A new AI product launches that sparks a market panic. It’s a new initiative, or an agent or something. Whatever it is, I can tell you one thing it can’t do.
Even with the best AI model in the world, the one thing it can’t do is its best work for you.
How do I know? Because the state of frontier models has gotten so advanced that to get the best talents out of any of them, you need different models to work together. The biggest weakness of Claude is that it only coworks with Claude
No one model family can do its best work for you without the talents of others. As models get more powerful, we’re seeing them specialize. The future state of AI is your best work will get done when there are different models working together.
In AI we call this “orchestration,” or sometimes, “the harness.” Whatever you want to call it, the superior performance of multi-model orchestration is vastly far ahead of anything any individual family of models can do.
Today we launched a product that is massively multimodel. 19 models are available in the backend, and the outputs have impressed me more than anything in AI for a while. But here’s how it was born.
In the beginning, we named our experiment ASI. It didn’t really stand for anything, but sometimes we called it “Artificial Super Intelligence.” It could’ve meant, “Another Slack Integration,” because that’s where it lived at first.
We thought of it as a digital worker. You message the worker in Slack with a task, and it goes to work. It delegates, creating subagents for different tasks, making files to help itself with the next step, finding vital information, and occasionally checking in, only as needed.
You could go to sleep and wake up with weeks of work done. Or, you could ask it to work for weeks, and it would reliably and repeatedly continue.
During this time, the world discovered Clawdbot, now OpenClaw. It was interesting, if you like malware reading your texts. We could see that ASI was safer and far more capable: it had a file system, connected to hundreds of tools, could read and write into files, write code, use CLI tools, browse the internet, and was capable of orchestrating every task to the right sub-agents which use the best model suited for that task. All within a secure development sandbox on the cloud.
In fact, ASI wasn’t really a remote digital worker. With access to a file system, a shell to execute code securely, and a browser to access web apps and connectors, ASI was essentially a computer. Your computer.
In 2011, Google had the idea that you could launch a computer with just the internet, files, and a bash terminal. It was called the Chromebook.
Google was right to understand the most important part of the computer is that it has web access. The web is where we work and communicate. It’s also our storage device, literally and figuratively. It holds vast amounts of data, information, and knowledge. Some shared, some personal.
If you think about the web as a storage system of knowledge, it’s always been very good at the WRITE function. The first blog was published in the early 90’s, and publishing and storing knowledge on the web has exploded ever since.
What Google underestimated with the Chromebook is that the web’s biggest problem as a storage system is actually the READ function.
To read knowledge from the web, we’ve relied on algorithmic search that’s more or less unchanged for 28 years. Search engines, as a way to access knowledge, have never been capable of deep research, synthesizing, or summarizing. Nor can search engines behave like an agent, to access any knowledge on the web that requires action to access.
Then came AI. Perplexity built the first answer engine, and the rest have followed. We’ve stayed focused on accurate AI, with intense focus on Deep Research. That’s because when AI can find, analyze, reason, and accurately present all of the knowledge on the open web, the READ problem is finally solved.
If the internet is the storage disk of the earth’s knowledge, Perplexity solved the read function.
Once the web is a fully functional read/write storage system, with agents that can navigate and accomplish tasks, the only remaining thing the computer needs is everything that’s personal. Persistent memory makes a computer personal. It has files and tools, and they’re your files and tools, private and tailored to your preferences. The computer must be able to work with them for you, personally.
Computers also create. They make real artifacts like code, documents, and files. They store the files, retrieve the files, write new files, access the files. AI can do all of these things.
Finally, when AI is doing all this work, you need different frontier models on hand for different jobs in the computer. Any computer that’s useful requires too many different skills for a single model to do well at all of them. Orchestration is essential to the computer.
It’s possible this means the hard code of the operating system becomes less of a moat for device makers as AI advances. I don’t know. What we do know is the computer is the orchestration system.
Steve Jobs said, “Musicians play their instruments. I play the orchestra.”
AI is similar. A Stradivarius has certain traits, and we’re amazed by its sound and craftsmanship. But it still can’t shake you like a bass drum, and the drum can’t move you like Mahler's Symphony.
Model makers are the Antonio Stradeveri’s of our time. I don't mean this in a disrespectful way. I’m in awe of every new model capability and a passionate supporter of the work at frontier labs. They push our future forward. But we now see the models are specializing, not commoditizing. When you build agents with them, they have different talents, like any worker. One is better at reasoning. Another at code. Another at creative writing. Etc.
What matters is how you orchestrate them.
Orchestrating across models and families is the only way to build a system versatile enough to handle real work. When you need deep research to inform a decision, you want a model optimized for accuracy and retrieval. When you need to write code, you want a model trained on billions of lines of production codebases. Whether you are managing a team of people, or a symphony of musicians, high quality work has always been this way. AI agents are no different.
For workflows, the differentiator is the harness. A planner breaks down complex objectives into subtasks. An orchestrator assigns those tasks to the right agents, enforces rules, and manages execution. Specialized sub-agents perform focused actions. Shared memory stores context and learnings for continuity. In other words, the intelligence is the computer.
400 years ago, a “computer” was an astronomer's apprentice. (The people who named comets).
Then, the computer was mechanical, then electromechanical, then digital. Around 1974, GUI came along. (Graphical user interface). With GUI, we could work with the computer in new ways. We could delegate to it and see its work. Later, the screen folded over the keyboard to be a portable, personal computer.
It’s a meme to say the next evolution of the computer is when Steve Jobs put the personal computer in everyone’s pocket. That’s dumb. It’s much more accurate to say he put a GUI in everyone’s face. Everyone is scrolling.
Over the last 49 years, the GUI has evolved from how we control the computer to how the computer controls us.
Exhaustion with this model is growing. The mad rush for Mac Mini’s during the OpenClaw fad was a perfect example. GUIs are necessary, but they’re not always necessary. AI models have finally become advanced enough to return the computer to its true meaning: to work autonomously, accurately. And asynchronously, to free us from watching it work.
The mouse, GUI, and keyboard are all important technologies, but they are interfaces. And unless you want a full self-driving horse, you have to admit the interface has never mattered as much as the computer.
AI models are becoming so capable that the products built around them have been a bottleneck for showing their true potential. The chat UI is good for answers, and agents are good for individual tasks. Meanwhile, the UI for entire workflows has always been the computer.
AI is now firmly multi-modal. It understands and generates many forms of data in a single coherent system. Jensen and others have said the future of AI must also be multimodel. They’re right, specialized models must collaborate like a team.
It may be more accurate to say the future of AI is massively multimodel. As AI replaces more of the function of the computer itself, the central activity of the computer will be massively multimodel orchestration.
We are only at the beginning. In 2025, a new frontier model entered the market, on average, every 17 days. With every new capability advancement in AI, another musician walks onstage to take her seat in the symphony. The conductor taps his baton. The computer goes to work.
https://www.perplexity.ai/try-computer
(for accessibility, this text is a copy of the Twitter Articles written by Aravind Srinivas, CEO of Perplexity)