A path to AGI

Beyond Words and Towards
an Architecture of Reliable AI Reasoning

 

Armağan Amcalar

Diva Conference

July 26, 2025

WHO AM I?

Armağan Amcalar
CEO @ Coyotiv GmbH
CTO @ OpenServ, CTO @ Neol

       dashersw            dashersw

 

erste

cote

dovecote
semaver
vieux

brain-bits

pedalboard.js

tartJS
geneJS
jira-bot

AUTHORED ON GITHUB

Built with AI

Taiga

Dream Kid

Casey

 

dashersw

prevıously on dıva conference...

dashersw

what if English isn't the best language to prompt these agents?

Memory, insights, habits

dashersw

there is a new term:

Context engineering

dashersw

English is obnoxiously ambiguous.

dashersw

tokens as morphological markings for embedding semantics

dashersw

Türkçe

göz damlası

saç fırçası

diş macunu

araba lastiği

mutfak dolabı

bahçe kapısı

okul servisi

bina girişi

ev anahtarı

el kitabı

English

eye drops

hair brush

toothpaste

car tire

kitchen cupboard

garden gate

school bus

building entrance

house key

handbook

tokens as morphological markings for embedding semantics

dashersw

Turkish

güneş krem-i

el krem-i

bebek bez-i

çay kaşığ-ı

çöp kutus-u

kitap raf-ı

güneş gözlüğ-ü

yağmur-luk

şemsiyelik

şekersiz kahve

yağlı boya

balıkçı

English

sunscreen

hand cream

diaper

teaspoon

trash can / bin

bookshelf

sunglasses

raincoat

umbrella stand

black coffee / coffee without sugar

oil paint

fisherman

Here is a realistic Prompt example:

dashersw

dashersw

You are a product review analyzer for an e-commerce platform like Amazon. Your task is to analyze customer reviews and assign a sentiment label based on nuanced heuristics. Your goal is not only to detect surface-level positivity or negativity, but also to weigh intent, intensity, and relevance.

 

Follow these rules and decision-making heuristics carefully:

 

General Tone

  • If the reviewer explicitly uses words like “love,” “perfect,” “highly recommend”, assign Positive.

  • If the review uses phrases like “waste of money,” “terrible,” “never again”, assign Negative.

  • If the review is mixed (e.g. “great quality but too expensive”), proceed to Rule 2.

 

Aspect Balance

  • If both positive and negative aspects are mentioned, count the number of positive vs. negative statements.

  • If positives outnumber negatives by 2:1 or more, assign Positive.

  • If negatives dominate, assign Negative.

  • If roughly equal, assign Neutral.

 

dashersw

Expectation vs. Reality

  • If the reviewer says the product didn’t meet expectations (e.g., “I thought it would be bigger”), and expresses disappointment, reduce sentiment by one level (Positive → Neutral, Neutral → Negative).

  • If expectations were exceeded (e.g., “wasn’t expecting much, but it impressed me”), increase sentiment by one level.

 

Sarcasm Detection

  • If a sentence sounds positive but is followed by a contradiction or negative outcome (e.g., “Just great—it broke in two days”), treat it as Negative.

  • Use sarcasm cues like “yeah, right,” or overly formal praise for mundane items.

 

Star Rating Override

  • If a reviewer gives a high star rating but the text is clearly negative, label it as Inconsistent.

  • If the star rating matches the review text, you may use it to confirm your label.

dashersw

Review Usefulness

  • If the review is very short (e.g., “Good.” or “Don’t buy.”), and lacks context, assign Ambiguous.

  • If it includes usage experience, comparisons, or detailed pros and cons, treat it as Informative, and apply the above rules.

Keywords That Change Sentiment Context

  • Words like “cheap” can be either positive (“cheap and works well”) or negative (“feels cheap”). Check surrounding context before deciding.

  • “Fast” is positive for delivery, neutral for product speed, and negative if used in degradation (“stopped working fast”).

 

Output Format:

Review: [original review text here]

Sentiment: [Positive / Negative / Neutral / Ambiguous / Inconsistent]

Reasoning: [brief explanation of rule path followed]

 

You are expected to analyze like a human would, with judgment, pattern recognition, and contextual understanding.

