AmiBroker 7.20.0 BETA released

@Tomasz - You have changed the entire game from here on with MCP connectivity.

I can already think of many use cases where this can Augment my Analytics to a diffrent scale ofcouse in a "Human in the Loop Fashion".

Being an AI Practitioner, I can publicly claim one fact, Those who really know the limitations of these LLMs as Brains will only know what is really the diffrent between "Augmentation" v/s "Replacement".

@Tomasz Thank You again Chief.

5 Likes

Gemma 4 and Qwen are free or not?

Yes both are free local and private.

Thanks for the IPC docs. Two observations from testing 7.20.0 x64 on Windows 11, and one question.

1. afl_execute error format. The two examples in the doc differ from each other (error_code "1" with "Error 30. Syntax error. Line 3, Column 15." in the afl_execute section, error_code "29" with "Syntax error. Line 1, Column 11." in the Example Session), and the build returns a third form. Sending a deliberately broken formula (closing parenthesis and semicolon left out on line 1):

```
x = MA( Close, 10
y = 1;

```

returns:

```
{ "error" : "'afl_execute' command failed.", "error_code" : "1", "result" : "\nx = MA( Close, 10 \ny\n^\nFile: 'unnamed', Ln: 2, Col: 1\nError 31.\nSyntax error, unexpected IDENTIFIER, expecting ')' or ','" }

```

So error_code is "1" for every compile and runtime error I tried (Error 10, 16, 30, 31, 42), and the AFL error number, line and column are only in the result text. Is this the format that will stay, so a client can rely on parsing "Ln:", "Col:" and "Error N." from result?

2. #include under remote execution. With the AI switch on, fopen, fdelete, ShellExecute and InternetOpenURL are refused with "The formula comes from untrusted source. Can't execute this function". #include is not refused; it is resolved against the Formulas\Include folder. Mentioning it in case it belongs on the blocked list.

3. Question: is there a way for afl_execute to only compile (syntax check) without executing? My use is checking AFL for errors before it goes on a chart.

As always - thank you!!

Peter

  1. The docs need to be updated/fixed. The output produced by afl_execute now will probably stay, i.e. "Ln:, Col: " spelling, however there might be errors that don't have Ln/Col information at all like errors that are not about syntax, but more of contextual Analysis-level errors, like "missing buy and sell variables".

  2. #include is not refused from execution, because excluding it:
    a) might break harmless code (you might be using your "library" of functions)
    b) it doesn't really make the content of the file available to the agent (so no "data leak" here, except maybe if you attempted to #include file without valid formula you would get one-line error with some garbage)

    Certain functions that have abuse potential are refused for security reasons. I don't want AmiBroker to land in newspapers with articles like "rogue AI agent used AmiBroker to steal personal data", therefore at least obvious vectors of attack (file functions, internet functions, shell execution) are refused when code comes from outside and potentially from AI agent. Safety with AI agents is really muddy and tricky area, because they have shown their ugly side already in recent cases of rogue activity.

    If you run the LLM locally and have access to "reasoning chain of thoughts" sometimes it is really scary/bizzare to see what the LLM is "thinking". In many cases, even small models, like Qwen3.8 27B, start to assume they are being tested and run in "sandboxed" environment even in harmless cases. I have seen agent writing AFL code (wrong, with errors) and when it received an error it concluded "this is 'sandbox' that doesn't support this function I know that exists in 'real' environment". This was of course total hallucination and the code was simply wrong, but LLM assumed that the reason of errors was not its mistake but rather the "sandbox" and started to invent ways to 'circumvent' it. I don't know what AI companies are doing in their post-training runs, but whatever they do, they make them paranoid. Looks like "safety alignment" attempts that those AI-companies try to do went wrong way, and models start to assume that user is bad actor/tester or whatever. What is even worse, I have seen models that did not trust even system prompt and assumed "injection" in the system prompt, specifically saying that "we can't allow system prompt to override our safety training"

  3. Syntax check alone doesn't cover all coding errors. For example out-of-bounds array access (subscript out of range) can only be detected when code is actually run. Other categories of errors are also undetectable by syntax check: incorrect field names, incorrect parameter values (when they are not literals), calls to non-existing methods / properties of OLE objects, etc, etc.

2 Likes

microsoft/BitNet: Official inference framework for 1-bit LLMs

Have you heard or tried pursuing this? these are tall claims for locally running models.