former tech research analyst
NYC/🌏
Joined June 2026
- Tweets69
- Following59
- Followers24
- Likes63
rob retweeted
Imagine how this would’ve all played out if Jensen’s initial message had been more strategic. He could’ve frontrun the later shoe-drops and forced people to really envision Nvidia without him. Would’ve been as simple as adding:
“I let my family down and had multiple relationships outside my marriage. This is no one’s fault by mine.
One of these women is now being investigated for national security concerns. While I wasn’t aware of any such issues at the time, I was obviously not exercising good judgment and I strongly support the independent investigation initiated by our board. I will cooperate fully and accept the findings when it’s complete.
In the meantime, to avoid distracting from our important mission, I’ve asked the board for a leave of absence. Most importantly, this will allow me to focus on my wife and family.”
I really think better comms could’ve changed some decisions that are now hard to reverse, and altered the course of history
the open vs closed ai models debate is a red herring. it’s should solely be american ai vs chinese ai.
i used to believe we should accelerate ai development no matter what and was strongly in favor of all open-weight models because i didn’t want a handful of closed american labs controlling the most powerful technology ever developed.
but now i’m convinced that america needs to win the frontier ai race even if that means closed american models dominate and end users having to pay more.
ai is a dual-use technology so models that accelerate scientific discovery, automate businesses, and improve productivity can also accelerate cyber warfare, autonomous weapons, and military intelligence.
one of america’s greatest asymmetric advantages right now in the ai race is compute. the us (& its allies) controls critical chokepoints in advanced semiconductor manufacturing, like asml’s euv lithography, tsmc’s chip fabs, and sk hynix/micron’s high bandwidth memory.
china cannot replicate that supply chain overnight, regardless of how much money it throws at domestic companies like huawei, cxmt, etc.
frontier ai compute remains severely supply constrained. every blackwell/rubin gpu allocated to a chinese frontier cluster takes capacity from an american lab, effectively helping our largest geopolitical competitor.
of course nvidia wants to sell chips to china. the revenue is enormous, and keeping chinese developers dependent on cuda is a compelling business strategy, for nvidia. so cuda lock-in is great for nvidia and its investors, not neccesarily for american national security.
the alt argument (made by lutnick, sacks, and others in the pro chips-to-china camp) is that restricting chips forces china to innovate faster. deepseek demonstrated how algorithmic efficiency can partially compensate for limited compute during the scare early last yr.
although concerning, algorithmic breakthroughs can diffuse globally in weeks like it did with mla and moe techniques after that deepseek scare. american labs adopted those same techniques and ran them on substantially larger clusters. advanced semiconductor manufacturing capacity, on the other hand, cannot be replicated as quickly.
the pro-america strategy is to keep frontier training hardware out of china and close the loopholes around overseas cloud rentals/third-country intermediaries, and only allow older/downgraded chips where the commercial benefits outweigh the strategic risks.
last point, it’s likely not true that american ai progress would somehow slow down without chinese competition. we already have openai, anthropic, google, xai, meta, and others competing intensely against one another here at home. there’s more than enough domestic competition to keep pushing the frontier forward.
ai is ultimately a winner-take-all race with enormous economic and military consequences and i would rather see closed american models win rather than open chinese models win.
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
so now looksmaxxing is being called the "feminization of society"
yet in the patriarchal ancient greece, the concept of kalokagathia was highly regarded, linking physical beauty with moral excellence and elite status. alcibiades was famously handsome and his looks and charm contributed to his political influence.
unfortunately good looks has significant societal advantages, decades of academic research has demonstrated that as well.
this is correct, and it's a symptom of a feminized society where appearance matters more than substance
appearing good is more important than doing good
looks and credentials matter more than competence
feelings matter more than facts
not offending anyone matters more than telling the truth
good intentions matter more than actual outcomes
etc.
basically idiocracy, but female
99
not true, the dems have become ai’s phantom menace because of the industry’s obsequious relationship with trump. ai has had a pervasive messaging issue and anti-ai sentiment is bipartisan.
1
81
lots of positives in anthropic’s s1, its operating performance is considerably more impressive than i initially thought. biggest issue however might be duration mismatch between its noncancelable infrastructure commitments and contracted customer revenue.
anthropic is locked into years of compute spending in the hundreds of billions $$$ while most of its revenue is dependent on token consumption at rapidly compressing prices. other frontier labs and open weights will continue to push anthropic to lower its prices.
so anthropic could have the best models and grow revenue 10x, yet it could see financial impairment if the cost of intelligence falls faster than it can monetize committed capacity.
Anthropic has filed an S-1 registration statement with the SEC related to its proposed IPO.
