Lee Flannery
@fullstackdrip.com
I help people make software. I'm also hilarious so there's that too.
leeflannery.com
leeflannery.com
If Lions/Bears were this month, the over/under would be like 80 points.
September 18, 2026 at 1:27 AM
If Lions/Bears were this month, the over/under would be like 80 points.
"Semantic search sucks" overgeneralizes it. The question is: does meaning-based retrieval find useful material other methods miss at an acceptable cost?
When it ignores the one word you absolutely need it to respect? Then semantic search is obviously ass. And more expensive too.
When it ignores the one word you absolutely need it to respect? Then semantic search is obviously ass. And more expensive too.
September 12, 2026 at 11:24 AM
"Semantic search sucks" overgeneralizes it. The question is: does meaning-based retrieval find useful material other methods miss at an acceptable cost?
When it ignores the one word you absolutely need it to respect? Then semantic search is obviously ass. And more expensive too.
When it ignores the one word you absolutely need it to respect? Then semantic search is obviously ass. And more expensive too.
Honestly, I don’t want AI training on some of y’all’s code either.
September 5, 2026 at 1:15 AM
Honestly, I don’t want AI training on some of y’all’s code either.
Developers sharing keys? Love that! Oh.
August 31, 2026 at 7:51 PM
Developers sharing keys? Love that! Oh.
"If he uses AI to apply how do I know if he can code?" you used AI to review the resumes how do I know you can read?
August 31, 2026 at 7:40 PM
"If he uses AI to apply how do I know if he can code?" you used AI to review the resumes how do I know you can read?
Debt can be forgiven easily enough, just not for the poors.
August 31, 2026 at 5:19 PM
Debt can be forgiven easily enough, just not for the poors.
If reopening a side project requires excavating a 70,000-token conversation from six months ago, you have a problem.
Stale sessions are psychotic. Your durable stuff is code, commits, agents dot md, maybe memory and a handoff.
Stale sessions are psychotic. Your durable stuff is code, commits, agents dot md, maybe memory and a handoff.
August 25, 2026 at 5:49 PM
If reopening a side project requires excavating a 70,000-token conversation from six months ago, you have a problem.
Stale sessions are psychotic. Your durable stuff is code, commits, agents dot md, maybe memory and a handoff.
Stale sessions are psychotic. Your durable stuff is code, commits, agents dot md, maybe memory and a handoff.
The Cursor/Grok Bot source map thing is a perfect example of where agent software is going.
Someone took apart a proprietary agent app and reconstructed enough of the harness to plug in entirely different models/providers.
The model is increasingly a dependency.
github.com/b-nnett/grok...
Someone took apart a proprietary agent app and reconstructed enough of the harness to plug in entirely different models/providers.
The model is increasingly a dependency.
github.com/b-nnett/grok...
GitHub - b-nnett/grok-bot-0.18-reconstructed: Unofficial source-oriented reconstruction and extension of Grok Bot 0.18.0 for macOS
Unofficial source-oriented reconstruction and extension of Grok Bot 0.18.0 for macOS - b-nnett/grok-bot-0.18-reconstructed
github.com
August 24, 2026 at 4:32 PM
The Cursor/Grok Bot source map thing is a perfect example of where agent software is going.
Someone took apart a proprietary agent app and reconstructed enough of the harness to plug in entirely different models/providers.
The model is increasingly a dependency.
github.com/b-nnett/grok...
Someone took apart a proprietary agent app and reconstructed enough of the harness to plug in entirely different models/providers.
The model is increasingly a dependency.
github.com/b-nnett/grok...
August 24, 2026 at 12:13 AM
I know who didn't make Ox Alpha. fullstackdrip.com/i-know-who-d...
I Know Who Didn’t Make Ox Alpha
Written August 23, 2026, before the reveal. No edits after. That's the whole point of a prediction post.
Free frontier AI model "Ox Alpha" dropped this week and developer internet is obsessed with on...
fullstackdrip.com
August 23, 2026 at 4:54 PM
I know who didn't make Ox Alpha. fullstackdrip.com/i-know-who-d...
Ox Alpha mystery meat LLM free and unlimited for a week. The compute capability available is insane. Context window is a million tokens but I'm not sure how efficient it is with it.
