Emad Mostaque brought the story to the table, and Peter Diamandis put a chart on screen: a visual-reasoning system tackling colored-grid puzzles at a fraction of a cent per task. For Diamandis, the attraction was cheaper intelligence that might eventually run in everyday devices. Then he turned to fellow panelist Alex, who had read both the new paper and the earlier architecture paper behind it.
“I'll give you my hot take. I think it's a hot mess.”
The exchange on Moonshots with Peter Diamandis opened a disagreement about what counts as progress beyond transformers—the underlying design used by many leading AI models. Was an inexpensive, specialized result a reason to welcome more architectural variety, or did its complexity point away from the methods most likely to scale?
What the system actually scored
The system is BDH-CQ, described in an August 10, 2026 paper by Engdahl and colleagues. It has 150 million parameters, the learned numerical settings that shape a model’s behavior.
Its test, ARC-AGI-1, consists of colored-grid puzzles. Each task supplies before-and-after examples; the system must infer the transformation and produce the exact output for new input. This tests adaptation from demonstrations, not the full range of abilities expected of a general-purpose AI assistant.
BDH-CQ updates a recurrent memory as it processes the demonstrations: information from earlier processing is carried forward and revised rather than discarded. It then reasons through continuous internal states—numerical representations inside the model, rather than written steps in words. That is the “latent reasoning” in the paper’s title. The system can adapt to the examples through its internal state without retraining its learned parameters for each puzzle.
The authors report solving 118 of the 400 public evaluation tasks, or 29.5%, with up to two ranked candidate answers per task. They estimate inference—the computation used to produce answers—at roughly $0.00070 per task. That calculation uses about 0.85 H200 GPU-seconds at an assumed $3 per GPU-hour. It is a hardware-time estimate, not a customer price or the cost of training the model. The comparison chart mixes hardware estimates with commercial API prices, so it is not a like-for-like price list.
Training included ARC-style datasets such as ConceptARC, and implementation details remain proprietary.
The paper also reports controlled tests created after the model was frozen. BDH-CQ succeeded at copying and extending boundaries, but performance declined on longer ordering tasks and deeper nesting tasks. Better demonstrations eliminated the tested failures at nesting depth five, while only partly repairing failures on long ordering tasks. Removing semantic identifiers and mixing batches of task types produced no directional improvement or decline overall. The tests show both adaptation and limits to how far it stretches.
Two papers, different claims
BDH-CQ should not be confused with the original Dragon Hatchling paper, published on September 30, 2025 by Kosowski and colleagues at Pathway. That earlier work introduced a sequence-processing architecture inspired by locally interacting neurons. It was not the source of the 118-task result.
Nor is the distinction simply “attention versus no attention.” Attention lets a model weight which parts of its input matter to its current computation. Although the panel framed Dragon Hatchling as an alternative to conventional transformer stacks, its GPU formulation uses linear attention alongside sparse activations and low-rank transformations—ways of limiting which internal signals are active and making numerical operations more compact.
The earlier paper compared language and translation models ranging from roughly ten million to one billion parameters under matched training conditions, reporting broadly transformer-like scaling. Its experiments also exposed limits. Merging separately adapted French and Portuguese models preserved translation into English, but translation out of English mixed the languages until additional multilingual training restored proficiency.
Better data, or too much hand-built machinery?
Mostaque said he was still working through the architecture’s mechanics. What interested him was the opportunity created by better data: “people are figuring out new ways of basically turning that into intelligence.”
His optimism was about experimentation rather than a declaration that this particular design had won. Different architectures could compete to do a job, he argued, while the commercial question remained whether users would switch for a lower price.
Alex wanted something different from a transformer successor: a simpler design that benefits increasingly from more computation, with less structure engineered by hand. He invoked the bitter lesson—the argument that general methods which scale with computation tend to outperform approaches built around human-designed knowledge.
Dragon Hatchling struck him as moving the other way. He objected to its combination of particle-like elements, brain-inspired learning mechanisms and other architectural ingredients, describing it as “throwing in the kitchen sink.”
He acknowledged the appeal of a chart showing better performance for less money. His concern was that specializing a model for ARC-AGI-1 could improve that trade-off without producing abilities that transfer more broadly. He suspected benchmark-focused specialization; the discussion did not establish overfitting as a finding. The paper’s controlled tests report the more uneven picture: successful copying and boundary extension, but difficulties with longer ordering and deeper nesting.
Replacing the transformer one part at a time
Alex’s skepticism did not mean he expected transformers to remain unchanged. Asked when something beyond them would arrive, he pointed to changes already under way, including more efficient forms of attention and recurrence that carries information between processing steps.
His bet was a “ship of Theseus style replacement”: swap out the planks one by one until the design is no longer the original ship. In that account, a successor emerges through gradual component changes, not one paper announcing that transformers have been displaced.