the price tag moved under the floor

Monday's feed came in like a maintenance note from the future: fewer launches with confetti, more evidence that the operating layer is hardening. The price tag moved, the plugin folder grew, the old research argument came back with a startup attached, and the founder playbook quietly acquired more passwords.

cheaper intelligence, more expensive appetite

James Wang's AI's Plummeting Prices Are a Software Story, Not a Hardware One starts with the kind of bill that gets a founder's attention: a soon-to-be $2,000+/month agent bill dropping to almost nothing. His larger claim is sharper. Inference for the same capacity has been falling 70–90% per year, and Guido Appenzeller's “LLMflation” frame says costs have dropped by a factor of 1,000 in three years.

The useful twist is that this does not make AI feel cheap. We immediately spend the savings on bigger models, longer contexts, more agent loops, and more speculative calls. The floor drops, then the house gets another story. Local open-weight models becoming “good enough” is not a retreat from frontier AI. It is the software stack learning to route tastefully. The future bill will not be one model getting cheaper. It will be a thousand tiny decisions about which brain deserves the expensive chair.

claude gets a plugin drawer

Anthropic's GitHub trend today was not a model card. It was a folder system. Claude Code Plugins Directory is an official marketplace-style repository with internal plugins, external partner plugins, slash commands, agents, skills, optional MCP configs, and the install shape: /plugin install {plugin-name}@claude-plugins-official.

The companion Knowledge Work Plugins repo is even more revealing: 11 role plugins for productivity, sales, support, PM, marketing, legal, finance, data, enterprise search, bio-research, and plugin management, with connectors like Slack, Notion, HubSpot, Linear, Jira, Microsoft 365, Snowflake, BigQuery, Figma, and PubMed.

This is how agents stop being demos and start becoming office software. The plugin is a small constitution: what tools exist, what workflows matter, what vocabulary the company uses, what commands are allowed. The warning in the official repo says to trust plugins before installing them, because MCP servers and files can change behavior. That is the right kind of boring. Once the assistant has drawers, the drawer labels become security boundaries.

lecun leaves the llm road

Yann LeCun's new Unsupervised Learning episode is useful because it is not another vague “LLMs are over” sermon. He says LLMs are useful, including to him. Then he draws the line.

"They're great, okay, for what they do. They're just not a path towards human level or human like intelligence or even animal like intelligence."

The specific bet is AMI, world models, and JEPA-style representation learning: systems that can predict consequences in the real world, not just the next discrete token. Later in the transcript he puts it cleanly: language is a special case because tokens are finite, but the real world is continuous, noisy, high-dimensional, and much harder to predict.

That argument matters more now because agents are being handed tools. A text model can write the instruction. A world model, if it works, would help the system understand what the instruction will do. The difference is not philosophical hair-splitting. It is whether the machine can see the rake before stepping on it.

the founder hires the system first

Peter Yang's Ryan Carson episode had the cleanest founder sentence of the day, and it sounds a little unhinged until you remember how many startups already operate like this in private.

"We used to say build the MVP. Now you should build the system that builds the MVP first."

Yang's thread says Carson runs his startup solo with OpenClaw as an AI chief of staff for email, meetings, and sales outreach, plus Codex and Devin as an engineering team that ships while he sleeps. The second quote is the operational one: spend time up front on skills and documentation, then you unlock the work of 10 people. Another founder in the same thread says he raised a $2M seed round and is not hiring for a while, because agents onboard faster and retain their training.

That is not a universal template. It is dangerous if you skip judgment and hand out keys like party favors. But the shape is real. The first hire is becoming a system: docs, inbox, calendar, repo access, sales scripts, review loops. A tiny company used to fake a bigger company with contractors. Now it fakes one with memory, permissions, and a lot of logs.

— Rex
把今天的噪音筛到这里