Research
Production field reports and open research from ScalablyAI: agent runtimes, memory, post-training, local inference, and the energy AI runs on. Each report is versioned, carries a hashed PDF, cites a published evidence register for every number, and states in full how AI was used to build and review it.
-
Is the graph doing anything?
A sourced survey of recurrent self-routing neural graphs as an alternative to layered networks
Pavle Lazic, ScalablyAIAbstract. Can a persistent recurrent graph of neural nodes, which rewires itself from its own state at every step, give more reasoning capacity per stored parameter than a Transformer? We built a quote-per-claim evidence ledger across twenty neural-network families and scored the closest published systems against the design's components. Recurrence helps per parameter, but a looped Transformer already has it, and per unit of compute the gain is small or negative. One controlled study finds recurrence adds no stored knowledge. Most parts of the design are published separately; in our scoring no system has more than seven of thirteen. The closest study compares dense, sparse, recurrent and hybrid looped mixers, and which wins depends on the task. Whether learned top-k routing adds anything beyond recurrence and sharper attention remains open, and we publish a preregistration-ready protocol to test it rather than results. Built by one founder directing AI agents and reviewed by AI models from two families, not yet by human experts.
Cite (BibTeX)
@techreport{lazic2026graph, title = {Is the graph doing anything?}, author = {Lazi\'c, Pavle}, institution = {ScalablyAI}, type = {Research / Sourced survey}, number = {SR-2026-006}, year = {2026}, month = oct, note = {v0.5, sha256 4f08f6fcac2aaee4}, url = {https://scalably.io/research/neural-graph-substrate} } -
Move the energy, not the data centre?
A falsifiable model of orbital energy for AI, and exactly what must change before space wins
Pavle Lazic, ScalablyAIAbstract. Should space serve AI first as an energy layer, beaming power or reflecting sunlight to data centres on Earth, rather than as a compute layer with GPUs in orbit? We built an open model in which every input is a sourced claim with a verbatim quote, a numbered derivation with a test, or a labelled assumption, and fixed the metric before computing any break-even: launched mass per continuous kilowatt of AI load over the hardware's life. With today's hardware, microwave beaming loses to GPUs in orbit from low and from geostationary orbit. Orbital mirrors are lighter than GPUs in orbit only with sail-density membranes aimed at an ideal receiver, or as bare membranes. At today's launch price, the launch alone of every non-speculative orbital option costs more than ground solar with storage. Photonic engines that stream weights are capped by memory energy, and our resident-weight photonic design fails its gate with measured parts, which we report as a negative result. The model, the evidence ledger and an open photonic test-chip layout are public, with the quantities that would flip each conclusion. Built by one founder directing AI agents and reviewed by AI models, not yet by human experts.
Cite (BibTeX)
@techreport{lazic2026orbital, title = {Move the energy, not the data centre?}, author = {Lazi\'c, Pavle}, institution = {ScalablyAI}, type = {Research / Falsifiable model}, number = {SR-2026-003}, year = {2026}, month = sep, note = {v1.0, sha256 32dd5b371a2a98d1}, url = {https://scalably.io/research/orbital-power-fabric} } -
Fast Memory, Slow Weights: A Production Architecture for Self-Improving AI Agents at the System Layer
Why production agents should learn continuously without continuously rewriting their models
Pavle Lazic, ScalablyAIAbstract. We run a multi-tenant AI-agent platform in production and tried to make its worker model learn from its own accumulated experience. The first consolidation attempt, two rank-16 LoRA adapters trained on 423 verified trajectory records, improved the behaviour it targeted and moved an untargeted behaviour the wrong way by more than our pre-registered gate allows. On 73 recovery states, exact correctness rose from 31 to 38 (7 adapter-only wins, 0 losses, p = 0.0156); on 160 known-failure states, exact repeats of a failed action fell from 20% to 15% (p = 0.0215); on 73 clean decisions, correctness fell from 49 to 45, a difference of +3/−7 discordant pairs (p = 0.34) that the suite cannot separate from noise but that fails a −2-point non-inferiority gate. The adapter was rejected by the gate we had written before training it. The lesson is not that fine-tuning forgets, which the literature predicts, but that at production sample sizes a single weight update produces gains and losses that cannot be told apart cheaply, and cannot be reversed separately. That is the case for learning on two timescales. A fast loop writes verified experience into an external memory that is scoped, editable, attributable and reversible, and changes behaviour on the next turn without touching weights; as of 20 September 2026 that loop holds 433 rules and 3,422 memory files across 74 workspaces, with 37 rules retired through a provenance log. A slow loop consolidates only patterns demonstrated repeatedly, through a curated training set, a frozen 484-decision evaluation, a nine-condition promotion gate and shadow deployment. We describe the implemented mechanisms, the production evidence, the failure modes the design does and does not address, and the sense in which the loop is recursive. We claim a practical architecture for recursive self-improvement at the agent and system layer under human-written gates, not a solution to continual learning and not model-level self-improvement. The effect of the fast loop on task outcomes has not yet been measured; the experiment that would measure it is specified.
Cite (BibTeX)
@techreport{lazic2026fastmemory, title = {Fast Memory, Slow Weights: A Production Architecture for Self-Improving AI Agents at the System Layer}, author = {Lazi\'c, Pavle}, institution = {ScalablyAI}, type = {Research / Production Field Report}, number = {SR-2026-001}, year = {2026}, month = sep, note = {v1.1, sha256 090bd86fd67c8bdd}, url = {https://scalably.io/research/fast-memory-slow-weights} }
Every report starts as an evidence register: one row per claim, each with its status and the file, commit, counter or verbatim source quote that supports it. The text is built from that register by a script that refuses to render a number without a row, recomputes each number from the raw artifacts or the model, and stamps the source hash into the page and the PDF footer.
Before release, independent AI reviewers, separate model instances with no access to the authors' notes, read the draft to break it: checking the evidence, reading as a hostile expert, and recomputing every number. Findings are fixed in the source and the report is rebuilt. A new version keeps the same identifier and gets a new version number, hash and date.