English
Andrew Zhang
186 posts
@xikhar @ValencianaAbel I think they forgot someone. Loteraĺy burned through weekly limits in ONE day now I have 6 days to get another limit that lasts for a few hours😭
English
With the $AAOI expansion announced today, here’s what it actually means for the company.
Their bottleneck was never demand. It was capacity. $AAOI has order visibility stretching well over a year out, and $1B+ guided for 2026.
The demand is already sitting there. They just physically can’t build fast enough to ship it all.
That’s what 400,000 sq ft in Pearland, Texas fixes. Every square foot converts demand they can already see into revenue they can actually recognize.
This isn’t speculative capacity hoping customers show up. The customers are already in line.
It also changes their position in the supply chain. This is US made 800G and 1.6T production at the exact node the industry needs most, right as hyperscalers push to onshore their optics supply.
When big companies want domestic transceiver capacity at scale, the list of options is short, and $AAOI just made itself a bigger part of it.
And the timing lines up with the cycle. 1.6T mass adoption hits late 2026 into 2027. This capacity comes online right into that wave.
They’re not building for today’s orders, they’re building for the ramp they can see coming.
Companies don’t expand on hope. They do it when the orders are visible.
English
Here is why I bought more $AAOI and keep DCA every week.
People are watching the multi-week drop from the $223.10 peak down to today's $113 area and panicking, thinking it's a permanent breakdown. It’s the same old story every time a stock pulls back.
But while people are panicking over the red candles, look at the actual buyers here. Rosenblatt literally just re-issued their $220 price target with an ironclad Buy rating. The analysts who actually understand the AI optical networking cycle aren't flinching at all.
The bears have thrown every piece of emotional noise they can at this level to try and crack the floor, but the tape is holding firm. The underlying setup is lining up perfectly if you actually look at the details.
First off, the executive selling noise is totally overblown. The recent Form 4 filings showing stock sales from the CFO and regional GMs caused a quick panic. But people completely missed that these were automated, programmatic sales done strictly to cover tax withholding obligations when their equity vested. It wasn't a manual dump, and the market has already swallowed that volume anyway.
Second, the "CPO Delay" everyone is crying about is actually a massive win for us. Generalists read the recent Morgan Stanley research pushing out Co-Packaged Optics deployment to 2028/2029 and thought it was bad news. They have it completely backward. Pushing out CPO dramatically stretches out the high-margin revenue runway for high-speed pluggable transceivers and LPO. AAOI’s massive capacity ramp up to 650k units per month across their 800G and 1.6T lines by December is exactly what next-gen AI clusters are starving for.
It all points to a major financial turning point. As production volumes cross the threshold into actual volume shipments for top-tier hyperscalers, higher factory output is going to seriously expand margins and cash flow. Combine that hyper-growth AI backlog with their non-data center revenue cushion, like the massive multi-year, one-million-home Mediacom broadband upgrade win, and you are pricing a diversified supplier at a profound discount.
We’ve seen this exact movie play out on $AAOI multiple times. It ranges violently, shakes out the late-stage leverage chasers, builds a structural base at key support levels, and then snaps back.
Copper has hit its physical limits for next-gen 500k+ GPU clusters. Photons are the only answer left, and AAOI owns the in-house laser manufacturing layer.
Let people day-trade the intraday noise. The base is built. Time to take us back to All-Time Highs.
Bullish AFK $AAOI
English
The space sector did the most counterintuitive thing possible. It got its biggest catalyst in history, and the whole group got flushed.
That was supposed to be the moment the entire sector re rated.
Instead it sucked the oxygen out of the room. Investors dumped every other space name to fund their SpaceX allocation, the ETFs bled, and the whole basket got treated as one high beta pile of profit taking.
$SPCX at 145, down about 33% from highs.
$RKLB at $81, down almost 46% from highs.
$ASTS at $74, down about 45% from highs.
$PL at $26, down about 49% from highs.
This is just a few of the many stocks with massive drawdowns.
Space got killed.
English
Rosenblatt is bullish photonics and you’re selling? $LITE $AAOI $COHR $FN $CIEN
The mid-quarter weakness in optical isn’t a broken thesis, it’s a better entry.
The whole bear case rests on China ramping indium phosphide capacity and crashing laser prices, except Rosenblatt’s own channel checks say the shorts pushing that story don’t even believe it themselves. They’re already planning to cover into the late-July prints and get back into optical in 2027 for the scale-up CPO wave, which tells you they’re renting the short rather than betting on it.
The names Rosenblatt flags:
$AAOI, the one I hold biggest. Beat-and-raise setup with 2Q26 revenue seen above $200M as 800G ramps into Amazon and Oracle, 3Q guide near $300M, and gross margins pushing past 30%. Target $220 against roughly $114, and it looks like the watershed quarter they’ve been building toward.
