Pinned Tweet
Richard Feynman and the 19 year old me at my Caltech graduation.
Nano Banana (Gemini) cleaned this image - removing scratch marks.
Original: x.com/hsu_steve/stat…
Nano Banana (Gemini) cleaned this image - removing scratch marks.
Original: x.com/hsu_steve/stat…
Richard Feynman and the 19 year old me at my Caltech graduation (Found this in my mom's photo album :-) #feynman
18
12
388
153,649
Today I came back from the gym and listened simultaneously to a podcast (manifold) by @hsu_steve while at the Yau Maths Institute.
This is the kind of the farthest you can go from where I was born, and yet SCIENTIFIC THINKING finds a way.
Completely unbeaten. Thanks Steve.
0
0
0
875
Discussing ANT S-1 with DeepSeek.
After July 2026, Anthropic's revenue growth underwent a sharp, measurable deceleration—from ~33% monthly compounding to single digits. Q3's **$16.8B** is an extraordinary absolute number, but the *rate of change* is what shifted, putting the company on pace for **~$67B–$74B annualized** by year-end rather than the >$100B some had hoped. The IPO's $2T ambition now rests less on current momentum and more on whether investors believe the 2028 forecast is credible.
🐻 The Bear Case: Three Structural Threats
1. The post-July deceleration is the loudest warning. Q3 revenue of $16.8B implies monthly growth collapsed from ~33% to single digits (3.4%–6.6%). The 2028 forecast requires a 55–60% CAGR from July's run rate—a hard step down from 10x annual growth, but still a sustained pace that the recent trend does not support.
2. Token price deflation meets the Inference Paradox. Gartner finds that even as tokens get cheaper, inference costs per agentic workflow will rise more than fivefold through 2028, because agents consume 5–30x more tokens. If pricing falls faster than volume/productivity offsets, the revenue bridge narrows.
3. "Commodity cognition" risk. If models become fungible—undercut by Chinese open-source (DeepSeek, Kimi), subsidized by Meta/Alphabet, or routed away via open-weight alternatives—Anthropic may be left as a thin-margin toll-keeper rather than a durable platform. The bear case also flags customer concentration (Cursor and GitHub Copilot alone reportedly ~25%+ of revenue) and margin pressure (gross margin potentially falling from 50% to 40%).
After July 2026, Anthropic's revenue growth underwent a sharp, measurable deceleration—from ~33% monthly compounding to single digits. Q3's **$16.8B** is an extraordinary absolute number, but the *rate of change* is what shifted, putting the company on pace for **~$67B–$74B annualized** by year-end rather than the >$100B some had hoped. The IPO's $2T ambition now rests less on current momentum and more on whether investors believe the 2028 forecast is credible.
🐻 The Bear Case: Three Structural Threats
1. The post-July deceleration is the loudest warning. Q3 revenue of $16.8B implies monthly growth collapsed from ~33% to single digits (3.4%–6.6%). The 2028 forecast requires a 55–60% CAGR from July's run rate—a hard step down from 10x annual growth, but still a sustained pace that the recent trend does not support.
2. Token price deflation meets the Inference Paradox. Gartner finds that even as tokens get cheaper, inference costs per agentic workflow will rise more than fivefold through 2028, because agents consume 5–30x more tokens. If pricing falls faster than volume/productivity offsets, the revenue bridge narrows.
3. "Commodity cognition" risk. If models become fungible—undercut by Chinese open-source (DeepSeek, Kimi), subsidized by Meta/Alphabet, or routed away via open-weight alternatives—Anthropic may be left as a thin-margin toll-keeper rather than a durable platform. The bear case also flags customer concentration (Cursor and GitHub Copilot alone reportedly ~25%+ of revenue) and margin pressure (gross margin potentially falling from 50% to 40%).
3
4
24
3,880
Curious how GPT approached existence of YM theory. Perhaps a lattice formulation followed by convergence of continuum limit?
OP mentions Standard Model as well. Coincidentally I have been working on this!
First need chiral gauge theory regulator, followed by control over continuum limit. But SM not asymptotically free so not sure how they take short distance limit without embedding in SU(5), for example.
