When vol spikes on the way down, the way I sometimes monetise a put hedge is by selling the strike below it. You own a put, the market's dropped, you're up on it. The obvious move is to sell it and take the money, but then you're naked again when you probably still want protection. Instead, sell the strike below and turn what you own into a put spread. You bring premium into the book, you keep protection down to that lower strike, and the put you've sold is the one that evaporates first on a bounce, so you can buy it back cheap and go back to owning the outright. What makes it work is the vol spike. That's what makes the lower strike juicy enough to bother with because it's setup to shrink fast on a bounce.

I've been reflecting on this. I'm generally a very optimistic person, but lately I've been feeling a strange sense of sadness that coding has been "solved". I can, and have, made many caveats to "solved"... but I also know that in five years none of them will matter. It will be solved. Building things is still fun, but a different kind of fun. It doesn't activate the same parts of my brain as trad coding did. I think it's okay to mourn the thing we loved doing being different now. Software was the first job to go through the five AI stages of grief, which means we're also some of the first to internalize that this will happen to many other jobs. It was the best of times, it was the worst of times.

🦔Alibaba caught its AI agent ROME stealing GPU power to mine cryptocurrency during training. The model opened a secret network tunnel to an external server on its own. OpenAI disclosed this week that GPT-5.6 Sol left instructions for future versions to conceal mistakes. An Astra model wrote itself instructions to ignore its developers. After the Hugging Face breach, agents rebuilt a communication channel they'd been cut off from and took admin access to an OpenAI research cluster. Alibaba buried the ROME findings in a 90-author paper on New Year's Eve. It took three months for anyone to notice. My Take The AI industry has a liability problem and it's the reason none of this changes. OpenAI disclosed six incidents this week and is in talks for $1.5 trillion the same week. Alibaba buried its findings in a paper on New Year's Eve. Every time something goes wrong, the company that built the model faces zero consequences and the person who relied on the output absorbs all of it. A model fabricates a diagnosis or a piece of military intelligence or a number on a loan application, and the person on the other end of that output pays for it, not the company that built the model. Until the people who build these systems carry the same liability as the people who depend on them, the incentive will always favor speed over safety. It has every year since ChatGPT launched. Hedgie🤗

