2025
July – September
-
One of the useful things I've experienced with AI development, is that I can create something as a learning process and then discard the code. I might re-do it with new understanding, implement it manually, or take a different path.
-
Good advice from @arvidkahl It's starting to feel like this is changing to: The best stack is the one AI knows well and you can learn or know well enough to guide it. https://x.com/arvidkahl/status/1610464330226483202
-
When I reflect on AI augmented programming and creation, it's this minimizing of various points of friction that seems most important. https://x.com/cmcollier/status/1972693945227378989
-
Prioritize minimizing friction Everything that involves creating and building has an iterative loop. Writing code, building systems and businesses, obtaining customers, creating paintings, writing, training ML models and more. The loop of creation. In general, the more iteration…
-
Inconsistency in systems and processes leads to friction. Sometimes this is good. Sometimes this is bad.
-
I've been thinking about AI augmented ops again this morning. Previously I wasn't giving infrastructure as code enough thought. Same with docker. There's also a long continuum from me administering a Linux host by hand for fun, to the indie hacker getting something running on… h…
-
An entry in the LX Mini micro-feed
-
There's a lot of opportunity still on the table for leveraging agentic workflows to increase quality and decrease cost and effort. So many different niches and markets. It may feel like it's late to the game, but it's not at all. https://x.com/mdp314/status/1971719038209536284
-
Where is vibe ops? I've been doing a lot of AI augmented development with Cursor. Sometimes the editor. Sometimes the background agents. The last week or two, the dev work has focused on building a new FastAPI backend service and trying out a corresponding Vue frontend web app.…
-
Well what do you know... Proactive. "This also points to what I believe is the future of ChatGPT: a shift from being all reactive to being significantly proactive, and extremely personalized." https://x.com/sama/status/1971297661748953263
-
If you're building a product, even if it's technical and aimed at engineers, aim for simple interfaces. Build for operators not engineers.
-
Composable Modular Computing
-
Where is proactive AI? AI is at our beck and call. We're told ChatGPT, Claude, Gemini, and friends are ready to help with many tasks. Just ask and you will receive! Additionally the world of tool calling is putting digital world levers at the ready. AI can take action for you, s…
-
Agentic Workflows are the new process engineering? Historically, people inside businesses would wade through ambiguity and turn different back-office work into a sequence of steps. Something they could teach others. Something the business could rely on as consistent in terms of…
-
Ephemeral indexes for agents Not all LLM apps are the same. Anthropic did a good job defining agents vs agentic workflows. That is just the beginning of the rabbit hole. There are some agentic systems which are asynchronous and more like batch jobs in terms of characteristics an…
-
Nice thread about supabase edge functions. It feels like there should be something in the web managed services world which sits in between serverless and full backend app servers. The middle ground between a single edge function vs a whole fastapi application. https://x.com/rick…
-
Our world is full of attempts to direct my attention, all to achieve the goals of some mysterious other. Where's the ecosystem of attention magnets which are working for me and on my behalf?
-
Designing APIs for LLMs as a new craft? Designing a traditional API, such as for a library, rest, or rpc service, has some underlying assumptions we've all taken for granted for a long time. Maybe the one taken for granted most often... there will be some procedural code which i…
-
Asynchronous, long running agents are a distinctly new workload with very interesting opportunities.
-
In summary: Listen to the voice of the customer anywhere you can find it, ship fast, make corrections as problems arise. https://x.com/julianeagu/status/1964704824299253888
-
"The rise of agentic AI systems is, however, ushering in a mass of applications in which language models perform a small number of specialized tasks repetitively and with little variation. Here we lay out the position that small language models (SLMs) are sufficiently powerful,…
-
Basic LLM usage and dev work has garnered excitement, but it seems like a broad, shallow pool. It feels a little like those web backend jobs where all you do is integrate data from external APIs inside a request thread in a microservice.
-
Managed Services for Agents This is something @mdp314 and I have been thinking about recently.
-
With @cursorai Tab mode, you get more code out than keystrokes in. It's great for a pattern of writing a simple comment or function signature, and having the code generated.
-
The best solution to the scale-to-zero problem is to have your system rest at zero by default. This doesn't align well with traditional cluster architectures.
-
This is a great session. @turbopuffer is bringing real clarity to their case, with fundamental systems thinking and clear-eyed recognition of trade-offs. Sounds like it must be super fun to be designing and building. https://x.com/jxnlco/status/1960741368491577487
-
Agent development centers heavily around the design of the tools. At first glance, the tools are just your traditional backend api. Upon further inspection, that isn't quite right. Your tools are a conceptual api representing the app primitives from the user perspective.
-
Nice line of thinking, about actualization and healthy computing. What sarv is doing is worth following. This is how I feel about recent AI advancement. We should be bending it toward positive personal outcomes. https://x.com/SarvasvKulpati/status/1959013946679337418
-
Tinygrad is doing some really cool work https://x.com/tinygrad/status/1958761612699111698
-
Why can't I search through job postings based on the tools they use for work?
-
Per previous threads, I think I'll make an example agent centered around playing tic tac toe. It has just enough structure, state, and complexity. Designing the tool calling API is fun to think about.
-
Agents are a bridge between natural language and an API.
-
The AI assistant integration in the @supabase web interface is well done. A good example.
-
As a programmer, there are many ways to go from idea to working code. Some of those ways are more compatible with AI augmented coding. This is one reason we see so many different reactions to things like Cursor, Amp, etc.
-
Once you create clear patterns in your codebase, my experience is that @AnthropicAI claude-3.5-sonnet + @cursorai are really good at extending functionality using that pattern.
-
tired: template repo wired: template prompt https://x.com/jobergum/status/1955998610278977719
-
"Oh. You're a sonnet 4 guy?" Programmers are sometimes subtly judged by which editor they use, in different situations. Next up, it will be about which model you prefer for augmented coding.
-
It's nice that @cursorai extended the gpt-5 max time. Thanks!
-
It was fun when work also included these build outs
-
Working code wins
-
Let's see if I can do this in the next week or so. https://x.com/cmcollier/status/1953868789684047907
-
Too much opus usage earlier this month in the background agents I guess. But I think gpt-5 is free this week?
-
Interesting. Looks like @cursorai added a memory, at the project level. This is a good one too as it has been a particular thorn in the work:
-
Always Be Questing https://x.com/yacineMTB/status/1953069442629259618
-
I'm building an agent and got some function calling working today. That stuff is really cool, and I've barely scratched the surface.
-
New Background Agent...
-
Historically I'm a cli guy. None the less, I'm enjoying the web interface for firing off @cursorai agents from my phone. Wish @AmpCode had something similar.
-
Technology is becoming fun again
-
Interesting problem yesterday: It was surprisingly difficult to convince Cursor + any model to write code for the OpenAI Responses API. I suspect it's due to the semantically muddy word responses being used as a formal name. Once I quoted "Responses API" it worked better.
-
Added a new entry in my logbook. I looked into the terminology around LLM tool use and function calling. https://codycollier.com/lx/2025/terminology-llm-tool-use.html?x=2025-07-04-b
-
This is a nicely written reference, from Elastic, for embedding memory sizing. Concise and clear in the details. https://www.elastic.co/docs/deploy-manage/production-guidance/optimize-performance/approximate-knn-searchensuredatanodeshaveenoughmemory I should build a little web a…