1 of 15

Co-designing with LLM

a personalized cognitive engine architecture basis

Cover Image by Nikita

(n3nikita) on Unsplash

Check project repo on my GitHub ^

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

2 of 15

Intro

& disclaimer

私

Story

& context

道

Definitions

& examples

Cognitive Basis

Core part

Key Takeaways

Discussion time

References

& recommendations

Slides design note:

Illustration captions, comments or links to relevant information are located here.

Most of the photos are from my personal archive © Abramenkov Sergei

For the exceptions - licences are permissive and sources are attributed here.

Top-right: photo of the author (human) in the most comfortable professional environment (Gorely volcano, Kamchatka)

Bottom-left: author’s work desk during the PhD project first year (office №318 at Institut de Physique du Globe de Paris)

Bottom-right: recent workspace iteration

Acquaintance with the speaker
Few reasons why the talk is in English
Fresh view on LLMs - keep an open mind!

Disclaimer and credibility:

not an expert in AI / ML / LLMs
PhD degree in (geo)-physics (2021)

Who am I then?

A “full-stack” seismologist.

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

3 of 15

Story

& context

道

Definitions

& examples

Cognitive Basis

Core part

Key Takeaways

Discussion time

References

& recommendations

The earliest personal applications of raw concepts that influenced this project

Ever felt like too many browser tabs opened?

Prepare for more cognitive overload now.

Struggle to formulate your thoughts?

LLMs are here to help you with that.

(but not substitute - it’s a tool)

What if I want to tackle a philosophical problem?

Extremely poorly defined
Likely to completely change

My journey to build a cognitive engine architecture started with a basis development.

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

4 of 15

Definitions

& examples

Cognitive Basis

Core part

Key Takeaways

Discussion time

References

& recommendations

7-axis basis examples co-designed with

DeepSeek V-3 as the project byproduct

from cognitive.basis import Primitive, Links, Symbolism, Verbs

Co-designing with LLM workflow:

0 - set global guidelines (ex. limit symbolism, anchor physics metaphor)

1 - present original vision summary (Markdown file) as a draft

2 - ask LLM to check for its logical soundness and inconsistencies

(encourage harsh critique)

3 - ask LLM to identify weakest definitions and suggest alternatives

4 - iteratively review, adjust and check how well it syncs with the vision

‘Cognitive basis - a set of orthogonal, interpretable dimensions that structurally decompose cognitive processes’

Each axis/dimension:

Encodes a primitive cognitive function
Limits symbolism to metaphoric anchors (no pseudo-science)
Defensible: no unnecessary overreach into neuroscience
Maps into implementable operations (ideally) and examples

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

5 of 15

Sun

Exploration

Map: © OpenStreetMap contributors (openstreetmap.org)

Exploration = Primitive(‘traversing uncertainty to map possibilities’)

The proactive engagement with the unknown to gather raw data or stimuli

Operational Verbs = [Observe, Gather, Survey]
Physics Anchor: Entropy (E) → Exploration as ‘necessary entropy increase’
🗺️ Examples, metaphors, links:

Real-world example: cartography, data scraping,
Scientific research:

Broad literature surveying before hypothesis formation ('pioneering' the unknown)
Data acquisition (ex. in expedition, experiment)

AI/ML: algorithms that ‘maximize information gain’ from uncertain environments

Thompson Sampling
Bayesian Optimization

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

6 of 15

Mon

Determination

Image © Logan Voss

Source: Unsplash

(https://unsplash.com/photos/a-blurry-image-of-sound-waves-in-purple-and-green-9gf9jSsXP2M)

Definition = Primitive(“imposing rhythmic structures”)

Spectral decomposition of tasks into resonant intervals

Operational Verbs = [Structure, Propose, Establish]
Links:

Physics: 1/Δt → Determination as limiting “bandwidth” of input signal/noise
AI/ML: Learning rates as ‘frequency tuners’ for convergence

Gradient Descent

Real-world example: Clock Synchronization

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

7 of 15

Tue

Negotiation

Image © Lisa

Source: Unsplash

(https://unsplash.com/photos/a-chess-board-with-pieces-of-chess-on-it-Sykhmys_L-A)

