Tory Green@MTorygreen · 2026/08/27赛道影响Memory is becoming a much bigger part of the AI scaling problem.
J.P. Morgan estimates DRAM prices could be up 400%+ from early 2024 through end-2026, while SK hynix expects the shortage to persist through 2030.
That pressure is already shaping both sides of the stack. Models are leaning harder on MoE, quantization and KV-cache efficiency, while Nvidia’s Rubin pushes to 288GB of HBM4 and 22 TB/s of bandwidth per GPU.
We spent years treating FLOPs as the main scaling variable. For long-context agents and inference-heavy workloads, squeezing more intelligence out of every GB may matter just as much.
AI:偏主题投资表达,释放对基础设施与新叙事的偏多信号。
Tory Green@MTorygreen · 2026/08/27观点输出Day zero!!
320B total params, just 18B active per token, and GLM-5.3 Flash is still landing near frontier coding/agent performance.
Stats:
> 84.3 Terminal-Bench 2.1
> 63.4 DeepSWE
> 48.8 AutomationBench
> #1 GDPval-AA v2 in its comparison
It’s also the same model that pulled 11T+ tokens on OpenRouter before anyone knew who built it.
That’s a strong case for where MoE is heading: keep a huge amount of capacity in the model, but only spend compute on a small slice of it per token.
And that’s where @ionet fits too.
Models are getting better at using compute more efficiently. Our job is to make the compute they still need easier to access for everyone.
Cheaper intelligence + more accessible infrastructure is a pretty powerful combination.
AI:延续个人公开立场,强调长期主义、执行效率与行业方向判断。
Tory Green@MTorygreen · 2026/08/27观点输出@ajs6888 true, this was def a memorable launch
AI:延续个人公开立场,强调长期主义、执行效率与行业方向判断。
Tory Green@MTorygreen · 2026/08/27赛道影响@adelbucetta @AnthropicAI curious to see what'll come out of it
AI:偏主题投资表达,释放对基础设施与新叙事的偏多信号。
Tory Green@MTorygreen · 2026/08/26赛道影响@ArtificialAnlys @Zai_org I’d love to know the actual serving economics here. Burning more reasoning tokens than GLM-5.3 while charging 7x less per task is either an incredible infra advantage or an incredibly aggressive subsidy.
Both are interesting, but they mean very different things.
AI:偏主题投资表达,释放对基础设施与新叙事的偏多信号。
Tory Green@MTorygreen · 2026/08/26赛道影响If this really is Fable 5.1 being tested, @AnthropicAI probably needs more than a normal point release here.
Open weights have gotten dramatically cheaper, the Pareto frontier moved twice this month, and Opus 5 already eats into the reason to pay Fable pricing.
So 5.1 kinda has one job: reopen a capability gap big enough that the premium feels obvious again.
If it ships with another 5–10% incremental bump, I’m not sure that changes much but if Anthropic has been holding back a real step-function jump (which they normally do), then this gets a lot more interesting.
Excited to see how this one lands.
AI:偏主题投资表达,释放对基础设施与新叙事的偏多信号。
Tory Green@MTorygreen · 2026/08/26观点输出@AndrewCurran_ twice in one month the Pareto frontier moved hard to the left. Qwen3.8-27B did it first and now GLM-5.3 Flash. quite impressive
AI:延续个人公开立场,强调长期主义、执行效率与行业方向判断。
Tory Green@MTorygreen · 2026/08/26观点输出22T+
https://t.co/whfVoL3JHJ
AI:延续个人公开立场,强调长期主义、执行效率与行业方向判断。
Tory Green@MTorygreen · 2026/08/26观点输出@OZinzinn don't think so
AI:延续个人公开立场,强调长期主义、执行效率与行业方向判断。
Tory Green@MTorygreen · 2026/08/26赛道影响RT @Zai_org: Introducing GLM-5.3-Flash
- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context…
AI:偏主题投资表达,释放对基础设施与新叙事的偏多信号。
评论
暂无评论