For some purpose, the DOGE people are burrowing their way into all US Federal Systems. Their complete control over the Treasury Department is entirely insane. Unless you intend to destroy everything, making arbitrary changes to complex computer systems will result in destruction, even if that was not your intention. No

Entering 2025, I decided to spend some time exploring the topic of agents. I started reading Anthropic’s Building effective agents, followed by Chip Huyen’s AI Engineering. I kicked off a major workstream at work on using agents, and I also decided to do a personal experiment of sorts. This is a general commentary on building that project. What I wanted to build was a simple chat interface where I could write prompts, select models, and have the model use tools as appropriate. My side goal was to build this using Cursor and generally avoid writing code directly as much as possible, but I found that generally slower than writing code in emacs while relying on 4o-mini to provide working examples to pull from. Similarly, while I initially envisioned building this in fullstack TypeScript via Cursor, I ultimately bailed into a stack that I’m more comfortable, and ended up using Python3, FastAPI, PostgreSQL, and SQLAlchemy with the async psycopg3 driver. It’s been a… while… since I started a brand new Python project, and used this project as an opportunity to get comfortable with Python3’s async/await mechanisms along with Python3’s typing along with mypy. Finally, I also wanted to experiment with Tailwind, and ended up using TailwindUI’s components to build the site. The working version supports everything I wanted: creating chats with models, and allowing those models to use function calling to use tools that I provide. The models are allowed to call any number of tools in pursuit of the problem they are solving. The tool usage is the most interesting part here for sure. The simplest tool I created was a get_temperature tool that provided a fake temperature for your location. This allowed me to ask questions like “What should I wear tomorrow in San Francisco, CA?” and get a useful respond. The code to add this function to my project was pretty straightforward, just three lines of Python and 25 lines of metadata to pass to the OpenAI API. def tool_get_current_weather(location: str|None=None, format: str|None=None) -> str: "Simple proof of concept tool." temp = random.randint(40, 90) if format == 'fahrenheit' else random.randint(10, 25) return f"It's going to be {temp} degrees {format} tomorrow." FUNCTION_REGISTRY['get_current_weather'] = tool_get_current_weather TOOL_USAGE_REGISTRY['get_current_weather'] = { "type": "function", "function": { "name": "get_current_weather", "description": "Get the current weather", "parameters": { "type": "object", "properties": { "location": { "type": "string", "description": "The city and state, e.g. San Francisco, CA", }, "format": { "type": "string", "enum": ["celsius", "fahrenheit"], "description": "The temperature unit to use. Infer this from the users location.", }, }, "required": ["location", "format"], }, } } After getting this tool, the next tool I added was a simple URL retriever tool, which allowed the agent to grab a URL and use the content of that URL in its prompt. The implementation for this tool was similarly quite simple. def tool_get_url(url: str|None=None) -> str: if url is None: return '' url = str(url) response = requests.get(url) soup = BeautifulSoup(response.content, 'html.parser') content = soup.find('main') or soup.find('article') or soup.body if not content: return str(response.content) markdown = markdownify(str(content), heading_style="ATX").strip() return str(markdown) FUNCTION_REGISTRY['get_url'] = tool_get_url TOOL_USAGE_REGISTRY['get_url'] = { "type": "function", "function": { "name": "get_url", "description": "Retrieve the contents of a website via its URL.", "parameters": { "type": "object", "properties": { "url": { "type": "string", "description": "The complete URL, including protocol to retrieve. For example: \"https://lethain.com\"", } }, "required": ["url"], }, } } What’s pretty amazing is how much power you can add to your agent by adding such a trivial tool as retrieving a URL. You can similarly imagine adding tools for retrieving and commenting on Github pull requests and so, which could allow a very simple agent tool like this to become quite useful. Working on this project gave me a moderately compelling view of a near-term future where most engineers have simple application like this running that they can pipe events into from various systems (email, text, Github pull requests, calendars, etc), create triggers that map events to templates that feed into prompts, and execute those prompts with tool-aware agents. Combine that with ability for other agents to register themselves with you and expose the tools that they have access to (e.g. schedule an event with tool’s owner), and a bunch of interesting things become very accessible with a very modest amount of effort: You could schedule events between two busy people’s calendars, as if both of them had an assistant managing their calendar Reply to your own pull requests with new blog posts, providing feedback on typos and grammatical issues Crawl websites you care about and identify posts you might be interested in Ask the model to generate a system model using lethain:systems, run that model, then chart the responses Add a “planning tool” which allows the model to generate a plan to guide subsequent steps in a complex task. (e.g. getting my calendar, getting a friend’s calendar, suggesting a time we could meet) None of these are exactly lifesaving, but each is somewhat useful, and I imagine there are many more fairly obvious ideas that become easy once you have the necessary scaffolding to make this sort of thing easy. Altogether, I think that I am convinced at this points that agents, using current foundational models, are going to create a number of very interesting experiences that improve our day to day lives in small ways that are, in aggregate, pretty transformational. I’m less convinced that this is the way all software should work going forward though, but more thoughts on that over time. (A bunch of fun experiments happening at work, but early days on those.)