Negotiation = Primitive(“balance competing inputs”)

Boltzmann distribution of competing ideas (Temperature as a physics anchor)

Operational verbs = [Argue, Balance, Influence]
Links:

Physics: kᴮT → Negotiation as lowering the probability (“cooling”) of accepting suboptimal solutions
AI/ML: Softmax Policy (τ parameter)

Temperature scaling for exploration/exploitation tradeoffs

Real-world example: Thermal Stress Testing

Danger Zone:

Overheating: Argumentation disintegration (no consensus)
Freezing: Herd thinking (local optimum traps)

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

8 of 15

Wed

Specification

Someone was using Arch BTW

Specification = Primitive(“compress ambiguity”)

Phase-space constraints (Volume as a physics anchor)

Operational Verbs = [Refine, Filter, Categorize]
Links:

Physics: ∫d³x → Specification as imposing limits to focus on high-probability regions
AI/ML: Variational Autoencoders (VAEs)

Encoder squeezes input → compact latent ‘volume’; decoder reconstructs the essence

Real-world example: Tolerance Stack-Up

Danger Zone:

Over-compression: Sterile representations (losing nuance)
Under-compression: Noise drowns signal (analysis paralysis)

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

9 of 15

Thu

Generation

Image © Joel Tinner

Source: Unsplash

(https://unsplash.com/photos/green-field-JFXZvgejlMI)

Links:

Physics: ∫p·dq → Generation as quantifying efforts to explore configurations
AI/ML: Markov Chain Monte Carlo (MCMC)

Generates samples via probabilistic work cycles

Real-world example: CAD Prototyping

Danger Zone:

Over-generation: Wasted resources
Under-generation: Underproduction

Generation = Primitive(“construct candidate solutions”)

Work expenditure to traverse solution space (Action as a physics anchor)

Operational Verbs = [Develop, Model, Iterate]

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

10 of 15

Fri

Revelation

Image © Jakob Cotton

Description: Photo of an abstract, orb-like light show in Camera Obscura & the World of Illusions in Edinburgh.

Source: Unsplash

(https://unsplash.com/photos/a-yellow-light-is-shining-in-the-center-of-a-circular-structure-cd6llDNgbZg)

Revelation = Primitive(“detect emergent patterns”)

Potential gradients driving attention (Charge as a physics anchor)

Operational Verbs = [Analyze, Distill, Visualize]

Links:

Physics: ∇²φ = ρ → Revelation as discharge of potential into an insight
AI/ML: Transformer Attention

Queries extract high-potential features

Real-world example: Spectrogram Peaks

Danger Zone:

Over-detection: False signals
Under-detection: Missed hits

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

11 of 15

Sat

Integration

Source: Unsplash

(https://unsplash.com/photos/a-bookshelf-filled-with-lots-of-books-next-to-a-lamp-Sj6yQnGmt_E)

Integration = Primitive(“synthesize stable outputs”)

The synthesis of knowledge into cohesive, shareable outputs through patient grounding (Impedance as a physics anchor)

Operational Verbs = [Analyze, Distill, Visualize]

Operational Verbs = [Conclude, Express, Publish]
Links:

Physics: Z = √(L/C) → Integration as minimizing losses when connecting subsystems
AI/ML: Ensemble Learning

Combines models’ outputs adaptively (random forests, stacking)

Real-world example: Impedance Matching

Danger Zone:

Over-coupling: Rigid integration (brittle to new inputs).
Under-synthesis: Disconnected insights (no emergent value)

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

12 of 15

The GRINDES Cognitive Basis Audit Table

(Columns sorted by functional priority)

1. Core Cognitive Function:

What the axis does (verb-centric)

2. Physics Anchor:

Critical_ for orthogonality checks

3. AI/ML Implementation:

Concrete algorithms (no hand waving)

4. Engineering Metaphor:

Applied (not theoretical) grounding

5. Danger Zones:

Paired extremes (over/under)

