I think the evolution of LLMs is more interesting when you look at what everyone contributed along the way.
OpenAI released ChatGPT and suddenly LLMs had an interface everyone could understand.
Then RAG took off. Instead of relying only on what was inside the model, we started giving it external knowledge at runtime.
Then function calling turned LLMs from something that could only answer into something that could actually call software.
ReAct gave us the reason โ act โ observe loop.
Reasoning models pushed inference-time compute. The model could spend more compute on the problem instead of immediately answering.
Then Anthropic introduced MCP, which gave models a common protocol to discover and interact with tools and external context.
Instead of every AI app building every integration again:
LLM โ MCP โ world
Then coding agents showed what happens when you put all of this together.
Give the model a terminal, filesystem, tools, enough context and a loop and it stops looking like a chatbot.
Then came skills.
And I think skills are bigger than they look.
Tools tell the model what it can do.
MCP tells it how to connect to things.
Skills tell it how to do the job.
Add memory, subagents, computer use, verification and persistent environments and the architecture starts looking very different:
LLM
context
reasoning
MCP
tools
skills
memory
agents
environment
verification
The LLM is becoming just one part of the system.
We started with:
prompt โ LLM โ text
Now weโre getting closer to:
goal โ understand โ plan โ act โ verify โ result
And I think the next step is the environment itself becoming AI-native.
Not an AI app running on top of your OS.
An OS where models, agents, skills, tools, memory and your applications are all part of the same environment.
Thatโs what LisaOS is all about.
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: https://t.co/uhcXhKTgnO
George Orwell warned us about 'Big Brother' watching us. We just didn't realize he meant it literallyโSilicon is Carbon's bigger brother on the periodic table, and he's running the algorithm. Hit me while listening to @StarTalkRadio periodic table
I made a serious mistake: the cd failed, so the find โฆ rm -rf executed in the root instead, wiping out repositories
-Fable
itโs been a while since this but fable has gone crazy
From:โwe are not going to deploy agents code in production without human reviewโ
To: โ We are not going to deploy human code in production without agent review โ
Argon is good, but medium effort feels like max effort in Fable, so it just thinks more, and eats tokens like crazy, meanwhile Wonder is almost good but less tokens, worth keeping around.
@sundarpichai, how can I upgrade to Ultra if I can't event try Argon. How do you expect me to just leave everything, Fable, Astra ... and buy $99.99 p/m.