Leopold retweeted
Thoughts on memory:
Anyone who thinks compute will stay in shortage while memory slides into oversupply must be smoking something. Compute and memory are destined to tighten together or deflate together. One cannot die while the other survives.
By the same logic, the pods shorting memory and going long HDD are out of their minds. How do you think HDD isn't joined at the hip with memory? You really think HDD stays safe if this cycle rolls over?
In the end, all of this comes down to nobody having conviction about what happens after 2027. Historically, we have watched P peak one to two quarters before supply ramps, then Q expands and the entire profit pool collapses. But everyone forgets that this cycle runs on nonstandard products, unlike PCs, smartphones, and general purpose servers, and that it was triggered by factors different from past cycles.
This is because of the boomers who still see memory as an industry where you make a killing in the upcycle and lose your shirt in the downcycle. These self proclaimed veterans view memory through the lens of their past experience, and since they drive the market and set the stock prices, memory trades at such a low valuation. What I keep saying is that a sharp collapse like in the past is impossible. A gradual decline is possible. But even if Samsung and SK Hynix margins were cut in half from current levels, they would still trade at a P/E in the teens. That is still roughly half of HDD valuations.
Personally, I think that as memory prices fall, other sources of demand will absorb the supply at each step down. In particular, I expect personal AI demand, led by local AI, to emerge. Looked at another way, for local AI to take off, the hyperscaler driven spike in memory prices has to be brought under control. At current memory prices, local AI can never diversify.
Just as mainframes gave way to PCs and the cloud gave way to smartphones, each time carrying the cycle from P into a Q expansion phase, I think this AI cycle will also move from frontier labs to local open source AI. Based on these historical precedents, I believe local AI will take over the memory cycle from the frontier labs and carry memory demand.
I am not saying memory prices will hold for more than a decade. Memory P will fall at some point. But what is different from the past is that new consumer devices and new consumer demand, such as open source driven local AI, will support the floor for memory.
So the market right now keeps discounting these stocks out of fear of a memory peak nobody can time. This discount is being driven by boomer investors who dismiss the idea that AI supports the downside and keep pushing the view that earnings will collapse and the trough will revert to past levels too. My point is that this is completely wrong.
Leopold retweeted
100% agree. Like many others, @t1alpha and @t1acapital are making rapid progress in moving local and only using cloud/frontier for hardest 5%. The local compute token supply is exploding.
From my perspective, it’s broadband 2.0, but possibly much worse, and will require pretty strong evidence otherwise.
I get it, there are nuances for tech-bros who are deep in the weeds. Everyone has an anecdote about why I’m wrong.
Take a step back and be a commodity-bro who’s seen enough supply/demand cycles. Tokens are a commodity. Currently, the token factories are out-stripping token demand. Pricing is the adjustment factor. Except token factories are mostly fixed-price businesses. They have take-or-pay. They have commitments. They will produce tokens at a loss rather than turn off.
I’m not saying that token demand is slowing. It’s still growing at astronomical rates. I’m saying that token production is now out-pacing demand. The cycle will turn when demand accelerates or token production slows. There’s like half a trillion $ of GPUs waiting to be installed. I don’t think token production slows anytime soon.
Yeah, I get it. Jevon’s and all. It’s great for AI consumers and innovators. But the guys owning token factories are gonna all go bust in the process, and that’s like $2T of capex to date. You cannot even restructure these things and crush the debt. They’re just stranded by new, and more efficient chips. They’re always going to be stranded by new chips. More efficient Models will strand the more expensive ones. It’s a race to the bottom against fixed costs. It’s the most godawful business misallocation I’ve seen in a while. Makes the Chinese look like prudent capital allocators🤣🤣
It’s like having iron rails when the new railroads run on steel. This bankruptcy cycle will be epic…
Leopold retweeted
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
Leopold retweeted
The memory trade became so consensus even the Holy Father knows the base case.