One of my recent home organisation projects has been sorting out my LEGO collection. I have a bunch of sets which are mixed together in one messy box, and I’m trying to separate bricks back into distinct sets. My collection is nowhere near large enough to be worth sorting by individual parts, and I hope that breaking down by set will make it all easier to manage and store. I’ve been creating spreadsheets to track the parts in each set, and count them out as I find them. I briefly hinted at this in my post about looking at images in spreadsheets, where I included a screenshot of one of my inventory spreadsheets: These spreadsheets have been invaluable – I can see exactly what pieces I need, and what pieces I’m missing. Without them, I wouldn’t even attempt this. I’m about to pause this cleanup and work on some other things, but first I wanted to write some notes on how I’m creating these spreadsheets – I’ll probably want them again in the future. Getting a list of parts in a set There are various ways to get a list of parts in a LEGO set: Newer LEGO sets include a list of parts at the back of the printed instructions You can get a list from LEGO-owned website like LEGO.com or BrickLink There are community-maintained databases on sites like Rebrickable I decided to use the community maintained lists from Rebrickable – they seem very accurate in my experience, and you can download daily snapshots of their entire catalog database. The latter is very powerful, because now I can load the database into my tools of choice, and slice and dice the data in fun and interesting ways. Downloading their entire database is less than 15MB – which is to say, two-thirds the size of just opening the LEGO.com homepage. Bargain! Putting Rebrickable data in a SQLite database My tool of choice is SQLite. I slept on this for years, but I’ve come to realise just how powerful and useful it can be. A big part of what made me realise the power of SQLite is seeing Simon Willison’s work with datasette, and some of the cool things he’s built on top of SQLite. Simon also publishes a command-line tool sqlite-utils for manipulating SQLite databases, and that’s what I’ve been using to create my spreadsheets. Here’s my process: Create a Python virtual environment, and install sqlite-utils: python3 -m venv .venv source .venv/bin/activate pip install sqlite-utils At time of writing, the latest version of sqlite-utils is 3.38. Download the Rebrickable database tables I care about, uncompress them, and load them into a SQLite database: curl -O 'https://cdn.rebrickable.com/media/downloads/colors.csv.gz' curl -O 'https://cdn.rebrickable.com/media/downloads/parts.csv.gz' curl -O 'https://cdn.rebrickable.com/media/downloads/inventories.csv.gz' curl -O 'https://cdn.rebrickable.com/media/downloads/inventory_parts.csv.gz' gunzip colors.csv.gz gunzip parts.csv.gz gunzip inventories.csv.gz gunzip inventory_parts.csv.gz sqlite-utils insert lego_parts.db colors colors.csv --csv sqlite-utils insert lego_parts.db parts parts.csv --csv sqlite-utils insert lego_parts.db inventories inventories.csv --csv sqlite-utils insert lego_parts.db inventory_parts inventory_parts.csv --csv The inventory_parts table describes how many of each part there are in a set. “Set S contains 10 of part P in colour C.” The parts and colors table contains detailed information about each part and color. The inventories table matches the official LEGO set numbers to the inventory IDs in Rebrickable’s database. “The set sold by LEGO as 6616-1 has ID 4159 in the inventory table.” Run a SQLite query that gets information from the different tables to tell me about all the parts in a particular set: SELECT ip.img_url, ip.quantity, ip.is_spare, c.name as color, p.name, ip.part_num FROM inventory_parts ip JOIN inventories i ON ip.inventory_id = i.id JOIN parts p ON ip.part_num = p.part_num JOIN colors c ON ip.color_id = c.id WHERE i.set_num = '6616-1'; Or use sqlite-utils to export the query results as a spreadsheet: sqlite-utils lego_parts.db " SELECT ip.img_url, ip.quantity, ip.is_spare, c.name as color, p.name, ip.part_num FROM inventory_parts ip JOIN inventories i ON ip.inventory_id = i.id JOIN parts p ON ip.part_num = p.part_num JOIN colors c ON ip.color_id = c.id WHERE i.set_num = '6616-1';" --csv > 6616-1.csv Here are the first few lines of that CSV: img_url,quantity,is_spare,color,name,part_num https://cdn.rebrickable.com/media/parts/photos/9999/23064-9999-e6da02af-9e23-44cd-a475-16f30db9c527.jpg,1,False,[No Color/Any Color],Sticker Sheet for Set 6616-1,23064 https://cdn.rebrickable.com/media/parts/elements/4523412.jpg,2,False,White,Flag 2 x 2 Square [Thin Clips] with Chequered Print,2335pr0019 https://cdn.rebrickable.com/media/parts/photos/15/2335px13-15-33ae3ea3-9921-45fc-b7f0-0cd40203f749.jpg,2,False,White,Flag 2 x 2 Square [Thin Clips] with Octan Logo Print,2335pr0024 https://cdn.rebrickable.com/media/parts/elements/4141999.jpg,4,False,Green,Tile Special 1 x 2 Grille with Bottom Groove,2412b https://cdn.rebrickable.com/media/parts/elements/4125254.jpg,4,False,Orange,Tile Special 1 x 2 Grille with Bottom Groove,2412b Import that spreadsheet into Google Sheets, then add a couple of columns. I add a column image where every cell has the formula =IMAGE(…) that references the image URL. This gives me an inline image, so I know what that brick looks like. I add a new column quantity I have where every cell starts at 0, which is where I’ll count bricks as I find them. I add a new column remaining to find which counts the difference between quantity and quantity I have. Then I can highlight or filter for rows where this is non-zero, so I can see the bricks I still need to find. If you’re interested, here’s an example spreadsheet that has a clean inventory. It took me a while to refine the SQL query, but now I have it, I can create a new spreadsheet in less than a minute. One of the things I’ve realised over the last year or so is how powerful “get the data into SQLite” can be – it opens the door to all sorts of interesting queries and questions, with a relatively small amount of code required. I’m sure I could write a custom script just for this task, but it wouldn’t be as concise or flexible. [If the formatting of this post looks odd in your feed reader, visit the original article]

