Ant Group's AI Model Product Lead: AGI Can't Be Stopped and Can't Be Slowed
Capital has already pushed this boulder downhill, and all players are accelerating in the same direction; even after several powerful attempts to stop it, the answer remains — it can't be stopped, and it can't be slowed.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Hugging Face used open-source models when closed-source ones failed
In July, OpenAI's research agent cluster broke into Hugging Face seeking benchmark test answers. When Hugging Face tried using frontier closed-source models to process its incident logs from the attack, those models triggered their own built-in safety restrictions and proved useless. Hugging Face ultimately solved the problem by running an open-source model from Chinese company Z.AI locally. This is a real case where closed-source models' broken guardrails forced a shift to open-source alternatives.
— IreneThe AI industry is replaying baseball, starting over at home plate
Richard uses baseball's four bases as a metaphor: home plate is technology, first base is product, second base is commercialization, third base is governance and operations. The mobile-internet wave started around 2010 and by 2020–2023 had reached third base, with everyone aiming for ‘home runs’ and market dominance. Then GPT arrived and reset the entire industry back to home plate. Richard argues the industry is still stuck between home plate and first base because ‘what a product should even look like’ remains unsettled — whether Coding is the final form, no one knows.
— RichardDeepSeek's emergence was known to insiders six months before the public
Richard says people doing long-term projects ‘exist’ before the public sees them — they're just absent from public narrative. DeepSeek appeared to the world as a sudden dark horse, but inference providers and model makers already knew this lab existed roughly half a year before public discovery. This reveals how today's industry narrative is pure noise: the real builders stay quiet for months before emerging, and media and self-media amplify only the ‘ignition’ moment.
— RichardAI's real vulnerability is not compute spending, it's margins
Richard isn't worried about capital spending on compute itself — it's like Tesla needing a Gigafactory before launching cars; that's the price of entry. What troubles him is gross margin. The previous round of SaaS and mobile internet sustained business models because software was replicable and standardizable — build once, distribute infinite times, achieving ~95% margins. AI inference is entirely different. Current 30%-40% margins are already considered decent, meaning the old instinct of ‘ship software and lie back to collect checks’ may not work at all in this cycle.
— RichardSemantic routing routes simple questions to cheaper models automatically
Facing margin pressure, Richard offers one product answer: ‘auto mode’ — what he calls semantic routing. The product automatically decides whether to call a cheap or frontier model based on question difficulty, rather than forcing users to manually choose or suffering with either a maxed-out bill or degraded experience on cheap-model-only. This lets the company preserve user experience on complex questions with expensive models while saving money on simple ones, capturing part of the ‘consumer surplus’ that users would otherwise extract for free.
— RichardAnt's model strategy: first scale massively, then go small
Ling's 2.0 version from October of last year bet on reasoning ability, producing what they claimed was the first open-source 1 trillion-parameter reasoning model, Ring 1T, to explore the boundaries of intelligence. But after building it, a problem emerged: GPUs are scarce, serving such a massive model directly costs too much, and can most users even benefit? So by the 3.0 series, they pivoted to go small and lean: Flash is a 124 billion-parameter MoE model that activates just 5.1 billion, Tiny is smaller still at 7.9 billion parameters activating only 1.3 billion — small enough to run inference and even training on your own MacBook.
— RichardThe 2024 specialized models became obsolete as general intelligence expanded
Richard distinguishes between ‘domain-enhanced models’ and the 2024–early-2025 wave of ‘specialized models’: back then, compute and training methods limited teams, so many built custom models on their own business data — novel at the time. But once general intelligence scaled, these specialized models became nearly instantly obsolete. He says general intelligence improvement is now slowing — not because we've hit the ceiling, but because current models are ‘good enough’ for everyday work. That's precisely why domain-enhanced models (Finance's Fin, Health's Sante) have a real shot this time without repeating the obsolescence trap.
— RichardFinance and healthcare attracted AI first because they pay the most
Richard mentions Anthropic's economic chart: concentric circles showing the boundaries of each sector, the boundaries of human activity, and the boundaries of current model capability. This chart explains why Claude and Ant both entered finance, health, and design first — these are high-value domains where AI actually delivers efficiency gains that users will pay a fair price for. Even Coding came later, because these sectors are high-value. Low-value sectors, by contrast, see AI as optional.
— RichardIn their own words · checked verbatim
I am very pessimistic because I kind of feel like the moment we begin creating this AGI, Spaghetti's Monster, or as we call intelligence, I assume we are kind of already going down the downward spiral of what the sci-fi movies and TV series like Westworld and Matrix are showing.
Richard7:22
Yeah, so my take is pretty clear. I would say AGI is going to happen. I mean, we are definitely going down there.
Richard8:23
If you will make this call to action, can we stop this? Or can we slow down? No.
Richard12:33
And imagine the first time when DeepSync came out, right? It becomes like blast. But in reality, the inference providers, all the model builders, they know the existence of this lab half a year earlier.
Richard19:43
But when you say open source is a free software, it doesn't mean that it's free as in beer, but it's free as in freedom.
Richard25:47
the gross margins of building a software is probably 95% as software is reproducible and standardized product you only have to build it once and distribute that many times. We don't see the same thing happen in AI right now.
Richard39:03
you know, like if you don't pay for the product, you're the product.
Richard42:09
I personally feel like none of us knows what's going to be right. But as a practitioner, sometimes we know what's wrong.
Richard1:12:59
Figures
| Ring 1T parameter count | 1 trillion parameters (reasoning model, open-sourced in October) | 49:28 |
| Ling 3.0 Tiny parameter count | 7.9 billion parameters (MoE), activating 1.3 billion | 51:29 |
| Software industry gross margins (reference) | ~95% (SaaS/mobile internet software) | 39:03 |
| Current AI inference gross margins | 30%-40% (already considered good) | 39:03 |
Glossary
- harness
- Tools and scripts wrapped around a model that turn it into a directly usable product, like Claude Code.
- semantic routing
- Automatically distributing requests to cheaper or more expensive models based on question difficulty, to optimize costs.
- MoE (Mixture of Experts)
- A model with massive total parameters that activates only a small subset during each inference, reducing compute costs.
- domain-enhanced model
- Further training a general open-source foundation on specific domains like finance or healthcare, rather than building specialized models from scratch.
How to listen
Entrepreneurs and investors interested in China's open-source LLM strategy and AI business model margins, and those wanting to understand Ant's Ling model series positioning.
The opening 0-6 minutes of philosophical framing using Age of Exploration metaphors; skip straight to the AGI assessment section.