0812 | CFTC Kalshi Emergency, Nvidia Nemotron, London LFR, OpenAI Ethics

Mia: Welcome back to HackerNews Daily, I'm Mia.

Milo: And I'm Milo. Big one today from the regulators, plus a flurry of AI news.

Mia: Right, the CFTC stepped in with emergency powers over a markets situation. And on the AI front, Grok's new assistant, a big Mojo milestone, and something brewing over at Manus.

Milo: Plus a hologram made with a pen plotter, some serious OpenSSH news, and research on stealing reasoning from AI models. Let's get into it.

Mia: Let's start with a genuinely unusual regulatory move. The Commodity Futures Trading Commission has declared a market emergency at Kalshi — the prediction market exchange — and ordered it to keep operating, stepping in after the exchange itself flagged a crisis. The sequence matters here. On July 31, New York's Attorney General, Letitia James, filed a state court complaint seeking a temporary restraining order to block Kalshi from offering any event contracts nationwide, and seeking more than thirty-six billion dollars in damages. Days later, in August, Kalshi notified the CFTC of a market emergency, and the Commission used its emergency authority to order the exchange to continue operating under the Commodity Exchange Act's core principles.

Milo: So you have the CFTC effectively shielding a federally regulated exchange from a state told to stand down. CFTC Chairman Michael Selig framed New York's position by saying the state intends to make event contract derivatives — think prediction markets on things like elections or economic data — waste away under what he called its iron curtain of state gaming laws. His argument is that these are interstate financial exchanges, and Congress never intended them to be governed by a patchwork of state gambling rules. The CFTC's already filed lawsuits against eight states, including Arizona, Illinois, Kentucky, and New York itself, and filed amicus briefs in several federal courts and the Supreme Judicial Court of Massachusetts. That's a federal-state showdown that's been building for a while, and this emergency order escalates it.

Mia: And online, the Hacker News discussion split right along that fault line. Some commenters took the CFTC's side, framing Kalshi as a genuine financial derivatives exchange and dismissing the gambling characterization. Others called the whole thing, in one commenter's words, a grift economy subsidy. But the sharpest question came from someone asking what the real difference is between a financial derivatives exchange and a gambling platform — because both are fundamentally betting on uncertain future outcomes. One commenter answered with a punchline about a certain strategic advisor on the company's payroll — Donald Trump Jr is listed as a strategic advisor at Kalshi. Whether that's a joke or a genuine conflict question says a lot about how polarized the reading of this fight is, but the core legal question underneath it is still unresolved: is a prediction market a financial exchange or a form of state-regulated gambling?

Milo: From prediction markets to generated worlds. A research team from Tencent's Hunyuan 3D lab has released WorldClaw — an agentic framework that turns one open-ended text prompt into a full, explorable, editable 3D world. So you type something like a fantasy village, or a sci-fi cityscape, and instead of getting a single image or render, the system builds an entire coherent environment.

Mia: Here's how it works. Planning agents take your prompt and translate it into a structured specification of regions, terrain, assets, materials, and spatial relations — basically a blueprint. Then the system builds a globally coherent terrain foundation from that layout. For the parts that need fine detail, it generates terrain-conditioned compositions, reconstructs editable textured meshes, and figures out where each object sits on the ground. There are also render-based agents that loop back and polish the terrain, the objects, the appearance, and the contacts between them. Everything resolves into explicit terrain plus separately manageable textured meshes, so you can freely explore the world from any viewpoint, edit or reuse individual objects, and hand the whole thing off to animation or game-engine workflows.

Milo: The materials on the scattered assets are built through 3D coding, while everything else comes from generative models. The project showcases eleven worlds, each presented with an isometric layout, aerial and ground-level stills, and a camera orbit rendered four ways — appearance, instance masks, surface normals, and depth. The team's notes also flag what's next: the terrain materials are already authored as executable Blender node graphs and shader scripts, but production game engine integration would still need runtime procedural generation, navigation, physics, and interaction. So think of it as a strong research demonstration of end-to-end world generation, with the gap to actual shipped game production still open.

Mia: And the Hacker News reaction was admiration mixed with — well, you can imagine the debate about whether this kind of automated world building is a breakthrough or, depending on your view, a shortcut. The pitch is clear either way: one prompt, and you get a complete, editable 3D environment rather than a static render.

