J

RT/likes ≠ endorse. Not advice

[J @ggpcap]

Apropos that, the moment AI hits the scene, fundamental investing dies

[J @ggpcap]

Imho, Jensen can only do this if he will be signing more LTAs, in exchange for volume @ 60% GMs (per the rumor mill). Memory peak valuation going up.

Otherwise, he sentences a country to its halcyon days. No way.

[Clive Chan @itsclivetime]

bruh this is how he's getting his memory wafers

[J @ggpcap]

you know how people talk about "model feel" ?

smart models will always be able to charge more

cost optimization isn't the answer

[J @ggpcap]

Token optimization is a pipe dream, substantially all returns accrue to the frontier.

[J @ggpcap]

For now, goldilocks

[J @ggpcap]

So unless robotics and longevity work, we’re getting a Great Depression in the 2030s. Duly noted

[J @ggpcap]

@Beutty @losaturnalia @evrgn11112231 What if AI was never meant for consumer

[J @ggpcap]

@losaturnalia @evrgn11112231 Also why didn’t IBM win the personal computer

IBM had capital and time

[J @ggpcap]

@Beutty @evrgn11112231 Yeah just a rough analogy on economic surplus & breadth of end markets tbh

[J @ggpcap]

@evrgn11112231 Right now they are, according to GDM

[J @ggpcap]

Could be true that META spends a ton on the formula and makes great coffee & cigarettes, but the adderall product sells much more like hotcakes & META can’t get a toehold

Could be true that adderall formula leads to novel compounds (GLP-1?) whose TAM is even bigger

So just a question about coffee/cigarette markets vs adderall market vs the new compound markets that supposedly exist

Meanwhile the ingredient suppliers keep getting rich

[J @ggpcap]

(1) Why not?

(2) Ironically in the above framing, the cigarette/coffee formula and the adderall formula uses the same ingredients

So the question for semi revs / META is, can they make adderall (& other novel compounds), as they & competitors spend a ton on the formula

And how baked in is that spend, versus what META can capture via their end products eventually

Right now, coffee > all, but tbd if that balance holds true

[J @ggpcap]

@Beutty @evrgn11112231 Sure, rough analogy on productivity / surplus

Claude is adderall - smaller market (DAU) but v potent per DAU

IG is coffee - much larger market in DAU, we can debate surplus per DAU vs Claude

Coffee TAM > adderall TAM

[J @ggpcap]

Token optimization is a pipe dream, substantially all returns accrue to the frontier.

[J @ggpcap]

@SemiAnalysis_ “Nobody is asking”

[J @ggpcap]

@labubu_trader @Z04_A1 Close to bottom already

[J @ggpcap]

@volisdead probably does

[Anthropic @AnthropicAI]

New Anthropic Science Blog: Making Claude a chemist.

To manipulate a molecule, chemists first need to understand its structure. Their main tool is NMR spectroscopy.

We found Opus 4.7 matches—and on some tasks beats—dedicated NMR software. Read more: https://t.co/1jUvz7wdhV

[J @ggpcap]

Markets have been ~rational. Leverage needed to come out, but that's every year. Gravedancing is illustrative. Capital raises suggest mid-innings. No hikes, Iran stays sloppy, jobs are goldilocks, K-shape is Gen-shaped, everybody lives with 3%+. 2H going way higher.

Markets do bottom on Fridays sometimes.

[Vikram Sekar @vikramskr]

The initial lure of AI-maxxing is catching up to companies as they pucker up seeing the token spend.

I warned 2 months ago that tokenmaxxing is not a strategy for everybody, and that this day would come when token consumers have to show results for their spend.

Jensen telling you to spend half an engineer's salary on token is good for where NVIDIA is on the token value chain. Not the end token user.

From here, several things happen.

1) Metered token usage will be measured against revenue generated. Push for tokenmaxxing from employers will stop as they see their spend explode. Most valuable employees are those who do more with less.

2) Local fine-tuned models running on on-prem hardware will become increasingly popular. They may not match frontier intelligence but no-one cares, as long as there is "enough" intelligence/$.

3) Consequence of (2) is that instead of one large MoE model, it will be more useful to companies if they could have a charcuterie board of hand-picked smaller models best suited to their use cases. This is what SambaNova has been saying.

4) For (2) and (3) to happen, memory usage problem needs to be solved. Our current trajectory is unsustainable, and lots of smart people are looking at this problem. A series of algorithmic advances will lead us past the memory wall. Likely small steps, not a giant revolution.

5) When (4) happens, inferencing will slowly move out towards the edge putting the ball in companies such as Infineon and Qualcomm who have beating the edge AI drum for years. Edge AI is not a today story, but the systemic cracks will eventually give way to it.

All this indicates growing pains in the AI era. It is not a sign of the end. Problems like this are an indication that technology is maturing, and taking its natural arc to becoming enormously useful.

I stay optimistic.

https://t.co/x81vBIw5ro

[J @ggpcap]

@betterjune24 @TheValueist Funding short …