New chatbot

Lately I've been playing with writing a chatbot library in Common Lisp.

My previous gemini bindings were getting unweildy. I wanted to add the ability to run LLMs on my local machine but it turned out to be really kind of kludgy, so I decided to start from scratch with multiple back ends in mind.

I've got it to the point where in supports multiple back ends, so now I can prompt local LLMs from Lisp.

Recently I added the ability to recursively launch chatbots that can call each other. Since the chatbots do not share their contexts, this greatly reduces the context bloat of thet main chat because it can spawn off subtasks to a minion and not pollute the main context. This also allows you to create a federation of chatbots, each of which specializes in some topic and is overseen by a controlling chatbot that talks to the user.

Chatbots can be serialized and checkpointed, so if one is carrying out an agentic task and Lisp crashes, when we restart the agentic tasks are restarted as well and pick up where they left off.

IT turns out that recursive chats are a useful abstraction once you figure out how to use them. Basically any prompt you may issue may also want to be issued by an llm and this enables that to happen. It allows you to run subprocesses that would otherwise put junk in your context, for example reading the contents of a lange number of files. If you put that into a rocursive chatbot, it could slurp up the files into its context without adding tokens to the parent chat.

You can use a recursive chat as a `smart component'. The recursive chat can have a specialized system instruction and can preload its context with relevant information specific to it. It's context doesn't get diluted by the caller's context

Anecdote or data point

I saw that there was some argument over how much slower slot access is than struct access, so I just decided to measure it naively. I made a two slot sruct and a CLOS version of a CONS cell with car and cdr slots and I ran LTAK using regular lists, `lists' made from CLOS conses, and `lists' made from structs. Here are the results:

D:\repositories\clos-benchmark>sbcl --script run-benchmarks.lisp
Benchmark: ltak over native cons cells, CLOS my-cons nodes, and my-cons-struct nodes
Inputs: x=15 y=9 z=4 repeats=35

Scenario                   min-ms     mean-ms      max-ms      ratio
--------------------------------------------------------------------
native standard               0.129      0.146      0.186
clos standard                 1.346      1.365      1.475       9.37x
struct standard               0.172      0.175      0.179       1.20x
native optimized              0.068      0.069      0.073
clos optimized                0.411      0.414      0.419       6.04x
struct optimized              0.068      0.069      0.073       1.01x

In this naive use case, structs are same as native cons cells, but CLOS objects are one ninth the speed of a struct or cons cell if you just use it unoptimized, and one sixth the speed if optimizations are turned on.

But the CLOS instance is more functional than the cons cell in mimics. For instance, I could add a slot to the class and all the instances would be lazily updated with the new slot. I can also subclass the CLOS class and the selector functions will continue to work. Finally, I can redefine the CLOS closs while I'm developing it and all the instances will be uppdated. THe machinery to keep all this running is costing us our factor of 9.

But this might be worth the cost if we are running on a network where the bulk of the time will be transmitting the answer down the pipe once it is computed. Taking a few extra milliseconds to compute the answer might be worth the convenience features of CLOS.

Controlled Unclassified Information

Back in the day, the US government had a program called SBIR (Small Business Innovation Research) that funded small businesses to do research and development. I recall sitting in our dorm in college, reading through a giant printed catalog of SBIR grants just to amuse ourselves by brainstorming solutions over bad pizza.

.

So, I got curious the other day: what does the SBIR landscape look like now?

I can tell you right now: do not even try to read an SBIR solicitation on your local machine. You are opening yourself up to a world of absolute, unmitigated pain.

You might think, what harm could there be in simply opening a file?

Well, in the modern compliance panopticon, any manipulation of digital information that comes from the govenment has the potential to spawn CUI (Controlled Unclassified Information). CUI is basically a digital pathogen; once you download that file, *anything whatsover* derived from it, including notes and metadata, instantly becomes CUI by association. The moment you read an SBIR on your computer, you’ve infected your system, rendering you subject to a nightmare of Byzantine federal regulations.

These days, the amount of beurocratic red tape surrounding CUI is insane. To even look at the file legally, you need a dedicated, air-gapped machine completely disconnected from the internet, conforming to a massive, expensive slew of NIST standards covering everything from hardware-level encryption to strict access controls. Alternatively you could contract with a cloud company that offers a pre-certified "CUI-compliant" environment.

And assuming you actually shell out the cash and jump through the hoops to set up this digital containment zone just to read a PDF, you must meticulously audit and account for every single action you take in its presence. Under current federal auditing logic, you are explicitly assumed to be attempting to defraud the government unless you can produce a mountain of paper proving otherwise. Want to bring in a partner to bounce ideas around? You can’t just "know a guy." You have to navigate a labyrinth of federal subcontracting regulations.

I had intended on amusing myself by reading some SBIRs and daydreaming about solutions that might involve Lisp (an impossibility in the modern enterprise stack for entirely separate, depressing reasons). Instead, I quickly discovered I did not even own the physical hardware required to even read an SBIR without running afoul of federal regulations.

I wanted to read some clever and inspiring engineering proposals. I ended up reading a lot of very dry and boring compliance regulations.

Regression

Last year I wrote some Lisp related AI apps. There was a syntax highlighter that used the LLM to determine how to colorize and highlight syntax, and a prompt refiner that takes a wimpy LLM prompt and creates more elaborate prompt from them.

I took the apps down last week. They were `vibe coded' and therefore approximate and had bugs (but that's to be expected), but they had a security hole where you could hijack the LLM processing with your own prompt turning my app into an open relay using my API key. Last week I discovered that my AI spend on video creation was becoming serious. This is odd because I never create AI video. It turned out that my app was being hijacked by a proxy in Luxembourg and was generating videos on my dime.

So I shut down the apps. I knew they had the potential of being abused, and I was willing to tolerate a small amount of abuse, but it didn't occur to me that syntax highlighter could be hijacked to generate gigabytes of video at my expense. Future applications will be careful to obtain the API key from the user.