The only AI Agent LLM in the World
AINIRO as a company might not be that large, but we've currently got the largest and smartest LLM in the world within our niche. To be correct, the Hyperlambda Generator is now probably the single largest DSL-specific LLM in existance, for a domain specific programming language, created exclusively to create tools for AI agents.
Feel free to correct me if you happen to know that I am wrong, but according to several highly credible source, our LLM based upon ~40MB of code, in ~50,000 examples, apparently is the largest LLM in existance for a DSL, outside of the common standards such as SQL, Python, etc.
According to ChatGPT the Hyperlambda files could have a strategical value of 20 million dollars. Having actually done the job, and therefore realise what a ridiculously difficult task it is to pull off, I kind of tend to agree with ChatGPT here.
First of all, such a DSL can only be built by somebody who knows the language 100% perfect - Implying for all practical concerns its primary creators or maintainers. However, I am digressing. Let's get down to business!
How it works
- You give it natural language instructions
- Our LLM generates Hyperlambda code
With the crucial point being that this Hyperlambda code can execute immediately, and/or be persisted as an "AI agent tool". Features? Everything you would expect a mature AI agent to be able to do.
- Web scraping
- Access and manipulate data in databases
- Load files, and interact with 3rd party APIs
- Send emails, publish web socket messages
- Etc, etc, etc ...
Basically everything all enterprise software developers do on a daily basis. And if you can find something we don't support, creating an extension in the thing is dead simple using C#!
Security
For SMB and enterprise clients, the absolute most important question you need to answer is how your platform is related to security. If you were to create something similar in C# or Python, then ask yourself the following question ...
How am I to execute this code without possibly compromising the entire server
The problem is that no other programming language on earth besides Hyperlambda can actually be executed securely in-process. All other languages are both too "powerful" and too time consuming to compile, interpret, or parse. In addition, they're also ridiculously complex in syntax. Since Hyperlambda is a DSL, we can restrict what it's allowed to do. For instance, it's literally impossible for Hyperlambda to install malware on your server. To put that into perspective, realise the following ...
- Hyperlambda runs on a "virtual file system" and can therefore not even in theory access anything outside of these files.
- Hence, no overwriting of binary files, executables, etc.
- Hyperlambda allows for "white listing" keywords, allowing different users to have different execution rights, based upon their trust level, secured with authentication and authorisation.
- This allows you to for instance have user A only being able to read from one specific database, while others can access more databases, or modify data in it, etc.
- You can see an example of this on our natural language API page, where we've white listed all core keywords, but only the HTTP GET verb, and no file access, database access, etc.
- Hyperlambda was specifically created to create tools for AI agents. This implies it's "a bajillion" times more secure, faster, easier to generate, and easier to understand than Python, JavaScript, C#, etc.
To understand the problem, realise the following (true) statement ...
With Hyperlambda an AI hallucination results in code that won't execute. With Python an AI hallucination results in "format C" ...
It is therefore 100% credible to claim that AI generated code created by our Hyperlambda Generator is somewhere between 1,000 and 1 million times more secure to execute than anything created by anything else, simply because of that we're the only guys with Hyperlambda.
Performance
This was actually the key reason why we created Hyperlambda. Python is simply too slow to be possible to use. To put that number into perspective, realise that Hyperlambda is almost 20 times faster than Python and 'Fast' API. In addition, Hyperlambda consumes about 5 to 10 percent of the tokens that Python would require on average.
Below is a graph illustrating performance difference from a standard web API returning records from an SQLite database, where both Hyperlambda and Python endpoints was invoked with 50 threads in a loop for 30 seconds. This implies that Hyperlambda has got roughly 20 times as much throughput as Python, and hence also scales roughly 100 times better.
Wrapping up
For a DSL-specific LLM, exclusively created to "solve the AI agent 'tool' space", an LLM with 50,000 high quality training examples is actually a really, really big deal. According to ChatGPT it might be worth as much as 20 million dollars. Its absolute minimum evaluation was $600,000. However, at the end of the day, none of that matters. What matters is that you can now deliver AI agents with any amounts of tools you wish, using nothing but natural language to generate such tools - Also "on demand" if required.
