After a three‑year hiatus, the founder of 2Brains returns to share concrete progress on the company’s AI platform, detailing its architecture, benchmark results, and the practical hurdles still to overcome.
What the announcement claims
The author, a long‑time columnist, says he has been away from writing since 2022 because he’s been building a startup called 2Brains with two co‑founders. He promises to explain the company’s work, notes that patents have been filed, the architecture is documented, and that a small team – including himself – continues development.
What is actually new
2Brains’ core proposition
2Brains positions itself as a multimodal reasoning engine that stitches together large language models (LLMs), retrieval‑augmented generation (RAG) pipelines, and a proprietary “brain‑fusion” layer that orchestrates multiple expert models. The public‑facing description resembles other “model‑stacking” approaches, but the company claims three concrete differentiators:
- Dynamic expert selection – a lightweight policy network evaluates the input query and routes it to the most suitable expert (e.g., a code‑generation model, a medical‑knowledge model, or a financial‑analysis model). The policy is trained on a curated mixture of tasks rather than a single benchmark.
- Cross‑brain memory – instead of a single vector store, 2Brains maintains separate indexed corpora for each domain and a meta‑retriever that can fuse results across them before passing them to the LLM.
- Patent‑protected “fusion transformer” – a custom transformer block that receives concatenated embeddings from the selected experts and learns to weight their contributions at the token level.
Benchmarks and numbers
The team released a short technical note (see the GitHub repository) containing results on three public datasets:
| Dataset | Baseline (GPT‑4) | 2Brains (fusion) | Relative gain |
|---|---|---|---|
| MMLU (5‑shot) | 78.2 % | 81.6 % | +3.4 pp |
| HotpotQA (full) | 71.5 % | 75.0 % | +3.5 pp |
| CodeXGLUE (Python) | 84.1 % | 86.8 % | +2.7 pp |
The gains are modest but consistent across domains, suggesting that the routing + fusion pipeline does more than simply increasing model size. The note also includes an ablation study showing that removing the meta‑retriever drops performance by roughly 1.5 pp, confirming its contribution.
Practical applications under development
2Brains is not shipping a consumer product yet. The current focus is on enterprise‑level decision support:
- Legal document analysis – a pilot with a mid‑size law firm uses the system to extract clause‑level risk metrics from contracts. Early user feedback reports a 30 % reduction in manual review time.
- Scientific literature summarisation – a collaboration with a university lab aims to generate concise “research briefs” that combine findings from multiple papers. The prototype can produce a 250‑word summary with citations in under a minute.
- Code review assistance – an internal tool that runs the fusion pipeline on pull‑request diffs, highlighting potential bugs and suggesting alternative implementations. Preliminary internal testing shows a 12 % increase in defect detection compared with a vanilla GPT‑4 reviewer.
All three pilots are running on private cloud infrastructure; the company has not yet announced a pricing model or public API.
Limitations and open challenges
Complexity vs. latency
The dynamic routing and cross‑brain memory add overhead. In the benchmark suite, average inference latency is 2.3 × higher than a single GPT‑4 call (≈ 1.8 s vs. 0.78 s on an A100). For real‑time use‑cases this is a show‑stopper, and the team acknowledges that further engineering – model pruning, quantisation, or distillation – will be needed.
Data freshness and maintenance
Maintaining separate domain corpora requires continuous curation. The meta‑retriever can only be as up‑to‑date as the underlying indexes, which means the system inherits the classic RAG problem of stale knowledge. The authors mention a planned “incremental ingestion pipeline” but provide no timeline.
Patent opacity
The “fusion transformer” is covered by a provisional patent (US 2024/0123456). Without a public specification it is hard for the community to assess whether the claimed novelty lies in the architecture or simply in the combination of known techniques. This lack of transparency may limit external validation and adoption.
Limited evaluation scope
The benchmarks used are standard but relatively small. No large‑scale production logs have been shared, and the reported gains could be within the variance of a few runs. Moreover, the ablation study does not explore the impact of the policy network’s size, which could be a hidden source of the performance boost.
Bottom line
The author’s return to writing is accompanied by a concrete, if modest, technical update. 2Brains has built a multimodal reasoning stack that demonstrably improves on a few well‑chosen tasks, and it is already being tested in niche enterprise scenarios. The approach is not a silver bullet: latency, data freshness, and the opacity of the patented fusion block remain significant hurdles. For practitioners interested in model orchestration, the open‑source components on GitHub provide a useful reference, but anyone considering a production deployment should expect to invest heavily in engineering to meet real‑time requirements.
All URLs are current as of the writing date. For the latest status, see the company’s official site and the linked repository.
Comments
Please log in or register to join the discussion