hammad 🔍

Last year at @tryramp I laid out three predictions for how language models would evolve. I was trying to clarify which bets might actually be durable over time. A lot of it is now starting to take shape. Here’s an update. Thread 👇

"Glad tidings to he who knows his own faults more than other people know it." — Ibn Hazm al-Andalusi

@trq212 Eid Mubarak!

Eid Mubarak! a day of celebration after a month of reflection

@scottbelsky @Suhail nuance - For anything custom - dedicated models, LoRAs, training etc. they are not allocating capacity. which is what I assumed op was interested in. this is 100% true. Sure “run existing model” is accessible.

@HammadTime @Suhail not true - they just signed deal with AWS and Ramp reports they are among the fastest growing spend across their card holders just a week ago

I am now at 5 GPU providers being completely sold out for a single node of 8xH100s. I don’t think people understand the gravity of what is about to come.

@scottbelsky @Suhail cerebras is turning people away because openAI bought all their capacity

@Suhail Cerebras just in time for this

dudes be like frontier models are really good at filesystems but not search because they were post trained on them my brother in christ they were postrained on both

I think the CPU/Motherboard analogy is funnily the most accurate. Subagents are forming the periphery. Which can give you durable market position?

@HanchungLee @code_star Yep

@code_star felt like this is a extremely specialized tiny asic / controllers. so in terms of computers, this is vrm or ec on motherboard. not as core as pch or usb/storage.

I have no idea what specialized model for context compaction means and I have like 5 papers and announcements to read before I can think about this. It’s crazy that for even a single model we may have a whole ecosystem of specialized models for optimization. Spec decode model, compaction. What comes after that?

@julianlehr 100%

A hill I'll die on: Current LLM chat interfaces are a regression from GUIs. Actions that used to be links, buttons, or keyboard shortcuts are now things I have to spell out in conversation. Why?

@thdxr perhaps a false dichotomy

i think a elon type of company is better suited to make inference 10x cheaper than make a frontier model

@r0ck3t23 yep - x.com/HammadTime/sta…

Perplexity CEO Aravind Srinivas just shattered the greatest illusion of the AI arms race. The entire market is waiting for a single, god-like superintelligence to win the entire board. The physics of compute are forcing the exact opposite outcome. Models are not converging into a single monopoly. They’re violently fracturing into hyper-specialized execution nodes. Srinivas: “Towards the end of 2025, what happened was models started specializing. Even within coding, which you think might be a specialization, OpenAI’s Codex models and Anthropic’s Claude models are very different in terms of what they’re good at.” Bet your entire enterprise architecture on a single AI provider? You’re hardcoding your own ceiling. You don’t want a generalized model that’s “okay” at everything. You want a swarm of apex specialists. One ruthlessly optimized for syntax. One for visual synthesis. One for predictive reasoning. The future is not one AI. It’s the instantaneous orchestration of the absolute best compute for the exact task at hand. Platform lock-in is suicide. Srinivas: “Enterprise users are always selecting multiple different models all the time. That’s actually one of the value propositions of the Perplexity product. You don’t have to feel locked into one model provider, you don’t have to have one horse in the race.” Traditional tech giants are desperately trying to trap users inside their specific algorithmic ecosystem. Winning operators completely bypass the vendor war by becoming model-agnostic. When the foundational intelligence of the world is leapfrogging itself every three months, brand loyalty is a massive liability. The operators winning the next decade won’t care whether OpenAI, Anthropic, or Google trained the model. They’ll plug into an agnostic orchestration layer that autonomously routes to whichever specialized network currently dominates that exact sector of the board. The highest-leverage position is no longer building the intelligence. It’s directing the orchestra. Srinivas: “This is one particular skill, writing is another skill, being good at images and videos is another skill. You can hope that Perplexity figures out which model is best for what purpose, and you just have to come to the product and use it.” Multi-trillion-dollar hyperscalers burning billions fighting the model wars. Sovereign orchestrator bypasses the entire war. Harvests the output of all of them. You don’t need to be an expert in the underlying architecture of a dozen different foundation models. You just need to command the routing engine. When AI transitions from a monolithic product into a fractured grid of specialized utility nodes, the ultimate monopoly belongs to the orchestrator that abstracts the complexity. Foundation model builders became interchangeable plumbing.

the vibers have found the formalisms, I repeat the vibers have found the formalisms

me when I explain to my wife how I subtly prompt the 14 Claude codes I have running in parallel work trees

@mktpavlenko @trychroma 100%

@trychroma The underrated part is isolation, not just speed. You can test new retrieval or ranking ideas on the same data without gambling with the prod collection.

Collection Forking on Chroma Cloud unlocks faster workflows on top of your data without the overhead of starting from scratch.

fork fork fork fork fork fork fork fork fork fork fork

i revisit this every so often as the most glaring reminder of how many technology dichotomies are not a question of if but when.

people.csail.mit.edu/tzumao/diffvg/ Time to bring it back

LLMs are so good at weird things

me: "can you use whatever resources you like, and python, to generate a short 'youtube poop' video and render it using ffmpeg ? can you put more of a personal spin on it? it should express what it's like to be a LLM" claude opus 4.6:

cool work