ByteLite LLC

Approaches to AI

A different bet from mainstream AI.

Where this research differs from generative AI, and where it does not, with published sources for every comparison.

In short

Breadth today, versus proof by design.

Mainstream generative AI is broad, capable and widely used. This research gives up that breadth for now to make every answer traceable, every gap visible and every behaviour governed.

  • Knowledge

    Mainstream: patterns learned from very large text collections.

    This research: explicit records, each with its source.

  • Unknowns

    Mainstream: can produce a fluent answer either way.

    This research: says it does not know.

  • Learning

    Mainstream: training and fine-tuning on data.

    This research: explicit operator teaching only.

  • Coverage

    Mainstream: very broad.

    This research: narrow, and not yet public.

Not claimed on this page

  • That the ByteLite LLC approach is better. There is no comparable evidence.
  • Any benchmark score, ranking or performance figure.
  • That mainstream systems lack reasoning, memory, structured knowledge or safety controls. Many combine them.

Comparison

Topic by topic.

Mainstream approaches compared with the ByteLite LLC cognition research. Numbers in brackets refer to the sources below.
Topic Typical mainstream systems Also found in mainstream systems ByteLite LLC cognition research today Intended direction
Core method [1][2] Large neural networks trained on very large text collections generate a response one token at a time. Many products combine the model with search, databases, code execution and other tools. No generative language model is used. Replies come from governed knowledge records, from structured information supplied with the request, or from fixed conversation templates, all through one governed processing path. Responses built from explicitly represented concepts and their relationships.
Breadth of capability [2] Broad, general-purpose language ability across writing, translation, coding, summarisation and question answering, used by very large numbers of people. Image, audio and video input and output in many current systems. Narrow. A research system with limited knowledge coverage that is not available to the public. Broader coverage, added only as fast as it can be governed and verified.
Knowledge representation [8][9] Most knowledge is held implicitly in the model's learned parameters. Knowledge graphs, databases and retrieval systems supply explicit knowledge alongside the model. Knowledge is held as explicit records that carry where they came from. Connected concepts and relationships as the primary foundation of the system, not an add-on.
Reasoning [6][5][7] Multi-step reasoning emerges from the model and is strengthened with prompting and with training methods such as reinforcement learning. Reasoning interleaved with tool calls, external checkers and symbolic components. Limited. The parts that would let the system reason over a request's structure are early, not verified and not connected to answers. Reasoning over explicit relationships, with each step traceable to established knowledge.
Retrieval and tool use [3][4] Retrieval-augmented generation and tool calling are widely used to add current or specialised information. Models that learn when and how to call external tools. The system uses its own governed knowledge records and information supplied with the request. It does not search the web or call outside tools. Requests for information from other specialised systems, such as visual or audio systems, through defined interfaces.
Provenance [3] A model on its own generally cannot identify which training data produced a given statement. Retrieval systems can cite the documents they retrieved. Every knowledge record carries a record of its stated source. How sources are trusted is still being decided. Every answer traceable to the exact records and knowledge generation it used.
Uncertainty [10][11] Models can produce fluent statements that are wrong, an effect widely studied as hallucination. Research shows models can often estimate their own confidence, and products add citation and verification steps. In tested cases, the system says plainly when it does not know about a concept. A precise account of which knowledge a task requires and which of it is missing.
Determinism Output is commonly sampled, so the same prompt can produce different responses. Lower-randomness settings, and deterministic tools called by the model, are widely available. The same request against the same knowledge produces the same response in internal tests on one machine. Determinism as a property that outside reviewers can confirm.
Learning over time [12] Learned parameters are fixed after training. Updating them requires further training, and continual learning is a known open problem. Memory features that store information outside the model. Knowledge changes only through explicit operator steps. The system does not learn from conversations. Continuing knowledge acquisition under governance, with provenance kept for every addition.
Governance [13][14][16] Behaviour is shaped by training with human or AI feedback, plus policy filters around the model. Organisational risk-management frameworks applied to the whole system. Parts of a governed processing path exist and pass internal tests. Complete governance has not been demonstrated, so public use stays disabled. Governance that is explicit, inspectable and independently verified before any public use.
Verification [15] Systems are evaluated on broad public benchmarks and in third-party evaluations. Holistic evaluation suites that measure many properties at once. Internal automated tests only. No independent evaluation has taken place. Independent evaluation against published criteria before any capability claim.

Sources

References.

  1. Vaswani et al., "Attention Is All You Need", NeurIPS 2017. arxiv.org/abs/1706.03762
  2. Brown et al., "Language Models are Few-Shot Learners", NeurIPS 2020. arxiv.org/abs/2005.14165
  3. Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", NeurIPS 2020. arxiv.org/abs/2005.11401
  4. Schick et al., "Toolformer: Language Models Can Teach Themselves to Use Tools", NeurIPS 2023. arxiv.org/abs/2302.04761
  5. Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models", ICLR 2023. arxiv.org/abs/2210.03629
  6. Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models", NeurIPS 2022. arxiv.org/abs/2201.11903
  7. DeepSeek-AI, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning", Nature, 2025. arxiv.org/abs/2501.12948
  8. Hogan et al., "Knowledge Graphs", ACM Computing Surveys, 2021. arxiv.org/abs/2003.02320
  9. d'Avila Garcez and Lamb, "Neurosymbolic AI: The 3rd Wave", Artificial Intelligence Review, 2023. arxiv.org/abs/2012.05876
  10. Ji et al., "Survey of Hallucination in Natural Language Generation", ACM Computing Surveys, 2023. arxiv.org/abs/2202.03629
  11. Kadavath et al., "Language Models (Mostly) Know What They Know", arXiv preprint, 2022. arxiv.org/abs/2207.05221
  12. Kirkpatrick et al., "Overcoming catastrophic forgetting in neural networks", PNAS, 2017. arxiv.org/abs/1612.00796
  13. Ouyang et al., "Training language models to follow instructions with human feedback", NeurIPS 2022. arxiv.org/abs/2203.02155
  14. Bai et al., "Constitutional AI: Harmlessness from AI Feedback", arXiv preprint, 2022. arxiv.org/abs/2212.08073
  15. Liang et al., "Holistic Evaluation of Language Models", Transactions on Machine Learning Research, 2023. arxiv.org/abs/2211.09110
  16. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", 2023. doi.org/10.6028/NIST.AI.100-1