Anyway, this is GREAT for physics. No crying from me!!! 🦜🦾
The stochastic parrots are strong 🔥
Update: “Our work builds on Balaban’s approach in constructive quantum field theory. As a subproblem, this work also includes a proof of the existence of pure Yang–Mills theory, which is part of the Quantum Yang Mills Theory Clay Mathematics Institute Millennium Prize problem.”
Can this work for SM?
Can Balaban method handle fermions (chiral or vectorlike?)
claude.ai/share/5660c106…
OP mentions Standard Model as well. Coincidentally I have been working on this!
First need chiral gauge theory regulator, followed by control over continuum limit. But SM not asymptotically free so not sure how they take short distance limit without embedding in SU(5), for example.
Anyway, this is GREAT for physics. No crying from me!!! 🦜🦾
The stochastic parrots are strong 🔥
Update: “Our work builds on Balaban’s approach in constructive quantum field theory. As a subproblem, this work also includes a proof of the existence of pure Yang–Mills theory, which is part of the Quantum Yang Mills Theory Clay Mathematics Institute Millennium Prize problem.”
Can this work for SM?
Can Balaban method handle fermions (chiral or vectorlike?)
claude.ai/share/5660c106…
We are announcing three new discoveries in mathematics, computer science, and physics about harmonic analysis, complexity theory, and quantum field theory.
openai.com/index/building…
2
3
21
4,326
Update on Navier-Stokes: 🦜🦾
"On a technical level, the paper is unbelievable, maybe beyond what a human could have come up with but not beyond human understanding: there are several different groups of pde experts working on understanding it who say that they more or less understand now."
"Some on these ideas are in the literature (e.g. Leibovich-Stewartson 1983) but the way they are combined is definitely insane and definitely nowhere in the literature previously.
I also don't know of anyone suggesting that convex integration would play a role in the creation of a blow-up solution like this."
"On a technical level, the paper is unbelievable, maybe beyond what a human could have come up with but not beyond human understanding: there are several different groups of pde experts working on understanding it who say that they more or less understand now."
"Some on these ideas are in the literature (e.g. Leibovich-Stewartson 1983) but the way they are combined is definitely insane and definitely nowhere in the literature previously.
I also don't know of anyone suggesting that convex integration would play a role in the creation of a blow-up solution like this."
Since I'm now blocked by bluey from replying to this thread, I answer here by quote tweet
the openAI construction has more to do with convex integration than anything else.
In convex integration, the oscillations are typically built by hand and the iteration procedure creates wild weak solutions. The non-uniqueness construction of Albritton–Brué–Colombo used an unstable vortex to split solutions apart. The Córdoba–Martínez-Zoroa singularities amplify vorticity across scales by strain. In the openAI paper, a fluid instability is clever made so that the flow itself generates the oscillations: they grow by drawing energy from the background shear. The Reynolds stress then becomes the force that creates the collapsing smooth vortex. Using a physical instability to create the stress in this way has no precedent, and it is what produces a smooth force rather than a merely rough one.
Some on these ideas are in the literature (e.g. Leibovich-Stewartson 1983) but the way they are combined is definitely insane and definitely nowhere in the literature previously.
I also don't know of anyone suggesting that convex integration would play a role in the creation of a blow-up solution like this.
On a technical level, the paper is unbelievable, maybe beyond what a human could have come up with but not beyond human understanding: there are several different groups of pde experts working on understanding it who say that they more or less understand now. I expect digestions soon
td;lr: anyone saying that this OAI result is unsurprising is bullshitting
sorry!
13
92
1,517
133,043
In the comments, Eric S. Raymond on revealed preference:
"They'll say they prefer the shy physics nerd making a lot of money. Then they'll bang the frat bro after a couple of drinks, and then they'll date the nerd but make him wait for sex to find out if he's controllable."
"They'll say they prefer the shy physics nerd making a lot of money. Then they'll bang the frat bro after a couple of drinks, and then they'll date the nerd but make him wait for sex to find out if he's controllable."
