-
products
2026-05-14
For the first time AI went viral thanks to the application, not the model; users pay for getting things done.
«For the first time, what really made AI go viral was not the model, but the application.»
-
safety
2025-04-25
DeepSeek is not omnipotent: text-only, high hallucination rate, and above all slow and expensive.
«DeepSeek is not omnipotent... high hallucination rate... its bigger problem is slow and expensive.»
-
open source
2025-02-14
↺ muutti kantaa
REVERSAL: Baidu announces open-sourcing the Ernie 4.5 series from June 30 (DeepSeek effect).
«We will roll out the Ernie 4.5 series and officially open-source it starting June 30.»
-
compute
2025-02-13
Today LLM inference cost drops over 90% every 12 months.
«Every 12 months, inference cost falls by more than 90%.»
-
products
2025-01-02
No super app yet, but AI penetration is not low and exponential growth continues in 2025.
«no super app so far, but AI penetration is not low and exponential growth continues in 2025.»
-
china us
2024-11-20 ~
Mainstream US treats AI like a Manhattan Project; we care more about applications.
«Mainstream US treats it like a Manhattan Project... For us, we care more about applications.»
-
products
2024-11-12
AI agents are the most mainstream form of AI applications and are about to explode.
«AI agents are the most mainstream form of AI applications, about to reach their explosion point.»
-
safety
2024-11-12
The biggest change in 24 months is that large models have basically eliminated hallucinations.
«Large models have basically eliminated hallucinations.»
-
products
2024-07-04
Without applications a base model alone — open or closed — is worthless; stop competing on models.
«Without applications, a foundational model alone — open or closed — is worthless.»
-
compute
2024-07-04
In 2023 over 100 LLMs competed in China, wasting resources, especially compute.
«In 2023 over 100 LLMs competed in China, a significant waste of resources, especially compute.»
-
open source
2024-07-01 ~
Peak closed-source stance: open source is essentially an 'IQ tax'; closed models are stronger and cheaper.
«Open source is actually an IQ tax.»
-
agi timeline
2024-05-23
AGI is still more than 10 years away; today's most powerful models are far from it.
«AGI is still quite a few years away, I think more than 10 years away.»
-
open source
2024-04-11
Open models' significance is minimal; closed models are stronger and cheaper; the best base models in China and US are closed.
«The best foundational models in China and the US are all closed-source.»
-
products
2023-11-15
Constantly rebuilding foundation models is a huge waste of resources; millions of AI apps are needed.
«Constantly redeveloping foundational large models represents an enormous waste of social resources.»
