2025
October – December
-
Computation within a substrate of state
-
There are interesting things that LLMs can teach us about humans and our own intelligence and cognition.
-
Are constraints for a system the same as embedding the system in a substrate of continuous feedback?
-
Agentic systems are permutations of LLM call and tool call. Kinds of agentic systems start simple and progress along a continuum of agency and shifting locus of control.
-
Wondering what's happening inside an agent? Take some time this weekend to get hands-on and tinker with a really basic agent. It can be simpler than you might have imagined. You can definitely make your own with a little python experience and an api key. Start with this minimal…
-
Working with LLMs to create? Then you're essentially mining in a high dimensional semantic space. So if you want to avoid the average, your words and input need to anchor you in an unusual space to begin the mining. Anchor words? Centroids of creativity?
-
AI augmented development could lead to a resurgence of monolith architectures. It's easier if you have all the context in one place, and there's no need for micro-services if you don't have dev team proliferation.
-
There must be at least one dev team somewhere inside a large bureaucratic organization, which quietly started using AI and now have months of work done and queued just waiting for internal approvals and processing. I want to hear stories like that.
-
Near future class: CS2047 Lexicons and Patterns: Learning the language of software for effective communication with AI
-
Guiding AI to generate operational utility programs is a really nice thing to have at hand.
-
Why did we start with AI generating code, instead of having it groom the backlog?
-
Swarms of encapsulated execution https://x.com/cmcollier/status/1996990320442040821
-
Everything is an experiment. Always be experimenting.
-
Be wary if you find yourself in a space which is high complexity but low impact.
-
composable units of thinking with intermediary scratchpad https://x.com/cmcollier/status/1996957414520672397
-
Encapsulated complexity with an interface Composable units of execution
-
This is on replay throughout my day this week. Dipping in and out. The lateral thinking, patterns and abstractions as first class citizens, and anchoring in empiricism and real world application is quite beautiful. Inspiring work from @drmichaellevin https://x.com/lexfridman/sta…
-
Bring back nuance https://x.com/vboykis/status/1993645066393661784
-
Generating code is like pouring cement. The outcome is only as good as the work you put into the forms.
-
I wonder what the token costs are for this. And not because there should be shade, just so we could really understand the potential of this business model. Certainly there's a real win expressed here for small teams moving with agility and willingness to discard old code (per… h…
-
Same, the bottleneck is still me. https://x.com/yacineMTB/status/1992257725385060795
-
Very interesting points here. If I can use natural language and I can understand the lego components (aka functions) without needing to understand all internals then... https://t.co/jcOBtTlU5G
-
Discover through doing
-
Controlled Shell, Generated Core Sometime around 2012, Gary Bernhardt started talking about the program design "functional core, imperative shell". It had a pretty big impact on my own programming style. It's a style which is well suited to python, especially when you learn that…
-
And here's a version which uses the openai agents library instead, making things much easier. https://github.com/codycollier/miscellanea/blob/main/agent-colorbot/colorbot.py https://x.com/cmcollier/status/1989560283992314146
-
The inner core of an agent program is relatively simple. I made a small agent which illustrates function calling and uses the OpenAI Responses API. https://github.com/codycollier/minimal-agent-core
-
When I aggressively refactor Agent written code, no-one gets their feelings hurt. At least I don't think...
-
Tickets are just the original prompts -- @mdp314
-
Today feels like it should be a Footage and Blue Ant kind of day
-
The bottleneck is still me AI augmented development helps me move more quickly and manage more things within a given week. That, in turn, further highlights the various frictions I have that are more human. My resistance to doing certain kinds of work, fear of the unknown, etc.
-
It's like leet code, but you're presented with a vibe coded program to refactor.
-
Cursor is the new C compiler? I never really experienced Assembly programming. Sure I took a class and learned a little. But even very young it was already Apple Basic. Later Python and friends. Some C. Now I feel like I might understand what the experienced Assembly programmer…
-
I feel like I might finally understand what an engineer felt like during the era of the C compiler. The guy that wrote in Assembly language for quite a while, and then lived through the wave of C programming. Doing AI augmented programming, using natural language to drive code g…
-
The data you used to fine-tune your question answering embeddings... Does it contain fact oriented questions, opinion discussion, both?
-
It's weird how quickly the industry jumped to chunking documents for generating embeddings. It destroys meaningful information. I rarely see people discussing this. It's especially weird that it diminishes context, yet it's a leading technique in context engineering.
-
Show is better than tell. Show.
-
In search, why do we spend so much up front compute to ingest, process, and index all documents, only to generate an inverted index or maybe some embeddings? There have been plenty of advancements in nlp and semantic representation. But not so much innovation in novel indexes or…
-
Why isn't there more focus and discussion around query enrichment? It seems like there are plenty of situations where it could be a simpler and more cost effective approach to get the same benefits provided by embedding retrieval and semantic search.
-
Where's the chapter in Getting Things Done about putting my todo list in the agent prompt?