Read more: anthropic.com/news/s1-sec
Legal disclaimer: A registration statement relating to these securities has been filed with the SEC but has not yet become effective. These securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post does not constitute an offer to sell or the solicitation of an offer to buy these securities, nor shall there be any sale of these securities in any state or jurisdiction in which such offer, solicitation or sale would be unlawful prior to registration or qualification under the securities laws of any such state or jurisdiction.
i don’t even think this inflation is primarily ai driven, but ai could become one of its biggest casualties (outside of hyperscalers)
there is debate whether hyperscalers can keep spending as financing costs rise. i think they will. the bigger risk is everyone else financing the buildout who doesn’t enjoy the same cost of capital.
we could end up in a strange situation where ai demand remains strong, semiconductor earnings keep hitting records, but leveraged infrastructure developers begin to fail as fed yields remain high/continue to rise.
the second order effects are that higher rates could accelerate ai infrastructure concentration as hyperscalers absorb distressed assets at discounts.
The markets are already moving over September CPI hitting 3.7%, but @BrianSozzi says none of that’s actually new news. 🥱
The real tell: Goldman Sachs thinks AI-driven inflation is near its peak, since memory chip prices could drop in 2027. Bad news for Micron and SanDisk, good news for everyone else.
i’ve never been convinced by miami as a tech/finance hub, despite how badly its loudest boosters wanted it to become one over the last 5 years.
the missing piece was always obvious: miami had no real talent incubator like stanford/cal in the bay or nyu/columbia in nyc.
despite going to harvard, ken griffin giving $3b to cmu, might actually change my mind.
my read is he’s treating philanthropy like capital allocation by bringing one of the world’s best cs institutions to miami and suddenly you have a credible path to producing the talent ecosystem miami has been trying to import.
anyway, happy for cmu today and proud to be an alum.
Today, Carnegie Mellon University and Ken Griffin, founder and CEO of Citadel and founder of Griffin Catalyst, announced a historic $3 billion gift to elevate Carnegie Mellon in Pittsburgh and launch Carnegie Mellon University Miami. #CMUMiami
Learn more: cmu.edu/news/stories/archive…
2
172
i asked my muse agent directly what my current usage would cost at public spark api equivalent pricing.
its estimate was ~$1.2k to $2.3k a month.
so even if meta’s actual inference cost is a fraction of that, they’re subsidizing power users pretty aggressively right now.
I am using Muse so much that I keep expecting it to tell me I have used up my capacity. This seems both encouraging but also perplexing to me. Maybe browsing the internet doesn’t take up that much compute???
1
134
if all of the allegations against kawhi leonard turns out to be true, we’ll have to put an asterisk on that clippers championship
130
we're entering a phase of runaway capability overhang in ai.
the frontier models keep making step function changes, but the average user is extracting a smaller and smaller fraction of what the models are actually capable of.
the chat interface is part of problem, and we can see that with the latest chatgpt update. but it's also an abstraction, education, and workflow problem.
the key challenge for ai companies over the next 12-18mo is prob not adding a few pts/basis pts to benchmark scores, but to close the gap between available capability and realized capability.
3
183
soccer will never be as popular in the us as it is everywhere else. america is the most dopamine addicted country on earth (see drug use stats). football provides dopamine spikes every few minutes even without a touchdown, a big hit, a deep throw, a fourth down stop. basketball scores nearly every possession. soccer asks a dopamine hungry audience to sit through 90+ minutes with little to zero dopamine hits.
the american mind cannot comprehend that europe has big box mega retailers. this carrefour is bigger than most walmart supercenters i’ve been to in the us
2
1,483
this is the classic fox and the grapes fable on cognitive dissonance. to reduce their discomfort of not being able to acquire the status of being associated with lux brands, they larp. but as soon as they have the means to acquire that status, they will, likely quietly, stop mocking the game and start playing it.
How do these kids afford their online luxury lives?
Decoding the LARP economy, and what it means for perceptions of luxury.
149
happy 250th birthday, america.
since 1776 the american economy grew faster than anything in recorded history. nobody else is close. the greatest run of wealth creation in human history.
1
89
the notion that 'ideas are easy, execution is everything' has flipped with ai. it has now become 'execution is easy, ideas are everything.'
and when i say ideas are everything, i mean good ideas. back then, having an idea by itself meant nothing. execution was the expensive part, so the only way to know if an idea was any good was to go build it.
that's the part ai changed. now anyone can execute on anything. so people do. i keep watching founders and product teams at companies of all sizes ship polished products for ideas that were never worth building, and they go nowhere. execution stopped being the thing that separated the good ideas from the bad ones, because now the bad ones get built too.
execution used to be the arbiter of whether an idea was good. now that anyone can execute, idea arbitration moves upstream, you have to know the idea is great before you even start.
