Data · Evaluation · Retrieval · Infrastructure

Research for the next horizon.

A frontier-AI research lab building knowledge and infrastructure for intelligence itself.

Intelligence is not accidental. It is engineered.

The substrate.

Data

The pipes underneath. How structured data moves into and out of language models, where the abstractions leak, and what 'good data' means in a frontier-model world.

Evaluation

What it means to know a model is actually good at something. Methods for tasks that resist clean ground truth, beyond the benchmarks that are already saturating.

Retrieval

How models find what they need. Indexing, ranking, and the open questions in retrieval-augmented systems as context windows keep growing.

Infrastructure

The compute and serving layers that make research-grade work into something usable. Where ergonomic tooling is missing. What the substrate looks like a year from now.

We work at the foundational layer of modern AI. The focus: how frontier models are built, the data and infrastructure beneath them, and frameworks for understanding where intelligence is going.

Three surfaces.

Labs

Prototypes, tools, and frameworks we build.

Papers

Formal write-ups and longer briefs.

From the lab

Notes and papers published from the substrate work.

Research notes

All notes

May 22, 2026Telasian Labs

The closing window for public benchmarks

Twelve months out, the existing public benchmark suite will be fully saturated by every frontier model. The next generation of evaluations is going to look completely different. It is not going to be cheap.

March 14, 2026Telasian Labs

The data substrate is where alignment actually happens

The alignment conversation centers on post-training: RLHF, constitutional methods, refusal training. The base model already knows what it knows. Alignment is mostly choosing which of those things to emphasize.

January 23, 2026Telasian Labs

What a benchmark cannot tell you

Saturation on a benchmark does not mean the capability surface is mapped. The gap between the published evaluations and the real cognitive frontier is widening, and the field is optimizing for the proxy.

Papers

All papers

May 27, 2026Telasian Labs

On Governance at the Substrate Level

Governance interventions at the deployment layer (model cards, API gating, refusal training) are downstream of the actual decisions that shape model behavior. The real governance lever is upstream: what enters training data, what evaluations gate releases, what architectural commitments lock in path-dependent behaviors. This paper argues for an upstream governance framework and identifies three interventions that would move the field more than any current deployment-layer policy.

April 19, 2026Telasian Labs

Compute, Data, and the Shape of the Next Generation

Scaling laws met the data wall in 2024-2025. The next generation of frontier models is characterized by synthetic data composition, post-training compute reallocation, and reasoning at inference. This paper maps the shifting compute economics and argues the dominant lever for capability gain is no longer pre-training compute scaling but inference-time reasoning depth, with second-order implications for capital allocation across the field.

February 8, 2026Telasian Labs

Architectures of Frontier AI: A Map of the Choices That Compound

A survey of architectural decisions publicly made by frontier labs over the past 24 months, with analysis of which choices compound across capability dimensions and which are marginal. The argument: architectural commitments are path-dependent in ways the field underestimates, and the visible convergence on transformer-MoE-with-attention-variants is masking a much larger decision space than the major labs are actively exploring.

Selected reading.

A short list of recent work from the labs shaping the substrate.

Interpretability

Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 SonnetTempleton et al., Anthropic, 2024

Reasoning

Learning to Reason with LLMsOpenAI, 2024

Alignment

Alignment Faking in Large Language ModelsGreenblatt et al., Anthropic, 2024

Long context

Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of ContextReid et al., Google DeepMind, 2024

Foundation

The Llama 3 Herd of ModelsDubey et al., Meta, 2024

How the lab works.

Substrate before surface. The data, evaluation, retrieval, and infrastructure layers are where the open questions live.

Rigor before reach. Deep research only. No takes, no commentary, no surface-level analysis.

Edit before we publish. Every note goes through the lab's eyes before it goes out.

Quiet before loud. The work speaks. The lab does not.

Reach the lab

Research conversations, collaborations, and reading suggestions go through the contact form.