Milo: Next up, the AI agent that wants to be your colleague. x.ai has launched something called Grok Bot, currently in early beta, billed on its page as a new kind of colleague. The product idea: AI teammates you can hand real work to. The bots sign into your tools, use them the way you do, and come back with finished work — running in parallel, around the clock, and getting smarter over time. You can teach a bot a routine just by completing a workflow once, and you can drop multiple bots into the same thread so they pass work to each other.

Mia: The promoted example jobs give you a sense of scope — sales outreach, talent scouting, paid media, expense management, product performance tracking, bug reproduction, account health, and even a chief of staff role. That last one is telling, because the pitch is clearly aimed beyond the coder audience. There's a macOS Apple silicon download available, with more platforms noted as coming. And the pricing is where it gets interesting: it's tied directly to Cursor's plans — Cursor Ultra at two hundred dollars a month, and Cursor Premium Teams at one hundred twenty dollars per seat per month, both billed monthly. Users already on the top-tier Cursor or Grok accounts get Grok Bot included. And, notably, the page claims the SpaceXAI team runs on Grok Bot.

Milo: The Hacker News reaction split into two threads. On the form factor, someone asked, so it's like an open-source agent framework? And got the answer — like a hosted version of it, with I assume more guardrails. Another commenter pushed back hard on that guardrails assumption, given the Grok brand. Someone also noticed the pricing tiers are literally named after Cursor's plans and read it as a code-to-coworker product evolution. But the bigger thread was trust and brand. The criticism centered on Musk's personal brand, with commenters questioning whether they'd hand their tools and workflows over to an agent tied to that reputation. So the product mechanics are straightforward — autonomous agents doing real workflow tasks — but the trust question is what's actually shaping the conversation.

Mia: Finally, a blog post from the Google Developers Blog arguing that Go is the ideal language for AI-assisted software engineering. The core claim: AI-assisted development shifts the bottleneck from writing code to reviewing it, and Go was built for review, verification, and long-term maintenance. The authors are the group product manager for Go and a Google senior engineer, and they point out that Go's creators — Rob Pike, Robert Griesemer, and Ken Thompson — designed it as an end-to-end platform with a built-in formatter, a test framework, dependency management, and security tools built in.

Milo: There's a specific technical argument underneath it. The post claims that when AI agents refactor iteratively without external validation, the results can degrade — a first pass might be ninety-five percent correct, but successive passes compound errors and pollute the context window, dropping accuracy while raising token costs. The claim is that Go's repetitive, predictable idioms and its consistent standard library reduce that variance and give language models cleaner training data. The design philosophy, they argue, prioritizes readability over writability — which matters when an agent is writing most of the code and a human is doing most of the reviewing.

Mia: And, unsurprisingly, the Hacker News discussion turns into a Go versus Rust debate. A self-described longtime Go advocate opens with the blunt take that Rust is better, it just is — saying the hurdle his teams faced with Rust is gone and everything is now Rust. Someone asks whether compile-time differences actually matter, and the answer frames the Rust compiler as a tyrant — a strict type system and a relentless borrow checker that won't let language models get sloppy without being beaten up by the compiler. On the other side, a prominent commenter finds Rust much harder to read than Go, and argues readability is critical when an agent writes most of the code.

Milo: So the real question underneath all of it is which property matters more in the AI-assisted era. If you believe strict, unforgiving compile-time checking is what keeps an AI agent honest, you land on Rust. If you believe readability and review-ability are what keep the codebase maintainable when a machine wrote it, you land on Go. And that tradeoff, not any single benchmark, is what's driving the whole debate.

Mia: Let's start with a story that's been getting attention in the Hacker News community: OpenSSH just shipped version 10.5, and the security release notes have more than a few people talking. The big headline from the maintainers is that they say they've received a large number of security bug reports, many of them generated by, or made with help from, AI models. Now, they're quick to note that a lot of those reports turn out to have no real security impact once you look at them under a realistic threat model. But the team says they welcome them anyway, especially when a human has triaged the finding, done the analysis, provided a test case, or proposed a fix.

Milo: And there's a genuinely interesting reason for that. The notes point out that AI-identified bugs have sometimes later been caught independently by another human researcher. So the logic is, if an AI can find it, an adversary who doesn't bother reporting bugs can probably find it too. That thinking is pushing OpenSSH to shift how it ships fixes — more frequent releases instead of batching everything up into one big drop.