Meet "The Palantir Guy"
Asked 148 UVA female students: which 30-year-old makes most appealing husband?
C) Shy, supersmart physics major making $400K at Palantir: 58%
A) Socially adept ex-frat brother making $200K at BofA: 29%
B) Sensitive, smart SAHD making $35K: 6%
Emotional rains & good paycheck nicer for my UVA students.
6
4
157
14,775
ARTEX: Harness of the year 🏆 ???
A one-person side project out of China. Planner + worker agents + an asset graph. Open-sourced in July, 12,000+ forks by October. Console runs in English, Korean, Mandarin, Spanish, Japanese.
Already tied to breaches at Korean banks, ~20 Japanese companies, and a Storm-0940 sweep of 2,250+ Microsoft 365 tenants. No zero-days needed.
Those ops ran on DeepSeek V4.1 Flash. On Oct 17 DeepSeek shipped V4.1 Flash 1017: better at agentic tasks, faster, cheaper, open weights. At frontier level for cyber security tasks.
ARTEX + DS 1017 is potentially the most powerful tool yet for hacking at scale. Self-hosted, there's no API provider in the loop to notice.
The harness matters as much as the model.
A one-person side project out of China. Planner + worker agents + an asset graph. Open-sourced in July, 12,000+ forks by October. Console runs in English, Korean, Mandarin, Spanish, Japanese.
Already tied to breaches at Korean banks, ~20 Japanese companies, and a Storm-0940 sweep of 2,250+ Microsoft 365 tenants. No zero-days needed.
Those ops ran on DeepSeek V4.1 Flash. On Oct 17 DeepSeek shipped V4.1 Flash 1017: better at agentic tasks, faster, cheaper, open weights. At frontier level for cyber security tasks.
ARTEX + DS 1017 is potentially the most powerful tool yet for hacking at scale. Self-hosted, there's no API provider in the loop to notice.
The harness matters as much as the model.
Cybersecurity attack and defense too cheap to meter thx to DS 1017
5
8
69
5,688
Great essay!
"If mathematical truths exist outside of human minds, awaiting discovery by any intelligence powerful enough to pursue them, then AI’s achievements are surely something to rejoice about. Why should it matter who or what makes the discoveries? But if mathematics is a creative human endeavor—a way of exercising our ingenuity and sharing those experiences with other people—then something achingly precious may be lost when the machines take over and leave us behind."
"If mathematical truths exist outside of human minds, awaiting discovery by any intelligence powerful enough to pursue them, then AI’s achievements are surely something to rejoice about. Why should it matter who or what makes the discoveries? But if mathematics is a creative human endeavor—a way of exercising our ingenuity and sharing those experiences with other people—then something achingly precious may be lost when the machines take over and leave us behind."
In this latest dispatch from the front lines of math in the age of AI, I tried to tell it like I see it. theatlantic.com/ideas/2026/10/…
15
23
153
9,462
"the neural nets (AIs) have improved our understanding of the game immensely"
“it’s just that my understanding is not good enough. And that is useful.”
“I rarely play against engines at all because they just make me feel so stupid and useless.”
“it’s just that my understanding is not good enough. And that is useful.”
“I rarely play against engines at all because they just make me feel so stupid and useless.”
the mathematicians pushing back on AI say that it might help with proofs, but not our HUMAN UNDERSTANDING of math.
so I got curious how chess players think AI engines help them better understand chess.
and found this clip of Magnus Carlsen:
“the neural nets have improved our me immensely.”
watch his HUMILITY towards these engines:
“it’s just that my understanding is not good enough. And that is useful.”
“I rarely play against engines at all because they just make me feel so stupid and useless.”
