Shin

Medical AI Startup, Ex-Deepmind

[Shin @Shin0kei]

@aleabitoreddit 😂😂😂😂 this is going to be even more rage baiting

[Shin @Shin0kei]

@Alisvolatprop12 곗냐고요... 페이블은 지금 ... 6개월전 모델인데

[Shin @Shin0kei]

@aleabitoreddit Actullay should be banned.... how is this legal?

[Shin @Shin0kei]

@aleabitoreddit Another fun one is to see people like @bubbleboi and other retards who thinks photonics have no purpose LOL

[Shin @Shin0kei]

@gurgavin Don't blindly copy... spreading rumors

[Shin @Shin0kei]

@ParadisLabs Hmm is the diff due to one searching reddit

[Shin @Shin0kei]

@Alisvolatprop12 신기하네요... 중국에서는 ���제 꽤 잘되는걸로 알고있는데

[Shin @Shin0kei]

@FirstSquawk People just don't understand... They are using images/videos for Gemini. Its obvious that in this field Gemini is #1.

[Shin @Shin0kei]

Everyone is dunking on Meta right now, and for once it has nothing to do with privacy.

People just realized that Meta, the same company pouring $600 billion into its own AI buildout, has quietly been running on Google's Gemini. The internet's reaction was basically: "Wait, META? Zuckerberg is paying his BIGGEST rival to borrow a model?"

But here is the part nobody is talking about.

Meta did not do this for fun. Gemini was simply better than their own Llama at the ugly, unglamorous work. We are talking about catching scams, ripping down harmful content, and moderating the endless firehose of posts flooding Facebook and Instagram every single second.

So Meta looked at its own models, looked at Gemini, and decided to pay Google to help police its own apps. That is how far ahead Gemini had pulled.

Then came the twist. Google just capped Meta's access. Not as a flex, and not out of spite. They literally ran out of compute. Demand for Gemini got so heavy that Google could not hand over the capacity Meta was asking for, so they started rationing it.

And it gets better. The shortage is so severe that Google is now renting compute from Elon Musk's SpaceX, reportedly to the tune of nearly a billion dollars a month. Sundar Pichai has openly admitted that Google's cloud revenue would have been even higher if they could simply keep up with demand.

Meanwhile, Meta is scrambling. It is pushing its own in house model called Muse Spark to cut the dependence, and it has quietly told staff to stop burning tokens and start being efficient. Quite the reversal for a company that was recently grading employees on how much AI they used.

So let that sink in. The AI race has stopped being about who has the smartest model. It is now about something far more basic: who can actually keep the lights on.

Because the smartest model in the world means nothing if you have nowhere to run it.

[Shin @Shin0kei]

@Alisvolatprop12 아직도.. 진짜 많은 사람들이 (99.9% 전체인구) 모델 하나 직접 안돌려봐서 일어나는 일이죠

[Shin @Shin0kei]

Chinese typically being Chinese. I feel like a lot of these people are sponsered by Chinese companies like Alibaba or Zai...

Also, I feel like the fact taht Zai reaching mythos level security is fake. Zai has come very close, but not really that level yet.

Like Elon said... the Actual usefulness hasn't really come that close in reality for complex questions.

[Shin @Shin0kei]

@firstadopter Mostly true except... we shouldn't directly trust the quote that it works as much as mythos level?

[Shin @Shin0kei]

@ylecun @pmddomingos Disagreed. You can recover it easily with the right hardware scale which NO ONE else has. ESP anthropic or openai

[Shin @Shin0kei]

I still wonder how Waymo got so much permission ahead when it is so clear how much it MALperforms everywhere... Assuming it probably is Google connections with politics?

[Shin @Shin0kei]

A follow-up on Apple's "hundred-year flood." Funny thing about floods: some people saw the clouds and built an ark. Others are standing in the rain acting surprised.

While Apple spent two years squeezing memory suppliers on price, the people who actually understood what was coming got on a plane to Korea.

Jensen Huang flew to Seoul, sat down with Samsung and SK Hynix, and locked up supply early. He reportedly pushed SK to deliver 12-layer HBM4 six months ahead of schedule, and Nvidia now holds the majority of SK Hynix's HBM output through 2026. Both SK Hynix and Samsung are already sold out of 2026 HBM. Sold out.

Elon Musk did the same thing. In July 2025 Tesla signed a $16.5 billion chip deal with Samsung, with a Texas fab dedicated to its next-gen AI chip, and xAI is now in talks with Samsung too. Musk even agrees with Cook that this is "the biggest price jump in anything I've ever seen." The difference? Musk locked his supply in BEFORE he said it.

Then there is OpenAI. The Stargate deal alone has Samsung and SK Hynix committing up to 900,000 DRAM wafers a month, roughly 40% of the entire planet's DRAM output, secured a year ahead with a long-term commitment.

See the pattern? The smart ones treated memory as a strategic asset and went and secured it, in person, early. Relationships, volume commitments, real partnership.

And Apple? Apple treated its suppliers as a vending machine to beat down on price, then switched to quarterly haggling, then started shopping Chinese memory for leverage, and is now "renegotiating contracts" and scrambling for allocation while raising your prices to cover the gap.