Intelligence must draw on memory. The mind must not simply compile data — as an algorithm now does more quickly than we can — but recall lived experiences, which contain depths of meaning that only the human soul can recognize, and connect more deeply with the goodness of these experiences. Everything that is good begins “with the Spirit” and thus possesses a breath, a dynamism and a sense of communion that, once experienced, are engraved upon the heart. vatican.va/content/leo-xiv/e…
Leopold retweeted
The hashtags I use are in reference to pieces I’ve written or published and they serve as bookmarks for me for when I want to revisit those pieces
grant-williams.com/cry-wolf-…
Leopold retweeted
The Kospi’s like your college friend who used to blackout every weekend but now has like kids and a mortgage
Buddy would’ve been up at least 9% back in the day
Leopold retweeted
Amazon cuts off Muse.
While I am bullish Meta and Muse, I think many people are overlooking the digital knife fight that’s about to occur
Nobody wants to get commoditized or layered here. Let the games begin
Leopold retweeted
Important: Houthis Used @claudeai @AnthropicAI for Missile Programming
Excerpt from the Detecting and countering misuse of AI: September 2026
GTG-87001: Disrupting a Yemen-based guided weapons engineering cell using Claude to develop guidance software
Summary
We identified a cell of threat actors based in northern Yemen running three weapons development programs: a guided rocket that used a commodity phone-class flight computer with final-phase homing guidance; a multi-stage ballistic missile with a stated range goal above 2,000 km; and a multi-variant missile (referred to as the “R2000” set) that included a hypersonic glide vehicle variant.
The actors used Claude Code in place of human software engineers to develop the guidance, navigation, and control (GNC) software that steers and stabilizes a flying vehicle. For example, they used Claude to integrate an open-source autopilot onto a phone-class flight computer, writing the control and position estimation software, tuning the control settings, running a firmware build pipeline, and performing a flight simulation. The actors managed several Claude instances at once, assigning each one a role, much as a lead would delegate work on a small engineering team: the actors tasked one instance with writing the code, another with research, and a third with reviewing the code the first instance produced.
Our safeguards blocked many of their requests, but not all of them. The actors used a variety of tactics to evade our safeguards, including hiding their goals and the products the software was meant for, and they split their work across multiple sessions so no single session revealed their full intent.
These actors carried out a sustained effort to develop guided weapons, including using Claude to design guidance software. We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket. This field test appears to have failed: within hours, the actors returned to Claude to work out why it failed.
We identified this activity as part of our internal investigations into suspected weapons development. We banned accounts associated with the actors and shared threat information with public- and private-sector partners to mitigate risks posed by the actors. Nevertheless, we have evidence that the actors had already built an offline simulation toolkit that does not rely on Claude or other engineering computing environments such as MATLAB.
anthropic.com/threat-intelli…
Leopold retweeted
MAHIYA NAMAN KAYO | Secretary of National Defense Gilberto C. Teodoro, Jr. responded to a note from the Chinese side during the Seoul Defense Dialogue (SDD) 2026 in Seoul, Republic of Korea.
#DNDPHL #DiKaPasisiil #SDD2026
Leopold retweeted
At the core of AI, you really just need:
1) a thing that does
2) connections into tools
3) a place to hold stuff
4) a spot to share and update
In slot 1, we're talking AI agents like Codex or Claude Code or GrokBot. Whether you have a daily driver or a mix and match, doesn't matter. Just an LLM-based system that can spin up hundreds or thousands of agents across multiple tasks at once.
In slot 2, we're talking toggling them toggles and connecting Codex to places like Gmail, Stripe, Magnific, Remotion, web browser, computer use, etc. Your risk tolerance will determine how many toggles you toggle. I now feel comfortable with both Codex and Claude Code having access to my GDrive, Fireflies, etc. and did not feel comfortable when the tool connections first came out.
In slot 3, when AI creates a document or a dashboard, where is it? This could be Google Drive, Dropbox, GitHub, your local storage, Obisidian. But it needs to be a place for your agents to build up their artifacts, task managers, and brain. Their memory. The more multiplayer your system is, the more IAM and cloud matter (aka local storage won't cut it).