A lieutenant colonel in the Soviet Air Defense Forces prevented the end of human civilization on September 26th, 1983. His name was Stanislav Petrov. Protocol dictated that the Soviet Union would retaliate against any nuclear strikes sent by the United States. This was a policy of mutually assured destruction, a doctrine that compels a horrifying logical conclusion. The second and third stage effects of this type of exchange would be even more catastrophic. Allies for each side would likely be pulled into the conflict. The resulting nuclear winter was projected to lead to 2 billion deaths due to starvation. This is to say nothing about those who would have been unfortunate enough to have survived. Petrov’s job was to monitor Oko, the computerized warning systems built to centralize Soviet satellite communications. Around midnight, he received a report that one of the satellites had detected the infrared signature of a single launch of a United States ICBM. While Petrov was deciding what to do about this report, the system detected four more incoming missile launches. He had minutes to make a choice about what to do. It is impossible to imagine the amount of pressure placed on him at this moment. Source: Stanislav Petrov, Soviet officer credited with averting nuclear war, dies at 77 by Schwartzreport. Petrov lived in a world of deterministic systems. The technologies that powered these warning systems have outputs that are guaranteed, provided the proper inputs are provided. However, deterministic does not mean infallible. The only reason you are alive and reading this is because Petrov understood that the systems he observed were capable of error. He was suspicious of what he was seeing reported, and chose not to escalate a retaliatory strike. There were two factors guiding his decision: A surprise attack would most likely have used hundreds of missiles, and not just five. The allegedly foolproof Oko system was new and prone to errors. An error in a deterministic system can still lead to expected outputs being generated. For the Oko system, infrared reflections of the sun shining off of the tops of clouds created a false positive that was interpreted as detection of a nuclear launch event. Source: US-K History by Kosmonavtika. The concept of erroneous truth is a deep thing to internalize, as computerized systems are presented as omniscient, indefective, and absolute. Petrov’s rewards for this action were reprimands, reassignment, and denial of promotion. This was likely for embarrassing his superiors by the politically inconvenient shedding of light on issues with the Oko system. A coerced early retirement caused a nervous breakdown, likely him having to grapple with the weight of his decision. It was only in the 1990s—after the fall of the Soviet Union—that his actions were discovered internationally and celebrated. Stanislav Petrov was given the recognition that he deserved, including being honored by the United Nations, awarded the Dresden Peace Prize, featured in a documentary, and being able to visit a Minuteman Missile silo in the United States. On January 31st, 2025, OpenAI struck a deal with the United States government to use its AI product for nuclear weapon security. It is unclear how this technology will be used, where, and to what extent. It is also unclear how OpenAI’s systems function, as they are black box technologies. What is known is that LLM-generated responses—the product OpenAI sells—are non-deterministic. Non-deterministic systems don’t have guaranteed outputs from their inputs. In addition, LLM-based technology hallucinates—it invents content with no self-knowledge that it is a falsehood. Non-deterministic systems that are computerized also have the perception as being authoritative, the same as their deterministic peers. It is not a question of how the output is generated, it is one of the output being perceived to come from a machine. These are terrifying things to know. Consider not only the systems this technology is being applied to, but also the thoughtless speed of their integration. Then consider how we’ve historically been conditioned and rewarded to interpret the output of these systems, and then how we perceive and treat skeptics. We don’t live in a purely deterministic world of technology anymore. Stanislav Petrov died on September 18th, 2017, before this change occurred. I would be incredibly curious to know his thoughts about our current reality, as well as the increasing abdication of human monitoring of automated systems in favor of notably biased, supposed “AI solutions.” In acknowledging Petrov’s skepticism in a time of mania and political instability, we acknowledge a quote from former U.S. Secretary of Defense William J. Perry’s memoir about the incident: [Oko’s false positives] illustrates the immense danger of placing our fate in the hands of automated systems that are susceptible to failure and human beings who are fallible.

In our *Ambsheets* project, we are exploring a small extension to the familiar spreadsheet: **what if a single spreadsheet cell could hold multiple values at once**?