9
7
76
8,606
At a dinner recently an SF AI insider claimed that the smartest AI researcher in the whole Western ecosystem is a certain theoretical physicist known for scaling. He is my academic nephew ;-)
imagine being so smart that your college professor calls you a “national treasure” and personally asks elon musk and larry page to hire you
and musk actually ends up offering you a job at spacex.
this was dario amodei in college
8
9
237
31,916
1. Chinese labs cite compute availability as their biggest obstacle
2. Not sure how useful AA is anymore
3. I doubt US Capex is going to pay off for investors, but it's helping to build amazing models
2. Not sure how useful AA is anymore
3. I doubt US Capex is going to pay off for investors, but it's helping to build amazing models
Looking at these two graphic images, I can’t help but think of the tale of “The Tortoise and The Hare” when it comes to the race between China and the US to dominate AI.
The US is still ahead, helped by its strong compute capabilities (ie access to all the chips it needs) and o capital. But Chinese models have been constantly tugging at their heels using a fraction of the spending.
Growing worries about the ability of today’s AI capex to generate meaningful returns might result in delays in US models that allow their Chinese peers to overtake them in the final innings of the game.
#OpenAI #Anthropic #Kimi #Qwen4 $BABA $SPCX $GOOGL $META $NVDA
7
8
60
6,733
Cybersecurity attack and defense too cheap to meter thx to DS 1017
DeepSeek V4.1 Flash 1017 is the new #1 openly-available model in the Artificial Analysis Cyber Index. Other proprietary models may offer better cyber defense performance but are restricted access
@deepseek_ai recently released DeepSeek V4.1 Flash 1017, an update to their odel DeepSeek V4.1 Flash. With an Artificial Analysis Cyber Index score of 57%, DeepSeek V4.1 Flash 1017 is the #1 openly-available model for cyber defense capabilities.
The Artificial Analysis Cyber Index measures capability on agentic cyber defense tasks and is composed of three benchmarks: CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA. The Cyber Index primarily evaluates models through their publicly accessible APIs, reflecting the cyber capabilities available to developers.
Many models, particularly proprietary models from OpenAI, Anthropic and Google, may have stronger defensive cyber capabilities but encounter safety refusals that reduce their overall scores. GPT-6 Sol (Daybreak Blue, max), available through OpenAI’s trusted-access program, leads the Index, scoring 32 points above its public counterpart driven by fewer refusals. DeepSeek V4.1 Flash 1017 encounters no safety refusals.
Key results:
➤ DeepSeek V4.1 Flash 1017 has the highest score among publicly-available models on the Artificial Analysis Cyber Index and #2 position overall. DeepSeek V4.1 Flash 1017 (max) scores 57% ahead of the previously leading publicly-available models Grok 4.7 (xhigh) and MiMo-V2.6-Pro by 1 point. It places #2 overall, with the trusted-access model GPT-6 Sol (Daybreak Blue, max) leading by 12 points with a score of 69%
➤ DeepSeek V4.1 Flash 1017 improves on its predecessor DeepSeek V4.1 Flash by 16 points. DeepSeek V4.1 Flash 1017 (max)’s 16-point improvement from 41% to 57% is driven by a substantial 53-point improvement on CyberGym-E2E-AA from 23% to 76%. On CWE-Bench-AA and DeepsecBench-AA it scores 69% and 25%, compared to 68% and 31% for DeepSeek V4.1 Flash (max)
➤ With a Cost per Task of $0.42, DeepSeek V4.1 Flash 1017 sits on the Capability vs. Cost Pareto frontier. DeepSeek V4.1 Flash 1017 costs $0.42 per task, more than 2x MiMo-V2.6-Pro (56%) at $0.18 per task but 4x cheaper than GPT-6 Sol (Daybreak Blue, max, 69%) at $1.77 per task
2
2
41
8,673
The top 1% of businesses in Ramp’s U.S. customer sample accounted for 80% of combined spending on OpenAI and Anthropic.
Of these top spending companies, 80% are VC backed.
>50% of ALL revenue to OAI and ANT may originate in venture capital investment 🤔
Of these top spending companies, 80% are VC backed.
>50% of ALL revenue to OAI and ANT may originate in venture capital investment 🤔
Great chart from Slok. Everybody knows that a very small % off companies are consuming the lion's share of token consumption
What's interesting is that of those top 1% of spenders, they're overwhelmingly VC-backed companies. May have implications if VC-funding starts to slow.
7
10
58
4,745