August 22, 2026 at 7:29 PM
Ox Alpha mystery meat LLM free and unlimited for a week. The compute capability available is insane. Context window is a million tokens but I'm not sure how efficient it is with it.
Models are becoming components. The engineering work is moving into harness design, context isolation, and evaluation.
August 22, 2026 at 6:09 PM
Models are becoming components. The engineering work is moving into harness design, context isolation, and evaluation.
Model wars are so Summer of 2026.
By Christmas people will care more about harnesses than models. I'm kinda cheating because it's almost true now: NVIDIA’s AVO result - state, memory, tools, feedback, supervision, recovery. Repost now so you can call it too!
developer.nvidia.com/blog/nvidia-...
By Christmas people will care more about harnesses than models. I'm kinda cheating because it's almost true now: NVIDIA’s AVO result - state, memory, tools, feedback, supervision, recovery. Repost now so you can call it too!
developer.nvidia.com/blog/nvidia-...
NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents | NVIDIA Technical Blog
A frontier language model is only one component of an AI agent. The surrounding agent system—often called a harness—determines how the model receives context, uses tools, maintains state…
developer.nvidia.com
August 22, 2026 at 1:21 PM
Model wars are so Summer of 2026.
By Christmas people will care more about harnesses than models. I'm kinda cheating because it's almost true now: NVIDIA’s AVO result - state, memory, tools, feedback, supervision, recovery. Repost now so you can call it too!
developer.nvidia.com/blog/nvidia-...
By Christmas people will care more about harnesses than models. I'm kinda cheating because it's almost true now: NVIDIA’s AVO result - state, memory, tools, feedback, supervision, recovery. Repost now so you can call it too!
developer.nvidia.com/blog/nvidia-...
LLMs get lazy about state representation. They’ll start with a loose object, then bolt on more states as the app grows.
That’s not always wrong!!!! But when state actually matters, give it a contract. Pydantic, Zod, something so the agent can’t just vibe its ass through invalid states.
That’s not always wrong!!!! But when state actually matters, give it a contract. Pydantic, Zod, something so the agent can’t just vibe its ass through invalid states.
August 20, 2026 at 12:31 AM
LLMs get lazy about state representation. They’ll start with a loose object, then bolt on more states as the app grows.
That’s not always wrong!!!! But when state actually matters, give it a contract. Pydantic, Zod, something so the agent can’t just vibe its ass through invalid states.
That’s not always wrong!!!! But when state actually matters, give it a contract. Pydantic, Zod, something so the agent can’t just vibe its ass through invalid states.
The baller IDE is a bunch of terminals.
cmux + OpenCode2, Git in one pane, bun running the app in another.
This is today's coding workspace. What does yours look like now?
cmux + OpenCode2, Git in one pane, bun running the app in another.
This is today's coding workspace. What does yours look like now?
August 19, 2026 at 6:19 PM
The baller IDE is a bunch of terminals.
cmux + OpenCode2, Git in one pane, bun running the app in another.
This is today's coding workspace. What does yours look like now?
cmux + OpenCode2, Git in one pane, bun running the app in another.
This is today's coding workspace. What does yours look like now?
OK the story of, Claude Code keeps track of obscenities? If that's still accurate the data ingest on fucks must be off the charts.
August 17, 2026 at 7:23 PM
OK the story of, Claude Code keeps track of obscenities? If that's still accurate the data ingest on fucks must be off the charts.
Calling it a "hiring spree" while your public headcount is rapidly declining is an incredible bit.
We have LinkedIn too, guys.
We have LinkedIn too, guys.
August 16, 2026 at 2:47 PM
Calling it a "hiring spree" while your public headcount is rapidly declining is an incredible bit.
We have LinkedIn too, guys.
We have LinkedIn too, guys.
Delta looks sick. Zed built a conversation and worktree replicated together, comments that don’t rot when the code changes.
Im ready for collaboration and for "chat box next to editor" to be over.
zed.dev/blog/introdu...
Im ready for collaboration and for "chat box next to editor" to be over.
zed.dev/blog/introdu...
