://t.co/b89862LIqj

[Carlos Granier @cgranier]

How I noticed: I run a Karpathy-style LLM wiki across six project lenses. The wiki had been collecting these saves for weeks. Today's OpenAI announcement landed alongside Isenberg's AI-native estimate and the pattern crystallized. Five sources from four lenses. A single-lens second brain would have seen the items. The wiki saw the pattern.

I don't have to be smart in real time. The wiki does the work.

Follow-up article in a few days.

[Carlos Granier @cgranier]

The integration layer just became the real moat in enterprise AI.

Five deployment stories in the last month. Four of them this week. All converging on the same operational moment from four incompatible incentives.

[Carlos Granier @cgranier]

Structurally, spending money on frontier models is the easy half. The harder, more durable investment is making the company itself legible to agents.

Documented workflows, structured records, decision systems, evaluation loops. As Isenberg framed it: AI-native isn't a tech label. It's an organizational one.