AaronWolf

[AaronWolf @ExoExplorer3]

@drwconvexity AI labs constantly improve GPU throughput to compensate for rising GPU rental price, so gross margin is intact. If throughput improves faster than cost of renting GPU, gross margin actually goes down. This is exactly what's happening inside frontier labs

[AaronWolf @ExoExplorer3]

会不会变成 long compute + short semi 的情况？感觉现在就是，hyperscalers现金流转负，但股价没有大跌，显然市场觉得未来的cloud backlog还是会兑现今天的负现金流，等于说市场还是相信AI需求持续的。但对capex二阶导有很大质疑，会增长但是会减速。而 semi 毕竟是capex的放大器，不是token proxy，所以更对capex增长的减速很敏感。

[AaronWolf @ExoExplorer3]

Hey Michael, I need an expert's take on this. My GPT-Astra matched a glass shard (B54 shard 6 at depth 75.50 m) from 168 CE sulfate event to Calbuco-group glass, specifically the mafic samples of CLD341A and CLD340C from Fontijn’s Table S3a (https://t.co/cW9Xa1CJPa).

Calbuco's Ca12 eruption is reported 31-296 CE and VEI≥4. Let me know what you think!

[AaronWolf @ExoExplorer3]

@ereliuer_eteer LLMs are computable; brains likely are too. If brains have qualia but LLMs don't, why and where is the boundary among computable processes? If substrate matters, what property of neural tissue, absent in silicon circuitry, makes qualia possible? Panpsychism seems less implausible

[AaronWolf @ExoExplorer3]

@ChinaMacroFacts @techeconomyana 不及预期是真的，现在波动性这么大，每个季度少几个或多几个$B不应该影响大的长期趋势。但现在市场已经高度警惕了

[AaronWolf @ExoExplorer3]

@lord_fed The Navier–Stokes moment for AI coding will be a swarm of 1000 agents autonomously building an entire enterprise software stack in just one week. I think we'll see this before the end of 2027. That will be the real SaaSpocalypse, not the early 2026 coding agents

[AaronWolf @ExoExplorer3]

@joinyellowbrick His fatal error is not realizing GPU rental price ≠ cost per token. Cost/token = GPU $/hour ÷ throughput. Kernel engineers boost throughput; RL researchers improve reasoning, reducing tokens per task. Both cut production costs without lowering GPU rental prices

[AaronWolf @ExoExplorer3]

@partners_road My reply reached char limit. This is about the pay in frontier labs specifically, kernel optimization is the most paid after RL research. Kernel engineers set the token throughput/GPU/hr, which is currently a limiter on frontier lab's revenue growth and token production cost

[AaronWolf @ExoExplorer3]

@GS_CapSF Ugly balance sheet, but if they prove profitability despite the spending commitments, which I suspect that they will, these are potentially $10T companies in the next few years

[AaronWolf @ExoExplorer3]

@ContrarianCurse For how much math this paper covers, it's kinda suspicious how little there is on deep learning theory and computational cryptography. Makes one wonder if OAI is deliberately holding back results with more obvious practical implications. That'd be their biggest edge after all

[AaronWolf @ExoExplorer3]

@BristOliver

[AaronWolf @ExoExplorer3]

@hoaxeye @tyrion79 @DarkCisum Coma aberration pattern from Hamburg observatory shows not all are artifacts. Astrophysical energetic events provide a good fit until the brightest and clustered ones (especially the Mar 4, 1951). That leaves undocumented upper atmosphere flash or explosion experiments by humans

[AaronWolf @ExoExplorer3]

@not_ellington Can you elaborate? Is it because linear attention?

[AaronWolf @ExoExplorer3]

@bubbleboi Last time I saw you this bullish about a stock was still $200 Sandisk and $30 Intel 😅

[AaronWolf @ExoExplorer3]

@JapanDeepValue1 @irbezek Retails certainly can. The 2015 China A-share bust is now widely considered to be caused by retail deleveraging. The magnitude and speed of the initial liquidation wave (−32% in 26 days for SSE Composite Index) were even worse than this KOSPI drawdown

[AaronWolf @ExoExplorer3]

@pcbanalysis @zephyr_z9 You previously mentioned low Dk glass is only getting worse and will last into 2028. Why HVLP is getting solved much earlier?

[AaronWolf @ExoExplorer3]

@illyquid Feel like it could be Kitagawa Seiki?

[AaronWolf @ExoExplorer3]

@ResearchQf If you look at this image closely, Feynman’s NVLink 8 is clearly called CPO, while NVLink for Rubin & Blackwell is just switches. I think here it’s implying that scale-up will go full CPO in Feynman.

[Jukan @jukan05]

Die stacking!!! custom HBM!!!

[AaronWolf @ExoExplorer3]

@rubicon59 1) Frequent write/erase shortens NAND lifetime. KV cache is very w/e heavy. Entire GPU must be swapped for a malfunction HBF
2) Above 70ºC, bits stored in NAND escape quickly. GPU interposer operates at 90ºC

Any announcement about HBF is futile, unless the constraints are solved