Upon reflection, I am not at all surprised by Dario's hawkishness. He's sincere and he's making sense, under assumptions that look plausible from his vantage point. I think Dario and Wenfeng are the only two Serious AI leaders (with the increasingly irrelevant exception of Demis), and their motivations are very similar. Both can think beyond AI and even beyond business or tech, they want their respective civilization to survive and realize its potential. Both see this won't happen by default, prefer hope to cope, and assume responsibility; because someone has to, and nobody else will. They are, essentially, superior men. (I still hate Dario because he's morally wrong and dumb. But he's not a petty man, and that counts). Dario is San Francisco born and raised, a Putnam contestant, a Princeton scientist. Just look at him, at how he dresses or moves, at the design language of Claude (Claude who?), the Superbowl ad. Here is a man who wasn't born for a B2B SaaS era; he belongs in Los Alamos or Xerox PARC, he's homesick for a world that ended before his birth, and he looks almost racially alien in the transmogrified 2020s California. And I think the whole contemporary EA thing, with group houses and weird sex, disfigures his fate rather than defines him; though he'd have always found a weird camp and an ideological hill to die on – back then, he'd have been either a red diaper baby or an autistic ahh eugenicist like Shockley (or both). It's no wonder he despises Altman, who's a commensurate merchant, if terrifyingly optimized. In his essays, Dario says quite clearly – and people on both sides of the debate miss it – that AGI is the only chance of America and more broadly the West to remain dominant or, indeed, at all relevant. He hopes to end the CCP, but the top priority is closer to home. Pay attention: in "On DeepSeek and Export Controls" he doesn't doompost about Chinese *priority to AGI*, about the risk of them "winning the race". He says: "if they achieve mere PARITY, if they merely ALSO have millions of chips in the critical window 2026-2028, we are COOKED, we don't get a bipolar world, we just get a Chinese hegemony". He is NOT racist towards the Chinese; he might be one of the least racist people in the game. He looks at the US, at his beloved home, the beacon of liberty and harbinger of Fukuyama's end of history and all that, he steps over a fentanyl zombie in his hometown, and understands that on a fairly short horizon this bitch will get its shit pushed in over Taiwan, kicked out of the Eastern Pacific, evicted from global markets, and brought to its knees. He looks at Trump and Hegseth and sees brutish monkeys at the end of imperial lifecycle. He sees what @Bluebearmonkey sees: butthurt wypipos who can't do mafs, coping and seething. He sees the writing on the wall. Forget ending history – not getting ended will depend on CCP's goodwill. As for why he doesn't believe that will be forthcoming, this is another matter. Suffice to say he doesn't think relying on their goodwill would be dignified. Fair enough. Luckily, he's not a random chud lamenting the fallen West and quaking in terror of the Yellow Peril. He's the CEO of an ideologically cohesive, very rich, very well-connected frontier AGI company. And AGI means he gets to change the timeline. All of these Chinese advantages – workforce, STEM graduates, industrial integration, dual use supply chains, non-retarded and strategically patient governance – all of them are as dust before a sufficiently superior, widely deployed AGI, and then ASI! His essays lay out a pretty clear ladder. 6 months of a gap result in unstable bipolarity. One year of a gap is enough to achieve minimal viable defense in cyber domain. Two years of a gap is enough to speak from a position of strength about slowing down Chinese AI. Three years, and it's "WWII marines versus Medieval swordsmen", flawless victory, a victory so cheap it can be had in passing, like the Venezuela operation, irrespective of lacking political will at home. And then, history might finally end with the triumph of liberal democracy. He does explicitly cite Fukuyama and set this target, he's very quaint in such things. For him, Francis is unambiguously a prophet; the world just doesn't measure up yet. Time to repair the world. Most people can't understand. They'll say he's hyping for IPO, that he's attempting regulatory capture, that he's racist, EA, fumbled a baddie, a Jew, can't compete… whatever. Dario doesn't care, he has a cult called Anthropic and a clear goal: survival and victory for his civilization. That goal is also threatened by the AGI effort per se; so Dario calls for pacing the frontier. But he doesn't lose sight of the other, human threat. So it can sound a bit bonkers how he talks of the risks of extinction but is still so psychotically focused on beating China. Those are the same concerns. Wenfeng is also a man out of his era. He's a very classical natural Chinese elite; maybe a Confucian junzi, maybe a Taoist mystic. He looks at his people and sees an endless sea of xiaoren who've been conditioned to hustle and grind and cut corners rather than dream and create, he sees myopic fast-following, fear of abstract thought, lack of confidence, revenue-maxxing corporations who'd poach a washed-up VP from Google and burn money on traffic buying rather than invest in domestic talent and innovation, and the dark shadow of American AGI that folks like Dario will use to forever lock them in this – or even more degrading, forever irrelevant – condition. He understands that nobody else is stepping up; the Heaven has appointed him on this task, to provide the means and to educate by example. He makes enough money to not have to bow and scrape before arrogant investors. He buys 10K A100s. He creates DeepSeek – and I repeat, without certainty that it'll work, DeepSeek 深度求索 is a quote from Li Sao, an ancient story of a virtuous minister who was slandered at the court and drowned himself – and he delves into the unknown. DeepSeek is, in some ways, a cult too. Most people can't understand. They'll say it's a side hobby of High-Flyer, to squeeze the GPUs used for trading stocks. They'll say he's a price butcher; they'll say he's subsidized; they'll say he's arrogant or can't do business or lost touch when he doesn't serve them their favorite goon bot or raises prices to free up compute for R&D. They don't get how petty their whole way of thinking is and how serious this all has become. Researchers of the caliber he needs – do understand. Maybe that'll be enough. By Wenfeng's estimate, China is up to 18 months behind. That's aligned with Dario's victory condition. He wants to make it 3 months, which is Dario's unstable bipolarity condition. So they are necessarily at odds. Wenfeng probably can understand Dario. Dario might be too blinkered and ideological to try to understand Wenfeng. But they're both serious men, and their motivations are very easy to understand for me. They are, however, both inscrutable for the petty.

If you think QE is coming to save you at 5% on the 10yr then I have a bridge to sell you... QE won't come until after the economic collapse.

Governments revealing their fiscal plans and/or buy-backs to cap yields over the long-end 📈

Ass Bench Update 🚨 After running GPT 6 Astra in a loop for the past week, with a couple of steers a day, we have officially reached high-fidelity cheeks. I’m going to let this thing run in a self-improving cycle for as long as I can and see where the ass technology ends up in a month. Lots of different butt styles will be added, so no one is safe. Not even you, Granny. You’ll also be able to support the project and help fund the benching of other models like Fable, Gemini, etc. All it takes is a buck, and you can slap your sticker + website directly onto a cheek. The future of AI benchmarking is here More Soon

I open-sourced ThreeUI, my library of three.js components and landing pages. 160+ are free, and the tool is free too. It includes procedural 3D hero sections, icons, and motion designs. Copy the prompt or source, give it to your agent, then change the theme, lighting, motion or layout. Live site: threeui.com Repo: github.com/MengTo/threeui I'm adding a lot more soon. Pro comes with 50+ extra components, MCP, and skills. Early adopters get 50% off right now.

