11,355 karma · joined August 11, 2008
I think the point stands: MoE, a myriad of complex attention approaches, shared layers, you name it. And making it all work together well is a huge trial-and-error pain even for small models, never mind getting to efficient hardware utilization.
> "Some neo-clouds worry that they can’t stray from buying Nvidia’s full stack of hardware for fear of being put in “Jensen jail,” meaning they might lose their allocations of Nvidia chips, said Adam Fisher, a partner at Bessemer Venture Partners."
> It seems Jensen Huang is effectively threatening neo-clouds implying that Nvidia could cut their chip allocations if they don’t buy Nvidia’s full stack.
The European Union downsized plans for data centers to support AI, outlining a call for tenders that includes smaller-scale facilities than originally envisioned.
The EU is seeking bids to build 4 data centers with at least 25,000 GPUs and 3 with at least 40,000 processors in the first phase.
The original concept called for as many as five data centers, each with 100,000 advanced chips - Bloomberg
The future seems more like Blindsight [1]: hyper-intelligent, completely unconscious systems outperform, out-manipulate, and out-compete human beings purely through automated efficiency.
It will be bittersweet when there's no human needed at the wheel but IMHO we are far, far from that. These models/agents are just mimicking human text and need guidance because they often get lost or stuck.
"Exclusive: OpenAI Losses Increased Nearly 8X in 2025, With Spending Hitting $34 Billion" https://news.ycombinator.com/item?id=48550465 (188 points, 2 days ago, 108 comments)
https://www.wheresyoured.at/exclusive-openai-financials/
Please stop posting Ars. It's just blogspam, sad as that is. We need to let it go.
"How Much Is Too Much? Controlling Administrative Costs through Effective Oversight" (2017) https://www.goacta.org/wp-content/uploads/ee/download/contro...
For the past 20 years the budgets ballooned out of control (alongside the student debt). Yes, this WH admin is anti-science but US academia is due some introspection.
Disclaimer: I'm not from US
Those are SM120 so no tmem/tcgen05 and lack of support in main libraries (it's like everybody is focusing on B300/SM100).
For that money I'd buy a single B300, similar total AI TOPS, similar GPU bandwidth aggregated, and only 25% less total memory (probably saved in less implementation complexity), half the energy consumption...
Also by having all SMs local they have the special L1-level interconnect. SMs can collaborate on the same GEMM. And a bunch of other nice features.
Or, you know, rent it.
https://xcancel.com/jukan05/status/2066752564809605584
DeepSeek Completes First External Funding Round, Raising $7.4 Billion
- DeepSeek has completed its first external fundraising round, raising more than RMB 50 billion (approximately $7.4 billion). The company was reportedly valued at over $50 billion.
- Investors are investing through a limited partnership (LP) managed by CEO Liang Wenfeng, rather than directly into DeepSeek itself. This structure allows Liang [CEO/founder] to maintain absolute control over the company.
- The China National AI Industry Investment Fund is the sole exception, investing directly in DeepSeek and receiving voting rights. The fund has committed RMB 1 billion.
- Ordinary investors receive no voting rights, but gain access to financial information and priority rights in future financing rounds. All investor stakes are subject to a five-year lock-up period, designed to discourage short-term profit-taking.
- Major investors include Liang Wenfeng (RMB 20 billion), Tencent (RMB 10 billion), CATL (RMB 5 billion), JD.com, NetEase, and IDG Capital (RMB 3 billion each).
- DeepSeek’s management reportedly conducts *due diligence* on the identities of LPs backing investment funds, aiming to prevent the entry of unwanted investors.
- DeepSeek had previously operated without external funding, but rising compute costs and intensifying competition for AI talent have increased the need for capital.
Still, LLMs are extremely powerful pseudo-AI [2] and will bring a pseudo-singularity. But the impact is still scary if a tiny fraction of humans are augmented 1000x. And as better models become exponentially more costly, only the money people will be able to afford the new models. This is a very likely scenario and scares me to the point I dropped all my projects to work on affordable LLM-based tools to make the difference at most 10x instead of 1000x.
To my elder relatives I explain it like: imagine we are farmers in the 17th century and suddenly out of nowere John Deere tractors, combines, etc. become available. But they cost more to run than all you and your fellow farmers have, so only a tiny handful of rich people take over everything.
I don't know ffmpeg but this resonates with my experience with other open source projects.
FWIW, for the equity part there's a proposal expected to pass for next year: https://www.eu-inc.org/ (but it doesn't address taxes, cross-border employment, or anything significant so it's mostly moot). The main goal is to attract native VCs.
AFAIK it's designed by lawyers and old money, with little to no input from tech entrepreneurs.
I think Evans is completely wrong. There are only 2 truly frontier models. (at least for now). And Anthropic seems to be leaving OpenAI behind so there might be only 1 in the near future. (which is scary/dangerous)