That is the real story. This was not bad luck. The allocation went to the companies that showed up, paid up, and planned ahead. Apple optimized for the lowest possible price, and is now learning that when the music stops, the cheapest customer is the first one left without a chair.

The smart ones are getting their chips. Apple is getting a press release about a "flood."

@elonmusk and @nvidia saw it coming and moved. @tim_cook @Apple read the same data and chose the spreadsheet. Now you pay for it.

[Shin @Shin0kei]

Apple is raising prices on almost everything and blaming a "memory shortage." Tim Cook calls it a "hundred-year flood… I've never seen anything like it in over 40 years."

Here's the part he leaves out: Apple helped build the dam that just broke.

On June 25, Apple quietly raised prices across Mac, iPad, Apple TV, HomePod and Vision Pro. Same chips, same storage, same RAM, just a bigger number on the tag. The official reason? Memory costs have gone vertical. And that part is true: contract DRAM rose ~90% in Q1 2026 and another ~60% in Q2, roughly 4x what it cost three quarters ago. TechInsights estimates the RAM in a single Pro iPhone could jump from ~$39 to ~$145.

But ask WHY memory got so expensive, and the story stops being convenient for Apple.

Rewind to 2022 and 2023. Memory prices crashed below production cost. SK Hynix posted a net margin of about -28%. Micron cut wafer starts ~20%. Samsung slashed output ~50%. The whole industry gutted capex and froze new fabs. That underinvestment is precisely why supply can't keep up today.

And who held the whip in those years? The buyers. Apple, Dell, HP, giants big enough to pit Samsung vs SK Hynix vs Micron against each other. DRAM is a commodity; volume wins; so they squeezed, contract after contract, until the makers were selling at a loss. Micron's own exec said it on the record: "We told a couple of the customers who were being very aggressive with pricing… that this is not constructive."

Translation: you pushed us below cost, we stopped building, so don't act shocked when the well runs dry.

So forgive me if "hundred-year flood" rings hollow. A flood is an act of God, nobody's fault. A shortage built on a decade of squeezing your suppliers until they can't afford to expand isn't weather. It's a bill coming due.

Now watch the table flip. Micron's gross margin just hit ~75%, guiding toward ~81% next quarter, roughly triple its 2024 level, with its entire 2026 HBM output already sold out. For once the people who actually make the chips hold the leverage. And suddenly THAT'S the crisis?

When buyers dictated below-cost prices for years, that was just "the market." When suppliers finally price to demand, it's a "tax" consumers must be shielded from, by charging consumers more.

Here's the uncomfortable logic: a buyer offers $50, a desperate seller takes $60, but someone else bids $80. Is taking the $80 "gouging," or were the years of forced $50 the real anomaly? A company maximizing profit isn't a scandal. It's Tuesday.

Memory makers aren't saints; they're maximizing too. But there's a difference between a supplier finally getting paid fairly and a ~$3 trillion buyer who engineered rock-bottom prices for years, then plays the victim while quietly adding 20% to your cart.

So when Apple tells you the price went up because of "memory," remember the fuller story: the shortage didn't fall from the sky. It was negotiated into existence, one ruthless contract at a time, and now you're holding the invoice.

Hold the buyers accountable, not just the chipmakers. @tim_cook @Apple, own the whole story, not just the convenient half.

[Shin @Shin0kei]

@KrisPatel99 LOL CXMT is same price as the other 3 mostly + they have it mandatory to use a lot of items that is made in china (meaning Apple has to go against Huawei)

It will be that Apple will not be able to see 50$ ram for 250$

[Shin @Shin0kei]

그건 아니죠ㅋㅋ. 정확하게 말하면 애플이랑 삼성이 손잡고 마이크론 망하게 한거라는건 모든사람이 다 아는데 그때 손해본걸로 공장을 못지어서... 통신이 메모리 ㅈ 되게 한건 다아는사실인데

이번에는 반대로 간 케이스가 된거죠.

아마 여기서부터는 하이킹은 힘들겟지만 정부차원에서 아마 이가격 근처에서 유지시키겟죠. 아니면 펩 지을 돈을 지원해준다던가

[Shin @Shin0kei]

The more I dig into ASTS vs SpaceX, the more I see a huge bear case on ASTS long-term. $SPCX / $RKLB vs $ASTS

Everyone focuses on ASTS having better satellite design (those massive phased arrays).

That’s real! They have a clear edge in per-satellite D2D performance today. But, I think the actual moat in this industry is shifting.

It’s no longer mainly about who builds the best satellite.

It’s about who can cheaply and repeatedly launch very large, complex satellites at high cadence.

That’s exactly what Starship gives SpaceX. Once it’s flying regularly, they can put up much bigger satellites with bigger arrays, way cheaper, and way faster than anyone relying on Falcon 9 or New Glenn.

ASTS is basically betting their entire future on specialized satellite design while being dependent on third-party launches.

SpaceX controls the entire stack: design + build + launch.

This reminds me of the GPU/AI world. Having great chip design is nice, but if someone else controls the critical infrastructure (like HBM supply or massive compute), they eventually win on scale.

Design matters. But launch capability is becoming the bigger bottleneck.

I’m bearish on ASTS relative to SpaceX here. $ASTS