$LITE (Lumentum) got tagged the “toughest setup,” but only because the book is already sold out through 2027 and into 2028 on hyperscaler pre-buys, which is the kind of problem you want. Target $1,300 against roughly $707.
$COHR (Coherent) has the 200G EML and internal laser ramp lifting margins, plus a roughly $1B a year 1.6T deal with Google and 11-figure NVIDIA agreements running into 2030. Target $425 against roughly $317.
$FN (Fabrinet) is seeing sequential Datacom growth with NVIDIA, new logos in Amazon and Lumentum, and a capacity ramp that’s on track. Target $750 against roughly $480.
$VIAV sold off on InP fears it has zero exposure to, which is just mispricing. Target $70 against roughly $41.
$CIEN should see DCI and scale-across orders reaccelerate, with Coherent Lite in the data center as the 2027 story. Target $720.
$ANET stays strong across Meta, Microsoft and Oracle while Google and Anthropic ramp, and Google could become a 10% customer in 2026, with revenue growth running north of 30% into the back half.
This is the interconnect layer. The people selling you the China fear are the same ones planning to buy it back in 2027, so I’d rather be there before them.
Long photonics ⚡️
English
how to meet a girl who never leaves the house (speaking from experience):
- say something interesting enough in her dms that it’s worth using one of the few hours she ever leaves the house
- offer to join her on the things she’s already doing instead of asking her to add something new to her schedule
- send her groceries, coffee, flowers, or dinner without making a big deal about it
- build a real online relationship over time before you ever ask her to come see you
- figure out the one event she gets dragged to every year and magically be there
- help her career in a way that’s actually useful
- offer to take something off her plate, whether that’s cleaning her house, organizing something, or handling an errand
- become the easiest person for her to say yes to
- never make her feel guilty for needing space or staying home
- be useful before trying to be romantic
- if she disappears for a few days (or months), don’t panic. pick the conversation back up like it never stopped
- don’t ask “when are you free?” ask “i’m already headed to trader joe’s. want anything?”
English
English
@fperrywilson @Yale given that you can just buy glp-3 without any prescription, this doesn't seem noteworthy
English
A team @Yale invented a fake patient and had him try to buy GLP-1 drugs on 49 different websites. 45 of them wrote him a prescription. A couple in under five minutes. 🧵
English
Was on a run this morning.
It was hot.
Ran by a construction site.
A few of the laborers asked if I wanted water.
Of course, I accepted.
They asked me about running in this heat. Joked that I should drink beer instead of water.
I ran up to the gas station and bought a 12 pack of Tecates and ran back by the job site.
Gave a few to the laborers, cracked open one, slammed it, and then another.
They all enjoyed their beers as well and genuinely appreciated it.
One of them offered me some fresh tortillas that his abuelita had made to help me fuel the rest of my run.
They were delicious. I was really touched by this gesture. I thought of my grandmother.
On the way home in the heat, I reflected about the opportunities that this country provides and how I understand why others will do anything to get here and I was simply lucky to be born here.
Walking in the door, I was smiling, thinking about my new friends.
People are people.
I immediately grabbed a protein drink from the fridge, sat on the couch, and thought about how there are great people from all walks of life.
Then, I called ICE and told them the address of the construction site.
English
been in ML research for 7 years, wrote a paper on compression & scaling laws, and passed openai's information theory interview
yet the latest 3b1b *still* gave me fresh intuition on entropy
either i'm an impostor or 3b1b is the greatest teacher of all time
English
@thetakencareof @TsarinaBomba airlines don't exactly make money from flying planes either
English
Are train seats packed just like economy seats on airplanes? How about the pressure from changing altitude, times for taxing, seat choices, the ability to walk around? I mean it’s not entirely just about money- there’s a certain comfort to trains. Not to mention the location of the airport vs train station. Rather you should compare profitability- train companies almost never make money (lose money all the time at tax payers expense) but airlines almost always do.
English
@Rheamund_ you usually wake up and see nothing due to the fog 9/12 months of the year
English
Seattle and Vancouver people.
You mean to tell me you guys wake up in the morning and experience this EVERY. SINGLE. DAY?
The Pacific Northwest has impressed me beyond words. You don’t ever see sights like these in the South.
Vancouver, British Columbia 🇨🇦 English
@zKatalystz @theojaffee @ryanhammer09 I don't think this is true. Pangram, in my experience, is extremely accurate. On my own writing it always scores 0. On human writing that attempts to emulate ai, it still scores 0, and if you tell fable to write like a human it still scores 100!