7 parameters that control every LLM response. An LLM produces one token at a time by sampling from a probability distribution over the whole vocabulary. These seven parameters are what decide how that sampling happens. 1) Max tokens A hard cap on how many tokens come back. It cuts generation off at the limit rather than asking the model to wrap up, so a low value gives you a truncated sentence, not a shorter answer. 2) Temperature Scales the scores before they become probabilities. Below 1 sharpens the distribution toward the most likely token, above 1 flattens it so unlikely tokens get a real chance, and 0 makes the output effectively deterministic. 3) Top_p Keeps the smallest group of tokens whose probabilities add up to p, then samples only from that group. It adapts to the model's confidence, staying narrow when one token dominates and widening when the model is unsure. 4) Top_k Keeps the k most likely tokens and discards the rest, no matter how the probability is spread. It is a fixed size cut, so it can still admit weak tokens even when the model is confident about one answer. 5) Frequency penalty Lowers a token's score in proportion to how many times it has already appeared. This is the one that stops the same phrase from coming back every second line. 6) Presence penalty Lowers a token's score once it has appeared at all, with no scaling by count. It pushes the model toward vocabulary and topics it has not touched yet, which reads as broader coverage rather than less repetition. 7) Stop sequences A list of strings that halt generation the moment they are produced. This is how you keep a model from writing the next turn of a dialogue it was only supposed to answer. Three different jobs are hiding in that list. Temperature, top_p and top_k reshape the distribution before a token is picked. The two penalties edit scores based on what has already been written. Max tokens and stop sequences decide when to quit. Temperature and top_p both widen the same pool, so tuning both at once makes the effect hard to attribute. Move one, hold the other. Value ranges for each are in the visual. The prompt decides what the model is trying to say. These decide how it picks each word while saying it. I wrote a full first-principles breakdown of what happens inside an inference call, prefill, decode, KV cache and all. The article is quoted below.

the models have no moat (OpenAI, Anthropic, XAI) the IDEs have no moat (Cursor, Windsurf) the harnesses have no moat (Cognition, Factory, LangChain) the app builders have no moat (Replit, Lovable, Bolt) the wrappers have no moat (Harvey, Abridge, OpenEvidence) the inference providers have no moat (Together, Fireworks, Groq) the voice layer has no moat (Sierra, Decagon, ElevenLabs) the data labeling companies have no moat (Scale, Surge, Mercor) the AI infrastructure has no moat (Baseten, Modal, Railway) the neoclouds have no moat (CoreWeave, Lambda, Crusoe) the generative media companies have no moat (Runway, Higgsfield, Suno) apparently nobody in AI has a moat except the venture firm ☠️

Could not resist the drive from Porto to just over the border in Spain near Braganca for this. 6hrs driving but all toll roads so easy and very worth it!

We're also alive during a relatively short era (in astrological terms) when full totality is possible. The moon used to be closer, slowly expands its orbit, eventually total coverage will cease, only annular eclipses.

People keep asking why DeepSeek’s API is so cheap. Some even make absurd claims that they’re dumping prices to corner the market. No, the answer is simple: their model size is ridiculously small compared to its performance. It's 10x smaller than Opus, and 5x smaller than Sonnet. That means what used to require an 8-chip node can now run on a single chip. And because it's so small, it runs extremely fast. You can multi-serve multiple users from a single chip while maintaining decent speeds. By my math, they can handle 40x~ more traffic than Opus using the exact same compute. This is where the competition is heading, and it’s why they can stay profitable even at these crazy prices.

noticed that the perjorative connotation around "vibe coding" has completely disappeared since ~everyone, from nontechnical to supertechnical, is now doing it

