Stubborn Persistence Might Win the Race – A Plain-English Walk-through of the Tsinghua AGI-Next Panel
Keywords: next step of AGI, large-model split, intelligence efficiency, Agent four-stage model, China AI outlook, Tsinghua AGI-Next, Yao Shunyu, Tang Jie, Lin Junyang, Yang Qiang
Why spend ten minutes here?
If you only have time for one takeaway, make it this line from Tang Jie:
“Stubborn persistence might mean we are the ones left standing at the end.”
If you also want to understand what the leading labs are really fighting over in 2026-27, read on. I have re-organised the two-hour panel held on 10 January at Tsinghua University’s AGI-Next summit into five questions people repeatedly type into search boxes.
All quotes and figures come from the on-site conversation; no outside spice has been added.
Table of contents
-
Who is splitting away from whom? – To C vs To B, vertical vs layered -
What could the next paradigm look like? – four road-maps from four speakers -
Where are we on the Agent journey? – a 2×2 matrix you can draw on a napkin -
Is there still a window for start-ups? – value, cost, speed test -
Odds of a Chinese company leading within 3-5 years – 20 % or stubborn persistence? -
FAQ – seven micro-details the audience always asks -
One cheat-sheet graphic (save to your phone)
1. Who is splitting away from whom?
| Dimension | Quote | Plain-English translation |
|---|---|---|
| To C vs To B | “Most people simply use ChatGPT as a better search box.” – Yao Shunyu | Consumers settle for “good enough”; enterprises pay a premium for every extra point of correctness. |
| Vertical vs layered stack | “To C products like tight coupling; To B wants Lego bricks.” – Yao Shunyu | End-users love one black box; businesses want pieces they can swap. |
Host Li Guangmi followed up: “Will Chinese firms copy the same two paths?”
Lin Junyang replied half-jokingly: “After Shunyu joins Tencent, maybe Tencent will pick up his genes.”
Take-away: the map can still be re-drawn by people, not only by capital.
2. What could the next paradigm look like?
| Speaker | Core view | Keyword |
|---|---|---|
| Yao Shunyu | Imagination is the bottleneck | Self-supervised learning is already happening, just hidden by the spotlight on pre-training |
| Lin Junyang | RL’s potential is not fully squeezed | Usable long-term memory still ~one year away |
| Yang Qiang | Theory must leap | Gödel incompleteness → models can never self-prove zero hallucination |
| Tang Jie | Efficiency will force the leap | Intelligence Efficiency = bigger intelligence gain per dollar |
A quick table of “efficiency”
| Old mindset | New pressure |
|---|---|
| 1 B → 30 TB data, linear gain | 1 B → 100 TB, diminishing return |
| 2 B CNY compute bill | Tiny margin, so “efficiency” becomes king |
Conclusion: if a new paradigm appears in 2026, it will be because the old one became too expensive, not because it stopped working.
3. Where are we on the Agent journey?
Yang Qiang sketched a four-stage model on the spot. Here is the back-of-the-napkin version:
| Human sets goal | AI sets goal | |
|---|---|---|
| Human writes plan | Stage 1 – today’s “copilots” | Stage 3 – AI suggests “you should do X”, you decide the steps |
| AI writes plan | Stage 2 – you say “book my Paris trip”, AI breaks it down | Stage 4 – full autonomy: goal + plan born inside the model |
Market rumour for 2026: an Agent that thinks quietly in the cloud for 3-5 hours and finishes one to two weeks of human work.
Reality check: late Stage 1, knocking on Stage 2’s door.
4. Is there still a window for start-ups?
| Dimension | On-stage words | Start-up decoder ring |
|---|---|---|
| Value | “AGI’s magic is the long tail – unsolved problems.” – Lin Junyang | If a simple API call can solve it, skip it. Hunt for questions Google still fails. |
| Cost | “Agent cost > API cost is a contradiction.” – Tang Jie | Do the maths: (token fee + idle time) < what clients will pay. |
| Speed | “Window = six months; then the base model absorbs it.” – Tang Jie | Ship in < ½ year or become a free feature. |
Quick self-test (need ≥2 “Yes”)
-
[ ] I own unique data or a locked-in vertical scenario -
[ ] I can ship in six months and still under-cut client budget by 30 % -
[ ] The problem is NOT “call API → done”
5. Odds of a Chinese company leading within 3-5 years?
| Speaker | Number | Rationale |
|---|---|---|
| Lin Junyang | 20 % (already optimistic) | U.S. compute 10-100× larger, spent on next-gen research; China spends most on delivery |
| Yao Shunyu | “Pretty high” | China replicates and locally optimises fast, but needs more risk-takers for first-ever ideas |
| Tang Jie | No number, gives the quote | “Stubborn persistence might mean we are the ones left standing at the end.” |
Bottom line: engineering and commercialisation no longer lag; the missing piece is the courage to jump into unmapped pits.
6. FAQ – seven micro-details
Q1. What exactly is blocking long-term memory?
A: Lin Junyang jokes that today’s systems only remember his name and nothing smarter. He estimates one more year for “true contextual memory”.
Q2. Cursor retrains every few hours on fresh user data – is that legal?
A: The panel only mentioned “it is happening”. If you copy the trick, check your own data-licence and privacy clauses.
Q3. How does Gödel incompleteness relate to large models?
A: Yang Qiang’s analogy: a model cannot grab its own hair to lift itself out of hallucinations. A theoretical breakthrough may birth a new compute paradigm.
Q4. Is “Intelligence Efficiency” measurable?
A: Tang Jie’s lab uses “intelligence gain per million tokens per dollar”. The exact formula is still internal.
Q5. Why is OpenAI still the most likely birthplace of a new paradigm in 2026?
A: Yao Shunyu’s view: commercial pressure dilutes innovation, but the talent density remains highest.
Q6. Academia’s opportunity – real or hype?
A: By end-2025 the compute gap between top universities and industry will shrink from 10 000× to 10× – enough for proof-of-concept work.
Q7. Are Gen-Z and Gen-Alpha really more adventurous?
A: Both Lin and Tang claim they “feel the vibe”, but no hard survey was cited – treat as anecdotal.
7. One cheat-sheet graphic
Split: To C good-enough, To B pays per IQ point
Paradigm: Imagination → RL → Theory → Efficiency, four tracks
Agent: Human goal → AI goal, Human plan → AI plan, four squares
Start-up: Value, Cost, Speed table, six-month window
China: 20 % odds + stubborn persistence
Closing – turn “stubborn persistence” into a to-do list
-
Working B2B? Calculate Intelligence Efficiency – how much extra IQ your client gets per dollar. -
Building a start-up? Stick three Post-its on your monitor: VALUE, COST, SPEED; tick a box every sprint. -
In academia? Stop complaining about GPUs; the 10× gap is now small enough to test wild ideas.
We started with Tang Jie’s warning that his generation risks being “skipped”.
The antidote is stubborn persistence – no hype, no hustle, just solid steps.
Save the cheat-sheet, close this tab, and take the next small but solid step.
Maybe at the next summit you will be the one on stage sharing the numbers we all quote.