-
What if embedding retrieval and semantic search is only worth it for code? Code is a nice domain for semantic search because there's less 1:1 mapping between words and semantics. Code is of course more structured, but there's a subtle layer of indirection between the string of t…
-
One of the more interesting things about this cursor post is how they are running their own indexing and embedding. Then they presumably still send their data to turbopuffer for the search engine side. I've wondered about the choice turbopuffer has made, to not provide… https://…
-
The migration is shaping up. https://orangewords.com/?ref=x-2025-11-04-b
-
I'm dusting off my hacker news search engine built on Vespa. Having fun migrating to a new and better FastAPI backend. Also migrated to a Vue frontend. I really enjoyed using htmx before. Now I want a clearer separation of concerns between frontend and backend without the html w…
-
Working with LLMs to create? Then you're essentially mining in a high dimensional semantic space. So if you want to avoid the average, your words and input need to anchor you in an unusual space to begin the mining. Anchor words? Centroids of creativity?
-
It's not the model's fault, it's mine. https://x.com/cmcollier/status/1980681583985111264
-
A West Texas Win. Love to see it. "The GPUs are expected to be deployed in phases through 2026 at IREN’s 750MW Childress, Texas campus, in conjunction with the delivery of new liquid-cooled data centers that will collectively support 200MW of critical IT load (Horizon 1–4)." htt…
-
Programming feels like sculpting or carving with lexical tools in a conceptual space.
-
It's just me and cursor against the world
-
Still wondering... https://x.com/cmcollier/status/1983207033366106339
-
When using AI augmented development, refactoring is just as valuable as when you're not. It's also just as satisfying.
-
I'm finding that my sense of organization for python modules and packages is just as useful if not more when I'm working with AI to develop a system. Having good encapsulation helps with ensuring effective context additions in Cursor. The things which make for better context are…
-
The part of search which doesn't get enough attention is winnowing and exclusion. The smaller the candidate document set, the easier it is to accomplish the recall and precision you might need for a task. Smaller sets are also more amenable to a more diverse set of techniques in…
-
MCP and the narrow waist lesson Is MCP making a mistake? Why are they trying to expand beyond the simplicity of tool call and response? I don't debate the benefit of creating a bridge between the unstructured language world of LLMs and the structured world of programming and dat…
-
Is it possible that generating semantic embeddings is one of the biggest wastes of compute and storage we've ever executed in the search industry? Or maybe I've just had too much coffee.
-
Sometimes, with vibe coding, you're just taking a different path to the same place.
-
What is the Zipfian distribution telling us about search ingestion? I think it was my friend Jeremy that first pointed out (to me) the Zipfian distribution and what it says about the distribution of words in a corpus. It's a neat insight into linguistic patterns, and a gateway t…
-
Always Be Experimenting You never know when something will work out and teach you a lesson.
-
Remember when people used to say python is good for fast prototyping before rebuilding in a more robust language? I think about that in this new world of vibe coding prototypes.
-
Where is the innovation in search indexes? There's been a lot of optimization variants, but overall it's mostly inverted indexes and HNSW indexes.
-
No one really talks about the big problem in large scale search. It's calcification. The document count is so high I can't afford the time and resources to reindex and try changes. The cluster is so large and tweaked we have significant cost and fear to try any new search system…
-
Create the new code. Observe the chaos emerge. Bring order to bear.
-
Some things required of me to develop good software systems are very similar regardless of whether I'm writing the code, helping a team write the code, or using AI augmented development with Cursor. Clear requirements. Clear constraints. An ability to communicate these accuratel…
-
The grok-code-fast-1 model in cursor is a pretty good model for refactoring single python modules.
-
I need a simple morning news brief which keeps me informed about model pricing and details for cursor. Or maybe for AI augmented programming in general. Examples: - I missed the chance to lock in free Auto by switching to an annual subscription by Sep 15 - grok-code-fast-1 is fr…
-
Goodbye to UI as we know it "...capture the jobs-to-be-done, terminology, and data sources users rely on today." Nail the lexicon, data sources, and alignment between user intent and tool API. This is what OpenAI is telling you if you want to build ChatGPT apps. The new world of…
-
When looking at the AI boom and deep learning specifically, what is the relationship between explore-exploit cycles and create-optimize cycles?
-
Using AI to program seems to be teaching us something about software architecture philosophy. About which architectures are more suitable for optimizing the blend of risk mitigation and development speed. I don't have a thesis yet but I can see signs of it. Risk mitigation drive…
-
It's been interesting watching the deep learning community mature and grow into something more deep and complex. And to now watch the still early growth of the hardware, robotics, and reinforcement learning communities.
-
The "For You" timeline is indeed different now. It used to feel like a bad rabbit hole. Now it appears to be a recommendation list with prioritization of my favorite follows, my network, and recency. Much better! Nice work @nikitabier and the grinding timeline engineers. https:/…
-
Encapsulate complexity
-
Recommend actions not information There's a lot of talk about the new Pulse from openai. Custom reports and news and what not, each day. Information just for you. At the same time there's a growing undercurrent of negative feelings about recommendation feeds in social media. The…
-
Scarcity of tokens and pay to play The cursor pricing has been a bit confusing over time. Recently I circled back to my account usage and their documentation. It turns out I have free usage of "Auto" until the end of my current billing cycle. They were giving that benefit in ret…