I think office is copilot now? Who knows lol
January 15, 2026 at 3:32 AM
I think office is copilot now? Who knows lol
The harness is the agent implementation that controls what gets fed into the LLM. The app around the ai magic black box. Claude, codex, copilot are all examples.
Gh is GitHub. Copilot has limited ability to find code if it’s not manually fed into the chat box
Gh is GitHub. Copilot has limited ability to find code if it’s not manually fed into the chat box
January 15, 2026 at 3:31 AM
The harness is the agent implementation that controls what gets fed into the LLM. The app around the ai magic black box. Claude, codex, copilot are all examples.
Gh is GitHub. Copilot has limited ability to find code if it’s not manually fed into the chat box
Gh is GitHub. Copilot has limited ability to find code if it’s not manually fed into the chat box
Yeah the copilot harness has some really bad gaps (it can only grab 10 records from a gh search for example)
If you’re stuck with it you can get it to do stuff if you manually curate the context carefully and it’s still quicker than the alternative, but it’s a painful experience imo
If you’re stuck with it you can get it to do stuff if you manually curate the context carefully and it’s still quicker than the alternative, but it’s a painful experience imo
January 15, 2026 at 3:27 AM
Yeah the copilot harness has some really bad gaps (it can only grab 10 records from a gh search for example)
If you’re stuck with it you can get it to do stuff if you manually curate the context carefully and it’s still quicker than the alternative, but it’s a painful experience imo
If you’re stuck with it you can get it to do stuff if you manually curate the context carefully and it’s still quicker than the alternative, but it’s a painful experience imo
It’s mind bogglingly good! Really moves the challenge away from “how do you build” to “what do you build”.
Whether it can be made available to non-computer-power-users is an interesting challenge.. it’s a table saw and you *can* cut your arm off with it.
Sandboxing and tooling are tricky
Whether it can be made available to non-computer-power-users is an interesting challenge.. it’s a table saw and you *can* cut your arm off with it.
Sandboxing and tooling are tricky
January 13, 2026 at 4:54 PM
It’s mind bogglingly good! Really moves the challenge away from “how do you build” to “what do you build”.
Whether it can be made available to non-computer-power-users is an interesting challenge.. it’s a table saw and you *can* cut your arm off with it.
Sandboxing and tooling are tricky
Whether it can be made available to non-computer-power-users is an interesting challenge.. it’s a table saw and you *can* cut your arm off with it.
Sandboxing and tooling are tricky
Yeah that’s what I was getting at. They’re obvs unprofitable on net but I *think* the per unit economics are also appallingly bad lol.
Selling more AI doesn’t help without changes to cost model or pricing
Selling more AI doesn’t help without changes to cost model or pricing
November 6, 2025 at 2:35 AM
Yeah that’s what I was getting at. They’re obvs unprofitable on net but I *think* the per unit economics are also appallingly bad lol.
Selling more AI doesn’t help without changes to cost model or pricing
Selling more AI doesn’t help without changes to cost model or pricing
Fairly heavy subs for OpenAI I think.. medium.com/@furqankhaan...
Not clear they’re profitable on those subs on the margin tho. Running the model gets pricy for heavy users
Not clear they’re profitable on those subs on the margin tho. Running the model gets pricy for heavy users
How OpenAI and Anthropic Are Cashing In on AI: A Look at Their Revenue Models
As artificial intelligence technologies expand in both functionality and market reach, companies like OpenAI and Anthropic lead the charge…
medium.com
November 6, 2025 at 1:13 AM
Fairly heavy subs for OpenAI I think.. medium.com/@furqankhaan...
Not clear they’re profitable on those subs on the margin tho. Running the model gets pricy for heavy users
Not clear they’re profitable on those subs on the margin tho. Running the model gets pricy for heavy users
Linear algebra presumably felt incomprehensible until we had a framework to break things into subspaces etc.
Do we think we’re just lacking something analogous to decompose LLMs into something that can be reasoned about or is it simply not possible to do so?
Do we think we’re just lacking something analogous to decompose LLMs into something that can be reasoned about or is it simply not possible to do so?
July 11, 2025 at 3:12 PM
Linear algebra presumably felt incomprehensible until we had a framework to break things into subspaces etc.
Do we think we’re just lacking something analogous to decompose LLMs into something that can be reasoned about or is it simply not possible to do so?
Do we think we’re just lacking something analogous to decompose LLMs into something that can be reasoned about or is it simply not possible to do so?
Pretty much, yeah! I think mcp might let you work around the mdc vs Claude.md problem - some format for @rules projects and a parsing engine that plugs into an mcp server you can integrate with the llm would would support a pretty arbitrary agentic system I think. Devil’s in the details though lol
June 3, 2025 at 4:10 PM
Pretty much, yeah! I think mcp might let you work around the mdc vs Claude.md problem - some format for @rules projects and a parsing engine that plugs into an mcp server you can integrate with the llm would would support a pretty arbitrary agentic system I think. Devil’s in the details though lol
Ah yeah that’s a much harder problem. Maybe a standard package you could pull in with the react library to get tacked onto prompts via mcp or what have you. Kinda analogous to a types package today
June 2, 2025 at 10:52 PM
Ah yeah that’s a much harder problem. Maybe a standard package you could pull in with the react library to get tacked onto prompts via mcp or what have you. Kinda analogous to a types package today
I’ve had decent luck with this using cursor, so I’d imagine Claude will do it as well. If you provide enough detail in your prompt on code structure and best practices it seems to behave.
Cursor uses mdc files to say “when in data layer use these libraries” type stuff so I think Claude would
Cursor uses mdc files to say “when in data layer use these libraries” type stuff so I think Claude would
June 2, 2025 at 10:43 PM
I’ve had decent luck with this using cursor, so I’d imagine Claude will do it as well. If you provide enough detail in your prompt on code structure and best practices it seems to behave.
Cursor uses mdc files to say “when in data layer use these libraries” type stuff so I think Claude would
Cursor uses mdc files to say “when in data layer use these libraries” type stuff so I think Claude would
Yeah that’s fair! Discover is an essential part of social media. I have yet to hear an effective solution for disinformation that preserves the value of these platforms
February 20, 2025 at 2:40 AM
Yeah that’s fair! Discover is an essential part of social media. I have yet to hear an effective solution for disinformation that preserves the value of these platforms
An algorithm is an editorial choice imo, if an opaque one
February 20, 2025 at 2:35 AM
An algorithm is an editorial choice imo, if an opaque one
Facebook, Twitter,and TikTok at least, right? There’s a lot of content in my feeds on those platforms that does not come from accounts or people that I explicitly followed
February 20, 2025 at 2:34 AM
Facebook, Twitter,and TikTok at least, right? There’s a lot of content in my feeds on those platforms that does not come from accounts or people that I explicitly followed