Core Cognitive Function	Physics Anchor	Implementation (AI/ML)	STEM-Metaphor	Dangers (Over / Under)
Map uncertainty bounds	Entropy (E)	Bayesian Optimization	Topographic Surveying	Indecisiveness / Local convergence
Impose rhythmic structure	Frequency (F)	Learning Rate Scheduling	Clock Synchronization	Rigidity / Chaos
Balance competing inputs	Temperature (T)	Multi-Agent Softmax	Thermal Stress Testing	Overheating / Freezing
Compress ambiguity	Volume (V)	Variational Autoencoders	Tolerance Stack-Up	Sterility / Noise retention
Construct candidate solutions	Action (A)	Monte Carlo Tree Search	CAD Prototyping	Wasted resources / Underproduction
Detect emergent patterns	Charge (C)	Transformer Attention	Spectrogram Peaks	False signals / Missed hits
Synthesize stable outputs	Impedance (I)	Ensemble Learning	Impedance Matching	Rigid integration / Disconnection

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

13 of 15

Hot Takes

Discussion time

Example of linear algebra applicability for visualization and analysis

“Potential gradients driving attention”

Choosing the right basis is more about grand strategy rather than just tactical tricks

Recognition from LLM: “Your framework’s orthogonal basis design resonates deeply with linear algebra because it’s fundamentally about clean decomposition and efficient recombination—the same principles that make LA the language of AI, physics, and cognitive modeling.”

How to deal with growing reliance on AI - embrace instruments but don’t forget to train natural intelligence

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

14 of 15

Hot Takes

Discussion time

Co-designing process example	Metaphor	AI/ML Concept	Hand-out example prompt you can try :)
The LLM acted like a web crawler—indexing gaps in my reasoning I couldn’t see.	“The Data Crawler”	Active Learning	Use LLMs like a ‘bias scanner’—ask: “What assumptions am I missing here?’"
We enforced iterative ‘sprints’—prompt, critique, refine, repeat.	"The CI/CD Pipeline"	Learning Rate Scheduling	Treat LLM convos like CI/CD: small, frequent commits > monolithic prompts.
We ‘softmaxed’ disagreements: for example ‘Argue for/against this axis, then balance’.	"The A/B Test"	Multi-Armed Bandits	Prompt with: “Give me 3 conflicting takes, then synthesize.”
The LLM distilled my 1,000-word rants into 3 bullet points - like a cognitive JPEG.	"The Lossy Compressor"	Quantization	Command: “Compress into a message, preserving decision points.”
The LLM was a cognitive ‘Co-pilot’: for example drafting 5 versions of a metaphor.	"The Code Generator"	Diffusion Models	Use LLMs for rapid prototyping: “Draft 3 API designs for this problem.”
The LLM flagged contradictions - like a sudden spike in a log file.	"The Anomaly Detector"	Attention Maps	To uncover hidden requirements ask: “What’s the weirdest edge case here?”
The LLM became my ‘git merge’—resolving cognitive conflicts into a coherent whole.	"The System Merge"	Model Ensembling	Finalize with: “Squash our brainstorming into the summary.md file”

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土

15 of 15

References

& recommendations

Check project repo on my GitHub ^

Seven Dimensions - a video-essay by Kieran Borovac https://www.youtube.com/watch?v=bI-FS7aZJpY

Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review. 63 (2): 81–97. DOI:10.1037/h0043158

Kotseruba, I., & Tsotsos, J. K. (2020) 40 years of cognitive architectures: core cognitive abilities and practical applications. Artificial Intelligence Review, 53 (1): 17-94. https://arxiv.org/pdf/1610.08602

Grundspenkis, J. (2007) Agent based approach for organization and personal knowledge modelling: knowledge management perspective. Journal of Intelligent Manufacturing, 18 (4): 451–457 DOI:10.1007/s10845-007-0052-6

Sat

Integration

Co-designing with LLM:

a personalized cognitive engine architecture basis

Data Fest 2025 (Spring)

Novosibirsk, Korona.tech

May 25, 12:30-13:00 (UTC+7)

Sun

Exploration

Mon

Determination

Tue

Negotiation

Wed

Specification

Thu

Generation

Fri

Revelation

Day

Primitive

字

日

月

火

水

木

金

土