In slot 4, you need a hub. When AI completes tasks for you or wants to ask you something, where is it doing that? I use Slack as my team's hub and we have enabled Claude Tag and built our own Claude Code apps as well.
Am I missing anything? Let me know.
Leopold retweeted
The promise of AI is currently framed as relieving humans of the tedious tasks that encompass work, allowing them to focus on the more rewarding and creative problems of novel discovery and idea generation.
This has been mostly true, but as AI gets better I think it has a high likelihood of becoming a comforting lie that we tell ourselves. The more capable AI becomes, the more humanity will have to resist being relegated to the middle management layer of its own progress. And the larger the risk that all human work becomes machine-like without any real sense of fulfillment or accomplishment.
As AI gets better at the stuff that goes beyond execution and into the realm of judgement, creativity and purpose, the boundary between AI work and Human Work is going to destabilize. Without a conscious effort to push back against this, it’s easy to see a future where humans are relegated to supervising systems they no longer fully understand.
“Drudgery” is defined by repitive, tedious tasks. But perhaps a better definition for it would be work with the absence of agency, and under that definition AI can certainly create more of it. Plenty of coders have already experienced a loss of meaning in their work as AI “solves” coding.
“What is joy without despair” is accepted to be a fair question. But what is self-actualization without the overcoming of friction or frustration. I do worry a bit for the future generations that may live an entire life without experiencing the qualia of success after repeated full-effort attempts met with failure.
If AI can execute our ideas, it’s likely that eventually it will be able to create better ones. Mastery and ownership of a creation being things of the past is hard to imagine.
The most valid pushback to this concern is that people still play chess despite engines existing, make furniture from hand regardless of IKEA etc.
I wonder if we’ll see an investment theme over the next decade that has to do with the “meaning economy”, or whatever people buy when their work is no longer their identity.
Maybe in the more near term, there’s a trade in the value of accountability. If the vast majority of work isn’t done by a human, who do you yell at / fire when the work is bad? Anyone selling the right to blame a human or to outsource accountability (maybe E&O insurers?) could benefit.
Leopold retweeted
When it comes to memory and HBM-specific, I think it's quite straightforward. If demand for AI continues to stay on a steep curve like it has been for some years now, there are only a few things that can realistically stop this memory cycle, as for the next 2 years there is not enough memory supply coming online:
1. A significant software change to the way AI models are inferenced. And I don't believe Kimi K3's hybrid linear attention mechanism is that, as it is still a hybrid with 3:1. (Also, there is the big question of hallucination rates)
2. Scaling laws hitting a ceiling for AI model development.
On the other hand, there is also a not-so-small chance that, even if memory supply improves significantly over the next 3-5 years, we might hit a new memory bull cycle adoption, with a new set of consumer AI devices that will again need memory (smartglasses, pins, necklaces, other devices) and/or humanoid AI.
Memory companies are doing LTAs because, at some point, even they want to limit their gross margin to not kill the value generation at layers down the AI stack (right now every layer is making a lot of profits), and they want to be a less cyclical industry, which arguably I think, because of AI, they actually are. This is the first time memory is being used to produce direct intelligence, which resembles more of a commodity than a fixed tech gadget with limited demand (people don't want/need to have 3 phones or laptops). One should expect that if the price of intelligence falls, demand for intelligence increases.
I do not believe memory companies will go back to gross margins that were seen in prior down cycles, as every tech company in the AI value chain is also looking to have strong relationships with these companies, especially since all know that we are still early in the AI cycle, and consumer AI and humanoids haven't even started, where memory is again going to be a key component.
Leopold retweeted
Twenty-seven years of the euro area's 3% deficit rule, and who kept it.
France 20 breaches. Greece 17. Portugal 16. Spain 16. Italy 13. Germany 10.
Croatia, the newest member, has zero.
Monetary union was sold as a hard money project. This is the enforcement record.