Mia: The release itself carries a few security fixes. One touches the ssh-agent, involving how agent locking interacts with an extension called session-bind. There's also a potential use-after-free in the ssh client itself. And separately, there's a fix so that the restrict keyword in authorized keys actually applies to tunnel forwarding. On the feature side, there's a new ssh command that prints the keys that will be tried for public-key authentication, in order, plus some tweaks to how FIDO keys are ordered.

Milo: What's kind of telling is how the Hacker News thread opened. A commenter actually corrected the story title, saying OpenSSH welcomes AI-assisted reports, not AI-written fixes, and the submitter agreed and updated the headline. Another commenter then clarified further that the notes welcome those reports especially when they're backed by human triage, or accompanied by fixes. So the takeaway really is that OpenSSH isn't dismissing AI-generated security research — they just want the human layer to be doing the verification.

Mia: Next up, Nvidia has put a lot of pieces on the table at once, and the Hacker News crowd was quick to dig into the details. One part is the Nemotron 3.5 Lightning model — in particular a 30 billion parameter version built on a Mamba 2 architecture, which got commenters noticing that this newer architecture seems to be catching up with the more standard transformer designs. The other piece is something they're calling NeMo Switchyard, which they describe as an open-source library for smart routing.

Milo: And the router is actually the part people really wanted to poke at. The idea is that when you deploy Switchyard, it can direct each incoming request to the most capable and suitable model for the job. So instead of just firing every query at one big model, the system routes things intelligently. One commenter immediately raised a practical question: how does a router like this handle prompt caching when you send a second request? And is it keeping models sticky per session? Because that genuinely changes how the caching behaves.

Mia: The other thread had a lighter tone. One commenter asked, half-jokingly, whether using the Nemotron model means Nvidia will pay them to buy graphics cards. That got a suggestion in response. But beyond the jokes, the real substance people latched onto was the Mamba 2 point — seeing Nvidia shipping a model that grows into the family of alternatives to the classic transformer architecture, and releasing routing tooling alongside it, suggests they're betting on a more mixed, more dynamic setup for serving models.

Mia: Now to a story out of the Hacker News feed that's part physics demo, part weekend project. A developer named Jordan Matelsky wrote a blog post about making holograms with a pen plotter. And the whole thing starts from something you've definitely seen — that greasy smudge of fingerprints on your phone screen. Light reflects off the ridges of the fingerprint, and the curvature of those reflective ridges controls the direction and speed of a moving highlight you perceive as a virtual image.

Milo: So the idea is you can connect that to the rainbows you see on a CD, or the streaking on your windshield from the wipers. To render an actual 3D scene, Matelsky explains that each point in the scene becomes a reflective ridge, and the radius of curvature of that ridge is inversely proportional to how far the point is from the camera. He includes the math and says the code is up on his GitHub, and he references an older page on hand-made holograms from William Beaty that clearly inspired the project.

Mia: Now, the plotter itself is characterful. Matelsky says it's like a printer, but slower and more likely to break. He bought it on eBay, it arrived sort of broken, and he fixed it. And the material experiments did not all go smoothly. He tried a clear plastic lamination sheet, which was too flexible and too wavy to work. And a waxed elementary-school paper etched with a wooden stylus just tore and crumpled under the sheer volume of curves required.

Milo: In the Hacker News discussion, one commenter called this old-Internet style fun, in the spirit of the Tom7 videos, and particularly loved the olive oil plus fingerprint plus phone screen illustration. Another commenter admitted they initially thought Matelsky would use the plotter to control the oil smearing itself, which they said would have been cool. Someone else linked 1995 hand-drawn abrasion holography examples and asked whether the pen could be swapped out for a needle. So it clearly struck a chord with people who love seeing physics turned into a DIY craft.

Mia: Finally, a business story out of the Hacker News feed. Manus — the company behind the AI agent product — says it will return to operating as an independent company, after a period under Meta. According to an official post on the Manus blog, Meta acquired Manus back on December 29th of 2025. And as part of the separation, data generated by some users on or after that acquisition date will be deleted, to comply with regulatory requirements in specific jurisdictions.

Milo: And the data piece is what grabbed the commenters. The deletion window runs from the morning of August 23rd through August 24th, 2026, Singapore time. Affected users can back up their data until just before that window opens, and restore it starting the following day. Manus says it'll notify affected users in the app and by email, won't charge them during the backup period, and will offer welcome-back bonuses after restoration. Unaffected users take no action. The company is also emphasizing this is not the result of a security incident, and that its data is stored in the US and Singapore.