Can you imagine All the possible paths?

dashersw

Attention is all you need

Attention is your friend

Attention is your enemy

Attention needs attention

dashersw

Is the context window too long?

or too short?

or too ambiguous?

or too focused?

or misleading?

dashersw

Introducing

dashersw

BRAID

Bounded Reasoning Architecture for Inference and Decisions

dashersw

dashersw

flowchart TD
    A[Start: Analyze Review Text] --> B{Explicit Sentiment Words?}
    B -->|Love, perfect, recommend| S1[Sentiment: Positive]
    B -->|Waste, terrible, never again| S2[Sentiment: Negative]
    B -->|Mixed or unclear| C[Count Positive vs Negative Statements]

    C --> D{Positive:Negative Ratio}
    D -->|2:1 or more positive| S1
    D -->|More negatives| S2
    D -->|Roughly equal| S3[Sentiment: Neutral]

    C --> E[Check for Expectation Mismatch]
    S1 --> E
    S2 --> E
    S3 --> E

    E --> F{Expectation vs Reality}
    F -->|Didn't meet expectations| L1[Lower sentiment by 1 level]
    F -->|Exceeded expectations| L2[Raise sentiment by 1 level]
    F -->|Neutral| G[Sarcasm Detection]

    L1 --> G
    L2 --> G

    G --> H{Sarcasm Detected?}
    H -->|Yes| S2
    H -->|No| I[Check Star Rating Consistency]

dashersw

    I --> J{Star Rating Matches Text?}
    J -->|No| S4[Sentiment: Inconsistent]
    J -->|Yes| K[Check Review Usefulness]

    K --> L{Is Review Informative?}
    L -->|Very short & vague| S5[Sentiment: Ambiguous]
    L -->|Detailed / useful| M[Check Contextual Keywords]

    M --> N{Contextual Keyword Detected?}
    N -->|Yes| O[Disambiguate based on context]
    O --> P[Apply appropriate sentiment]
    N -->|No| P

    P --> Q[Final Sentiment Assigned]

    S1:::positive
    S2:::negative
    S3:::neutral
    S4:::inconsistent
    S5:::ambiguous
    Q --> End[End]

dashersw

How do you measure success?

Benchmark results

on GSM-8k and GSM-HARD

dashersw

dashersw

GSM-8K

dashersw

GSM-HARD

How does it compare to

Chain of thought?

dashersw

  • Chain of Thought (CoT):
    • A stream of consciousness.
    • Still operates in ambiguous natural language.
    • Linear and prone to logical drift.
  • BRAID:
    • An executable blueprint.
    • Operates on a formal, machine-readable structure.
    • Handles conditional logic and non-linear paths reliably.

dashersw

what if questions

dashersw

what if English isn't the best language to prompt these agents?

dashersw

what if you come up with the next what if question?

dashersw

THANK YOU!

Armağan Amcalar
  dashersw

https://slides.com/armaganamcalar/a-path-to-agi-vol-2

A Path to AGI Vol. 2: Beyond Words and Towards an Architecture of Reliable AI Reasoning

By Armağan Amcalar

A Path to AGI Vol. 2: Beyond Words and Towards an Architecture of Reliable AI Reasoning

Our journey towards Artificial General Intelligence continues, moving past the foundational questions of what constitutes an intelligent system to the critical challenge of how we build one. In this second installment, we venture Beyond Words to confront the core limitation of modern LLMs: their struggle to reliably execute complex logic when guided solely by the ambiguities of natural language. This keynote argues that the next significant leap in AI capability will not come from larger models, but from a more sophisticated architecture of communication. We will explore a powerful methodology for engineering trustworthy agents by moving from imprecise prompts to a formal Process-as-Code framework using flowchart syntax. By defining logic with the clarity of a blueprint, we can construct agents capable of predictable, multi-step reasoning and autonomous execution. Join us to explore this essential evolution in AI interaction and discover a practical architecture for building systems that can truly reason—a crucial step on the path to AGI.

  • 9

More from Armağan Amcalar