1
101
i never want to root against a startup, especially in an industry as captured as insurance. so anyone who takes that leap has my respect.
but i don't think corgi will end well.
the company cites "$40m arr" for a carrier, but arr is a saas metric. that $40m is actually gwp, a liability they took onto their balance sheet, and the losses don't show up for 18 to 36 months. a 2 year old carrier growing premium fast with no loss history usually just means they're the cheapest quote, not the best priced risk. they say they were profitable last month, which is the most flattering possible moment to point at a p&l, before claims emerge. and they're underwriting d&o, cyber, trucking, payroll and sports all at once, completely different loss distributions, each needing its own risk expertise.
also don't love that they raised at $1.3b then $2.6b three weeks later, same lead (tcv), so no real market clearing price. and don't get me started on the cafes or the street runway fashion shows. yep, you read that right.
i hope i'm wrong, but i've seen this story way too many times.
strongly agree. building something doesn’t mean deferring every good thing until the exit. life is fragile, with a non-zero chance we’re gone tomorrow, so i lifemaxx in parallel and also try to enjoy the build itself
"I'll travel once I hit my next milestone"
​proceeds to delay every vacation, stay glued to a laptop on weekends, and burn out before 40
Meanwhile people outside the bubble:
390 days
4 continents
29 countries
…for just $57k ($28.5k per person)
Stop waiting for the perfect time. Start living.
You don’t take your bank statements to the grave
166
nfa.
the zero call ignores that this cycle is tracking the last two almost to the month. top in oct 2025, ~17 months after the halving, same window as dec 2017 and nov 2021.
both of those bears bottomed ~12 months after the top. that puts this cycle’s bottom at around october 2026, and new cycle keys off the april 2028 halving, roughly 22 months from here.
the tide is still retreating tho and so we still do not fully know who, if anyone, is out there swimming naked. the dats are my main watch. ~230 public companies on ~820k btc, strategy the bulk of it. if their premiums break they flip from buyers to forced sellers. could be this cycle’s version of the ico bust or terra/celsius/ftx. again, nfa
so more downside, sure. zero, prob no. base case is we are closer to the bottom than the top.
The Wall Street legend who called the dot-com crash just said Bitcoin is going to ZERO.
He's watched every major bubble for 60 years, and he thinks crypto is the dumbest one yet.
His name is Jeremy Grantham. He managed $165 billion at his peak, predicted multiple crashes, and gave away more than 90% of his fortune.
But he owns no crypto and has never owned any. He said he'll never advise a single person to buy it and that Bitcoin will "certainly go to zero," though it might take years to get there.
His reasoning was simple:
It's not a store of value, because a real store of value doesn't fall from $120,000 to $60,000 on a whim. And it's useless as actual money, because you can't walk into a shop and spend it without friction.
He said it does exactly one thing well, which is let people speculate, and called the rest of it an unnecessary piece of nonsense that mostly helps criminals move money out of sight.
You can agree or disagree with that - but the bigger point is what he said right after:
Since 1929, the Goldman Sachs's of the world have never once told their clients to get out of an overpriced market. They stayed quiet before the 1929 crash, before the 1972 collapse, and before the dot-com top. Grantham called it a flat-out betrayal of trust.
In 1999 he stood in front of 400 full-time market professionals and asked them whether the market returning to normal valuations would guarantee a crash. EVERY one of them said yes. Nearly all of them privately believed it was already coming.
At the very same event, the executives above those analysts were on stage telling clients to relax and stay invested.
These weren't outsiders either. They were the engine room of the biggest firms on Wall Street, the analysts doing the real work, and they already knew.
The people representing them on stage knew too. They simply had a business reason to stay quiet.
Warning a client early is bad business, because a client's patience runs out long before a bubble does. When Grantham warned his own clients, he lost half of them in just over two years for being right too soon.
So the industry settled on a rule it still lives by, which is that you can be wrong as long as you're wrong in a crowd.
Jump off the cliff with everyone else and you keep your job. Warn people alone and early and you get fired before you're proven right.
Now connect that to crypto:
Wall Street sells crypto today. Spot Bitcoin ETFs, crypto funds, custody products, and a fee on every single one of them. The same industry that stayed silent before every bubble for a century now earns money every time you buy the asset Grantham says is worthless.
The people who never warned you about the last bubble are the last people who will warn you about this one, because they finally found a way to profit from it.
So the real question isn't whether Grantham is right about Bitcoin. It's which of the people advising you loses money IF he is.
Do you agree with what he's saying?
359