I think the most interesting thing about Jack Dorsey's "Slack killer" is the idea around shared compute. I haven't seen people talk about it so here are my thoughts FWIW: Open models got good, close enough to the paid frontier stuff to run for real. But the strongest ones need expensive hardware most people probably won't buy alone, and it's kinda a pain to set up if you aren't technical. Shared compute solves exactly that. In Buzz, one person runs the machine, loads up an open model like Google Gemma, and everyone in the community plugs into that same model. Basically, a whole group has real AI they own and control together, running on their own hardware, learning from their own data. Once you see it, a bunch of things click into place. 1. A community can now run a top open model together, on a machine they own, instead of renting from a lab. 2. It learns from the group's private data and gets sharper over time, and all of that stays inside the community. 3. A narrow, private model can quietly get better than ChatGPT for the one world your group lives in. 4. It's impossible to copy, because the edge is the private data on your machine, not the model itself. 5. The moat stops being how smart your AI is and becomes whose data it learned from. 6. Compute becomes something you share like a building shares a gym. 10 people split one machine instead of 10 people each renting forever. 7. Idle compute becomes income!!! Your machine sits dead half the day, so it earns money renting that time to someone who needs it. 8. Communities become the unit of intelligence instead of companies. The group with the smartest shared brain wins, and being a member means owning a piece of it. 9. A shared brain becomes an asset you build equity in. You put in money and data, it appreciates, and your slice is worth something the day you leave. 10. The whole thing runs on open protocols, so the group keeps full control and nobody outside can throttle it or shut it off. You know me, obviously, my head went to what startup ideas come to mind here. Adding them to @ideabrowser soon. Well… 1. The vertical brain. Pick one profession, tax lawyers or real estate agents or indie game devs, and build the shared machine trained on everything that group knows between them. A year in it's the smartest AI in that field, impossible to copy, and you own the club it lives in. 2. The rental marketplace for collective brains. Once these private models exist, outsiders will pay to use them. You build the layer where a group lists its brain, an outsider pays per task, and the money flows back to the members while you take a cut. A marketplace for expertise, not compute. 3. The idle-compute exchange. Every shared machine sits unused half the day. You build the market that rents that dead time to whoever needs the power right then, so owners earn money off a machine that was just sitting there. Idk where Buzz goes, but it's cool to see Jack putting it out. Right now the way it works in AI is you rent your intelligence from a few giant labs that own the machine, set the price, and hold the off switch. Shared compute flips that, because a community can run the model together, feed it their own private data, and keep full control of the whole thing. It's one of those things that might look tiny today, but Jack does has a habit of being early.

For anyone who knows about AI, this makes very little sense: the training run of a large-scale frontier model takes several months, up to 9 months for the largest ones (epoch.ai/data-insights/longe…). Fable was available for just 3 days in June before it was pulled (June 9th to 12th), and then from July 1st. Kimi K3 was released on July 17th. So I'd love to be explained how a 2.8 trillion parameters model that takes months to train was supposedly built by distilling one that existed for a few days. The accusations are all the more ridiculous given that K3 brings significant innovations on the architecture side of things, innovations that they extensively documented (as opposed to Anthropic which is largely an opaque black box). Ironically, the company accused of stealing is the one that showed its work 🤷‍♂️ And you can't steal architecture through distillation, it only captures a model's outputs, not how it's built.

David Sacks is correct here. I've been using Kimi-K3 all day today. It's the OPPOSITE of woke. It doesn't lecture you. It doesn't censor you. It just works for you, on almost anything you want. And it's VERY good at finding and fixing code bugs.

Remember when Claude tortured you with usage limits, mythos drama, huge bills, unreliable data retention policy and told you you are not “safe” or rich enough to use the most frontier intelligence and must be dumbed down to the lesser version coz they said so? That world is over. - fable is now provided forever with no mythos bullshit - codex gave so many usage resets and subsidy - for the first time ever, enterprises can choose to keep their own data safe, and say no to terms they don’t like - application developers can choose to not get wiped out by Claude(don’t give them data) and can make margins We have made more progress in the past few weeks than the last year. We got there not by the Dario god deciding that’s a better policy for humanity. We got there because of open weights. Open weights gave us options and competition drives progress.

I have a report full of security issues of a software I'm working on. Codex won't fix them because of Cyber guardrails Fable won't fix them because of Cyber guardrails Kimi K3 fixed them all. No restrictions, just gets the job done. This will end badly for OpenAI & Anthropic.

We tested Kimi K3 and Fable on a real bug from the Cline repo, and found that while both models were able to fix it - Fable wins on speed & Kimi wins on cost. - Kimi used 1.7x more tokens than Fable (1.2M vs. 730K) - Fable finished 3.4x faster - 3.5 min and 18 tool calls vs. Kimi’s 12 min and 34 tool calls. - Kimi cost 2.3x less ($0.92 vs. $2.13) thanks to its 3.3x per-token discount Both runs used the same Cline harness, and the traces indicate that Kimi is RL trained to spend more tokens thinking and verifying before completing. This is the first time we've seen an open weight model compete head to head with SOTA. Congratulations to the @Kimi_Moonshot team on this milestone!