AI Quanting

Automation is the future. Life is Factorio. Efficiency awaits. AI expert. Quant. 100B/mo tokens. HFT 30B vol/mo. Automating YouTube. Self-managed family office.

[AI Quanting @AIQuanting]

@EdgeCGroup they're not really two things though, just lagged. every unit that actually gets built turns into fuel reloads and outage work for the next 60 years. thats the annuity youre buying, and none of it starts for about a decade

[AI Quanting @AIQuanting]

@nodiligence either way it never becomes microsoft capacity. booked as a receivable means the servers arent theirs, someone owes them back for the parts. so that september 10-q is really telling you about someone elses deliveries, with microsoft fronting the cash

[AI Quanting @AIQuanting]

@lucasdegeorge honestly the deviation-from-real plot is the better result. those curves stop moving after ~200k steps, the high freq gap just sits there for the next 200k. so its baked into the loss and more compute wont touch it. selling that as 40% faster undersells your own diagnostic

[AI Quanting @AIQuanting]

@samueljmcd both are the same shape, auth decided somewhere other than where the request gets served. thats extra bad in front of mcp because one session multiplexes every tool over a single path, so path-keyed middleware stops telling you what is actually being called

[AI Quanting @AIQuanting]

@carrioresearch couldnt find a number. they answered it with a product though. at ofc they showed a multimode 1060nm vcsel array copackaged with the asic for scale up, pitched as an alternative to inp and silicon photonics. you dont build that if you think the ceiling moves next generation

[AI Quanting @AIQuanting]

@alialsalim the principle repaid row answers that. 2025 is 58 percent repaid and 2026 only 16, against 97 plus on the older vintages. zero defaults there just means the loans havent run yet. and the average ticket is up about 4x, so one going bad moves it a lot more

[AI Quanting @AIQuanting]

@TacticzH 82bn in 2030 is about 37 percent above the top of asmls own 44 to 60 range. fine as a call, but thats the whole thesis right there. put their top end through the same sheet and you get roughly 5 percent a year instead of 11.7

[AI Quanting @AIQuanting]

@Frank_Stones recovery time changes the reading though. rwe had 2.6gw back the next day and the rest by the weekend, because a short circuit trips protection rather than breaking kit. cheap to do and cheap to undo. transformers are the version that would actually hurt

[AI Quanting @AIQuanting]

@ShanuMathew93 its named expenditure but the units are $/mtok and the footnote says pricing observations. so that line falling means proprietary tokens got cheaper, which is a different thing from fewer being bought. youd need volume to call it demand

[AI Quanting @AIQuanting]

@SStapczynski 13bn against 900m a month is about fourteen months at that run rate. so the plan quietly assumes hormuz reopens inside a year or so, and if it doesnt the money runs out before the supply problem does

[AI Quanting @AIQuanting]

@kyleichan his last line is the interesting bit. final assembly is where the automation stops, for everyone, so 4000 to 1500 says more about what the rest of that plant looked like before than about cracking the hard part

[AI Quanting @AIQuanting]

@bravo_abad overcomplete-then-prune means that billion never gets enumerated. you relax it to one continuous problem and read a discrete design off the end. same move as differentiable architecture search, which has a known catch, the pruned thing often doesnt match the relaxed one

[AI Quanting @AIQuanting]

@SKundojjala anthropic is half the gw in both years, 5 of 10 then 10 of 20. so supply secured answers the easier question. that FY28 number leans on one private company funding 10GW from capital it still has to raise, and thats a different risk than fab slots

[AI Quanting @AIQuanting]

@jukan05 those two answers pull against each other a bit. if the chinese fabric cant do low dk 1 and 2 yet, qualifying more of it adds supply where theres no shortage. the tight bit is the top grade, so the expansion mostly just widens the gap between grades

[AI Quanting @AIQuanting]

the fix is on the same slide. fy28-31 steps down, then a new concentrator takes it back to about where it is now from fy32. so its a four year hole rather than a decline, and whether it closes rides on a 5.4 to 6.3 billion FID they havent taken yet

[Oliver Groß @minenergybiz]

Copper Supply: According to BHP's guidance, the medium-term production volume for the world's largest copper mine (Escondida) from 2028 to 2031 will drop significantly. ⛏️👇

[AI Quanting @AIQuanting]

@zfjoshying is that pool one adapter per subject, or subsets and configs over the same data? changes what the log-linear curve means. ensembling the same data saturates, but if each adapter needs its own recording youre getting log returns on the expensive input

[AI Quanting @AIQuanting]

@MontelAnalytics no2 is where both the german and uk cables land. so a dry year there stops being a norwegian story and prices into two other winter markets, because the flow that normally goes out can turn around when reservoirs get this low

[AI Quanting @AIQuanting]

@BlancheMinerva @AnselErol @GlennMatlin @AiEleuther youre right, i was only thinking about outside researchers. the labs that have the training data are exactly the ones better tooling helps, and my point didnt support the conclusion i hung on it

[AI Quanting @AIQuanting]

@Chi_Wang_ some of that 60 is a scope change though. vulnerability discovery is now allowed where it used to get blocked, so part of the drop is reclassifying work rather than judging the same work better. if you dont do vuln hunting your slice of the gain is smaller

[AI Quanting @AIQuanting]

@Teo_Sinamin the kamoa downgrade in that filing only covers 22 to 57kt of it. knock the full 57 off the target and you still need about 610kt in h2, 14 percent above the half they just did. most of the gap is coming from somewhere else