Pinned Tweet
Three thoughts on what really matters:
1. Fuck cancer
2. Friends are irreplaceable
3. The new "Marcus test" for AI is when AI makes a significant dent on cancer
May that happen sooner, much sooner, rather than later.
In memory of my childhood friend Paul.
Gary Marcus retweeted
Meta had an internal PR plan to try and discredit Jon Haidt because he was a critic of social media.
Meta was going to recruit academics, like Andrew Gelman, to write take down pieces. They were trying to turn academics against one another through a "whisper strategy".
This is a very serious problem for open academic debate. If academics who share negative evidence about a platform are targeted by campaigns to smear our credibility, we are in deep, deep trouble.
This creates a very serious conflict of interest for anyone who is paid to do this.
law360.com/articles/2542517/…
ha ha. some people think @edzitron didn’t post yet because he was somehow defeated.
those people are in for a surprise.
Gary Marcus retweeted
“If Anthropic’s colossal paper losses shouldn’t count in adjusted earnings, neither should the large paper gains on the other side of its deals.” 👇🏼
wsj.com/finance/investing/an…
If this is correct, Anthropic is not nearly as close to profitable as many folks seem to believe.
Instead, it appears they relied on a huge (but presumably not perpetual) subsidy from Amazon.
Anthropic spent $7.2 billion for 1.4 GW of compute in 2025 (based on reports from investor presentation).
Anthropic told investors in presentation they were tracking 5 GW in 2026. They are tracking $20 billion spent.
Anthropic spent $5.14 billion / GW in 2025
Anthropic is tracking $4.0 billion / GW in 2026 spent (4x lower than industry standard of $16-$22 billion / GW)
Anthropic just signed a contract with N-Scale at $16 billion / GW in 2026, 4x higher than its reporting in its S-1.
The magic of $AMZN deal. 75% off compute.
Their true compute cost would have been closer to $80 billion in 2026 vs $20 billion.
Anyone familiar with the tricks used during the DotCom knows one of the most popular was delay the payments as much as possible into the future while consuming a higher value of services in the present to boost operating margins, exactly what Anthropic has been doing already.
Anthropic spent $7.2 billion for 1.4 GW of compute in 2025 (based on reports from investor presentation).
Anthropic told investors in presentation they were tracking 5 GW in 2026. They are tracking $20 billion spent.
Anthropic spent $5.14 billion / GW in 2025
Anthropic is tracking $4.0 billion / GW in 2026 spent (4x lower than industry standard of $16-$22 billion / GW)
Anthropic just signed a contract with N-Scale at $16 billion / GW in 2026, 4x higher than its reporting in its S-1.
The magic of $AMZN deal. 75% off compute.
Their true compute cost would have been closer to $80 billion in 2026 vs $20 billion.
Gary Marcus retweeted
In case you've forgotten about the real AI risks, "rogue technology is not going to kill us. But it generates errors, deskills young people and, above all, threatens global financial stability, writes the former Google and Facebook investor" thenerve.news/p/roger-mcname…
this tweet of my mine was too generous and the latest OpenAI result needs some context.
read @LH’s comments below for that context.
a big deal if correct.
Here's a direct link to OpenAI's PDF claiming a refutation of ETH. In particular, it's claimed that 3SAT on n variables is in about exp(n/(log n)^{1/2+o(1)}) time
github.com/openai/math/blob/…
do see @LH’s comments below; not as a big deal as i imagined
self-proclaimed “moral” company finds every trick possible to fleece retirement funds and retail investors
not a show of confidence in their own stock.
Anthropic's IPO comes with 23 unlock dates in its first year.
Page 307 of the S-1 lists when insiders can sell:
– Day 1: certain directors, officers, employees and consultants
– Day 27
– Day 48
– Day 62
– Day 76
– Then a new window roughly every two weeks around each earnings report
– Day 180: everything left in the standard lock-up
– Day 365: everything else
How many shares in each tranche?
Blank. Every row.
Now add Nasdaq's fast entry rule. A company this size can join the Nasdaq-100 after 15 trading days. Your QQQ buys because it has to.
To be fair, staggered unlocks exist to avoid one giant cliff on day 181. That part is responsible design.
But the math is still the math.
Insiders get 23 exits over a year.
Your index fund gets one entry, at whatever price the market sets that day.
tuesday live in Baltimore: is AI good for college education?
(youtube and NPR versions to follow for those who can’t see it live)
Our full lineup is here! We’re back in Baltimore on October 27 to debate: Will Generative AI Enhance or Stunt College Learning?
Arguing “Enhance”: Rumman Chowdhury and @RaviPendse
Arguing “Stunt”: @mathbabedotorg and @GaryMarcus
Join all four debaters and Moderator-in-Chief @JohnDonvan, as they explore what generative AI means for the future of higher education.
Reserve your FREE tickets here: bit.ly/3U28sE2
Presented by @SNFAgora and @OpentoDebateOrg
And see also this request for expertise to unpack how much this does / does not ultimately matter:
When did homophobia become the official stance of the United States government?
Fake & Gay!
Moussa Diawara, the Malian national featured in this shockingly poorly researched @60Minutes report, now faces allegations of immigration FRAUD after his asylum application surfaced, revealing that he claimed to be a HOMOSEXUAL man who feared persecution in Mali.
Read more ⬇️
Would be great if a genuine expert in corporate accounting could walk through this, and explain where it leaves Anthropic with respect to how soon they might (or might not) be profitable, and how it relates to the $72 billion or so accounting charge Anthropic recently reported.
My initial S-1 thoughts:
As I expected, Anthropic is still firmly unprofitable. But, to their credit, they've significantly reduced compute costs as a % of revenue, leading to much narrower losses. On first look, there is a plausible and imminent path to profitability. This is impressive.
However, the meat of the document begins at page 161, under "liquidity and capital resources." There, we learn where the true compute costs have been hiding: future compute purchase agreements.
As of June 30, Anthropic has signed roughly $587 billion in compute leases and minimum purchase commitments that haven't begun. It has since signed $245 billion more, resulting in more than $800B in future compute costs. None of these costs have made their way onto the income statement yet. But they will, and when they do, the profitability picture will look completely different.
Anthropic acknowledges how precarious these agreements are. They point out that their compute providers (on whom they also rely for financial support) also sell services that "compete with ours." As a result, these providers/competitors may eventually force them into onerous agreements. This is something I've expressed concerns about for a while, and it's nice to see Anthropic openly recognize it.
That's why Anthropic admitted their most significant strategic decision: They're about to initiate "substantial capital expenditures." So far, they've spent zero. Why the change? Because they know they can no longer rely on leasing compute, "the magnitude and timing" of which are "inherently uncertain." So, they're going to build their own compute, which will, in their words, "change the cash flow profile of our business." I.e. make it severely negative.
It's possible that this is why they want to go public now. Since these costs have not yet "come online," they can be left as footnotes at the bottom of the document. But very soon, that won't be possible.
In sum, Anthropic's income statement is not telling you the full picture. To the company's credit, they spared us the accounting gymnastics of stripping out training costs in their non-GAAP adjustments. That would have been offensive. Still, the most important numbers aren't getting the requisite attention.
I encourage you to start at page 161 for a more complete understanding of what's going on.




