Introducing Delta - Zed Blog
From the Zed Blog: A multiplayer environment for coding with agents, from the creators of Zed.
zed.dev
August 16, 2026 at 12:00 AM
Delta looks sick. Zed built a conversation and worktree replicated together, comments that don’t rot when the code changes.
Im ready for collaboration and for "chat box next to editor" to be over.
zed.dev/blog/introdu...
Im ready for collaboration and for "chat box next to editor" to be over.
zed.dev/blog/introdu...
If you’re sleeping on GLM, I think you should really take a look.
GLM-5.3 shipped TODAY, and Z.ai says it’s 50% better than 5.2 on their coding benchmark. 5.2 has been my daily llm this week and I hadn't even thought about changing. It's moving fast out here and it’s a damn treat.
z.ai/blog/glm-5.3
GLM-5.3 shipped TODAY, and Z.ai says it’s 50% better than 5.2 on their coding benchmark. 5.2 has been my daily llm this week and I hadn't even thought about changing. It's moving fast out here and it’s a damn treat.
z.ai/blog/glm-5.3
z.ai
August 14, 2026 at 9:01 PM
If you’re sleeping on GLM, I think you should really take a look.
GLM-5.3 shipped TODAY, and Z.ai says it’s 50% better than 5.2 on their coding benchmark. 5.2 has been my daily llm this week and I hadn't even thought about changing. It's moving fast out here and it’s a damn treat.
z.ai/blog/glm-5.3
GLM-5.3 shipped TODAY, and Z.ai says it’s 50% better than 5.2 on their coding benchmark. 5.2 has been my daily llm this week and I hadn't even thought about changing. It's moving fast out here and it’s a damn treat.
z.ai/blog/glm-5.3
I understand the desire to be "I got Kimi running in Claude Code" guy but that day in the sun is over.
August 14, 2026 at 9:15 AM
I understand the desire to be "I got Kimi running in Claude Code" guy but that day in the sun is over.
Hand-jamming arbitrary LLMs into Claude Code right now is a cute party trick, but stop it.
Look at these numbers, man: Claude Code has a cache-miss rate more than 5x that of OpenCode 2.
Look at these numbers, man: Claude Code has a cache-miss rate more than 5x that of OpenCode 2.
given the amount of OpenCode Go traffic we're seeing it would help if you switched to opencode2
it will eat your usage less than any other client (except zcode lol)
definitely stop shoving foreign models into claude code
it will eat your usage less than any other client (except zcode lol)
definitely stop shoving foreign models into claude code
August 14, 2026 at 9:14 AM
Hand-jamming arbitrary LLMs into Claude Code right now is a cute party trick, but stop it.
Look at these numbers, man: Claude Code has a cache-miss rate more than 5x that of OpenCode 2.
Look at these numbers, man: Claude Code has a cache-miss rate more than 5x that of OpenCode 2.
I don't know if y'all know but OpenCode's "OpenCode Go" subscription is $10 a month and, I'm not like selling anything but if you haven't coded with GLM-5.2 ... or ... for something cheaper and competent, MiniMax M3, I'm just saying. This is a cheap way to try it out.
August 13, 2026 at 10:48 PM
I don't know if y'all know but OpenCode's "OpenCode Go" subscription is $10 a month and, I'm not like selling anything but if you haven't coded with GLM-5.2 ... or ... for something cheaper and competent, MiniMax M3, I'm just saying. This is a cheap way to try it out.
If your "best engineers" routinely spend entire days beautifully solving the wrong problems, I have some unfortunate news about your best engineers.
(From my LinkedIn contrarian-brain archives)
(From my LinkedIn contrarian-brain archives)
August 12, 2026 at 10:31 PM
If your "best engineers" routinely spend entire days beautifully solving the wrong problems, I have some unfortunate news about your best engineers.
(From my LinkedIn contrarian-brain archives)
(From my LinkedIn contrarian-brain archives)
Every dev has a task list app. Mine is branded after the Emmy-nominated Mark Ruffalo/Tom Pelphrey tour de force TASK. Season 2 is in the works! (Of the show. Not my task manager.)
August 7, 2026 at 5:31 PM
Every dev has a task list app. Mine is branded after the Emmy-nominated Mark Ruffalo/Tom Pelphrey tour de force TASK. Season 2 is in the works! (Of the show. Not my task manager.)