---
title: "Unique Friend | DeepSpectrum’s Zhang Tianhe: Data First — Traffic Isn’t the Decider for LLM Training"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2024-01-16T16:00:00+08:00"
canonical: "https://ffcap.cn/en/research/20240116a05w4h00"
source: "https://ffcap.cn/en/research/20240116a05w4h00"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20240116A05W4H00"
---

# Unique Friend | DeepSpectrum’s Zhang Tianhe: Data First — Traffic Isn’t the Decider for LLM Training

_Unique Research · Ranking / Unique Friend Q&A · January 16, 2024 · Guest: Zhang Tianhe, Founder & CEO, DeepSpectrum Intelligent (深谱智能)_

Inventory type kept as **ranking** per eligible inventory (Unique Friend Q&A package; not partner welcome).

## Guest

**Zhang Tianhe** — Founder & CEO, DeepSpectrum; partner, Tongxin Shang Tech. Renmin University economics M.A.; CASS economics PhD; early China CFA/FRM holders. 20+ years finance–industry–tech: ICBC HQ credit deputy director; ICBC Leasing MD; PE fund president; listed-company director; tech co-CEO — risk, industry analysis, valuation, data intelligence.

**DeepSpectrum** — joint venture of graph-DB leader Tongxin Shang and Zhang: (1) high-performance distributed graph engines for large clients; (2) KaaS platform aggregating commercial/industry intel for SMEs and individuals; (3) graph-intelligence XAI ecosystem for AI industry fusion.

## Q&A

**1\. Product and core scenarios?**  
After informatization and internet/big-data eras, we enter “deep data” — traditional DBs/frameworks fail on massive, complex, high-dimensional data. Graph DBs — high-dimensional, flexible, intuitive, efficient — become deep-data infra (finance, supply chain, healthcare, sentiment, industrial internet). Gartner: by 2025, 80% of data & analytics will use graph tech.

**XAI Yingtu (嬴图)** — China-built 4th-gen finance-grade distributed graph engine: supernode penetration, high-density concurrency, dynamic pruning, multi-level storage/compute acceleration — real-time HTAP graph DB + high-perf graph compute. Live at major banks for liquidity risk, ALM, smart audit, real-time anti-fraud. Graph tech (white-box, controllable) complements black-box LLMs toward explainable/controllable AI.

**2\. R&D / GTM; biggest 2023 challenge and win?**  
Productizing/tooling/platformizing on the engine — real-time fund-flow analysis, TP real-time fraud intercept demand extreme performance/latency/algorithms. Challenge: buyer budget cuts and deferred demand under operating pressure. Harvest: clearer customer cognition.

**3\. GenAI opportunities vs prior waves?**  
Opportunity mainly at the app layer — proprietary data and user conversion matter more; capital logic shifts; scenes fragment — small/beautiful apps bloom; large scenes harder. Data resources are prerequisite — hard to “get traffic first, then train.”

**4\. China landing challenges?**  
(1) Legal/policy uncertainty on data elements, privacy, IP. (2) Foundation-model melee unresolved — hard for apps to pick a base. (3) Bigger dimensional-reduction risk from domestic base-model giants than abroad.

**5\. Are LLMs still too weak? / XAI thesis**  
Cites media reports of Bill Gates saying GPT-5 won’t leap like GPT-2→4 — generative AI may be near a plateau; next decade = explainable AI (XAI). His read: (1) with enough GPUs, US giants already fed available corpus; (2) under same paradigm, tuning is mostly done; (3) next focus = alignment (values/ethics/law); (4) associative/fuzzy generation has intrinsic limits; (5) next wave = vertical models + scenes needing proprietary corpora, precision, explainability, controllability. AI should serve humans — trust, anti-discrimination, fewer errors, fewer risk holes: traceable, attributable, auditable, controllable.

**6\. Human–machine collaboration?**  
Revolutionary: simple/repetitive (even some creative) work shifts to silicon assistants; headcount needs change. Cost-cut via process/product redesign must come with culture and org redesign — a systems project.

_1 CN image; attach on ingest. Signup CTA in CN omitted._

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240116a05w4h00
On-site reading page: https://ffcap.cn/en/research/20240116a05w4h00