English
This is …. And I don’t say this lightly…. The single greatest piece of writing I have ever seen in my life
The Japanese have discovered unlimited chips & salsa and it’s beautiful
NOBUNAGA🇯🇵🏯_夏樹蒼依@japan_nobunaga
USA. A Mexican restaurant. We had not yet ordered anything, and the food was already arriving. Chips. Salsa. Unrequested. Free. I stopped the waiter. "We have not earned these." "They just come with the table, man." They come with the TABLE. In my land, hospitality is a debt. Every gift creates an obligation, weighed carefully, returned in the proper season with interest of feeling. Here, the gift arrives before you have even proven you can pay for dinner. This is not an appetizer. This is a declaration: we trust you. Eat. I ate with the gravity the moment deserved. And then — I must report this calmly — the basket emptied, and a new one appeared. "Did we…?" "Refill," the waiter said. "It's bottomless." Bottomless. They have wells of salsa. The supply lines of this nation are beyond anything my ancestors imagined. My friend warned me. "Don't fill up on chips, dude." Too late. I had accepted three baskets. Honor demanded each one be finished — an unfinished gift is an insult. By the time my actual food arrived, I was a ruined man. I was not hungry. I was not comfortable. I had been defeated by a courtesy. Generosity that arrives before the request cannot be repaid. It can only be survived. I know the rule now. I have made my peace with the basket. One basket. Two at the most. Who am I deceiving. There is no number of baskets I would refuse. The trust of a nation is in that salsa, and I intend to honor all of it.
English
@sfmcguire79 ok but what if only the top 20 students applied? is that not a huge confounder
English
Berkeley math professor:
“Today, the more successful a public high school is at preparing its students, the lower its graduates' chances of getting into top UC campuses like Berkeley and San Diego.”
Berkeley admitted 45% of applicants from a high school where nearly 94% of “students failed to meet the state standards in mathematics.”
It admitted less than 14% of applicants from a school where “nearly 100 percent of its students in AP Calculus BC pass the national exam with a perfect score of 5.”
The Free Press@TheFP
California universities dropped the SAT to help low-income and minority students. The policy is doing the opposite, writes Svetlana Jitomirskaya, a professor of mathematics at UC Berkeley. thefp.com/p/bring-back-t…
English
@deltaVee42 @a_karvonen @atticuswzf no because you can see tool calls. the compression/memorization abilities of llm is really crazy
English
@a_karvonen @atticuswzf Maybe it's just downloading it from the web when it notices it's relevant? So just (re)reading it like a human would, not memorizing.
English
I wonder if Anthropic is doing this?
Fable remembers niche appendix details from many of my papers such as the specific values of random non-standard hyperparameters.
Andrej Karpathy@karpathy
We have to take the LLMs to school. When you open any textbook, you'll see three major types of information: 1. Background information / exposition. The meat of the textbook that explains concepts. As you attend over it, your brain is training on that data. This is equivalent to pretraining, where the model is reading the internet and accumulating background knowledge. 2. Worked problems with solutions. These are concrete examples of how an expert solves problems. They are demonstrations to be imitated. This is equivalent to supervised finetuning, where the model is finetuning on "ideal responses" for an Assistant, written by humans. 3. Practice problems. These are prompts to the student, usually without the solution, but always with the final answer. There are usually many, many of these at the end of each chapter. They are prompting the student to learn by trial & error - they have to try a bunch of stuff to get to the right answer. This is equivalent to reinforcement learning. We've subjected LLMs to a ton of 1 and 2, but 3 is a nascent, emerging frontier. When we're creating datasets for LLMs, it's no different from writing textbooks for them, with these 3 types of data. They have to read, and they have to practice.
English
@__apf__ @GergelyOrosz data labeling actually costs quite a lot of money. I remember being able to pull in 5k week in 2024, and I think that was on the low end too. more generally, price has probably gone up since then.
English
@GergelyOrosz !! that's so wild it's hard to believe. how can that payroll math on cost of labeling possibly pencil out for them?
English
Just learned:
Software engineers used to do manual data labeling at Scale AI while Alex Wang was CEO. After he left, new leadership joined, and were HORRIFIED to learn this. Stopped it ASAP
Now at Meta, software engineers are assigned manual data labeling... see the pattern?
English
@ByrneHobart you can't just walk into nyu (not nowadays or ever really)
English
It's a puzzle for me, too! Why didn't I just stroll into NYU or Columbia and sit in on classes? In retrospect, it was really dumb of me not to do this.
English
The only people I've heard of who hung around a campus and audited interesting courses are Nicholas Decker and Sean Parker. If you're a believer in college-as-skill-acquisition vs college-as-signaling, this is a big puzzle!
English
why is chatgpt so good at math and coding and so terrible at being a human with any common sense
English
@synthwavedd tbf you can use opus for like 5 minutes every 5 hours anyways lmao
English
What should I cover next? Perhaps Hodge decomposition? Many possibilities!
Jack Ceroni@jack_ceroni
New blog post: my preferred method for proving Killing-Hopf: lucaman99.github.io/mathblog/killi…
English