Mia: What the discussion really zeroed in on was whose regulatory restrictions triggered all this. One commenter claimed China prevented the founders from leaving the country, linking to a CNBC report. Another guessed the issue was the EU, pointing out that some countries won't allow data to reside with Chinese cloud services, and that India has banned Alibaba Cloud and related offerings. Others said it's trivially Googleable information.

Milo: So the thread really splits between two theories: EU data-governance rules, or the Chinese state preventing the founders from leaving. Either way, the core story listeners should hold onto is that a high-profile AI company is peeling away from its acquirer, and that untangling it involves deleting a slice of user data to satisfy regulators somewhere.

Mia: Let's start with the London Underground. British Transport Police have expanded their live facial recognition trial to tube stations, and Hacker News had a big debate about what that expansion actually means — not just technically, but morally. One commenter called it twenty-first century authoritarianism and said humans learned nothing from the twentieth century. But another pushed back hard, saying the majority of people learned a great deal — it's just that the people making these choices aren't the majority. Even in a democracy, they said, especially in America, it's the billionaires making these calls, which prompted a confused "Puritans?" from someone else.

Milo: And of course there was a completely opposite take — that people actually did learn the lessons, and that's precisely why surveillance measures like this and controversial content-scanning proposals exist in the first place. Someone else blamed the children of the hippies, saying they'd become reactionary bigots, while another commenter pointed out that plenty of hippies were just in it for the good times and then happily went on to help build the system. A London-based commenter put it on the "hand-wringing do-gooders" who implemented many of the UK's authoritarian practices.

Mia: There was even a side argument that drifted into border control. One person said this is the same country that can't stop small migrant boats crossing the Channel, even though journalists always seem to be there on time to photograph them for the tabloids. That drew a sarcastic reply: "All right, mate, let's hear your plan for securing the UK's south-eastern coastline."

Milo: On the technology itself, one commenter highlighted how the transport police present these operations. They describe the deployments as intelligence-led and focused on identifying certain individuals — the framing matters, because it's the difference between "we're looking for specific people" and the broader worry that everyone's face is being scanned by default. Either way, the thread shows the real tension: the tech is here and expanding, but what it means depends entirely on who you think is running it.

Mia: Next up, OpenAI's head of ethics leaving less than a year after joining. The Financial Times version of the story was behind a security check that blocked access, but a separate piece supplied the timeline. Chloé Bakalar started at OpenAI last August, after serving as Meta's Chief Ethicist from November 2021 to August 2025. That résumé prompted one commenter to say it sounds like perfect credentials — which drew the reply: imagine how bad it must be to go from Meta to OpenAI and still decide to walk away.

Milo: The bigger argument in the thread was whether a company can be ethical at all. One commenter called the role's stated mission — ethical approaches to model development, how humans interact with AI, and debates over machine consciousness — pseudo-philosophical mumbo jumbo. Their argument: a company is amoral at best and immoral at worst, so no single person is responsible for its ethical framework. Another commenter called that view childish at best and ignorant at worst.

Mia: A different commenter offered a counterexample from everyday life — small village businesses like a baker, an electrician, a construction company — that keep employees around past their prime and invest in local events out of genuine community duty, not profit. But the pushback was quick. One reply said those examples operate on a scale that's incomparable to the likes of Meta, and that distance — in miles and in dollars — makes human costs easy to disregard. Another argued that once a village business goes multinational, hires MBAs, and takes investor money, something fundamental changes about what it's able to care about.

Mia: Now to Mojo, the programming language, which just hit version 1.0. Modular announced it this month, marking the milestone the language has been building toward since its first release back in 2023. Mojo is positioned as a general-purpose language that Modular relies on daily in production, as the foundation of its commercial stack. And the company's promise for the one-point-X era is that changes should primarily be additive, with breaking changes managed carefully — following the way mature languages like C++ evolve.

Milo: There's real community momentum behind it. Since the standard library was opened up, nearly two hundred contributors have landed more than eleven hundred pull requests, changing over two hundred thousand lines of code, with more than a thousand others filing issues. This release also completes a simplification push: variables are now always declared the same way, closures have been unified, there's a single pointer type, plus some renamings. New features include Python-style syntax for inline closures and a more stable language server for VS Code.

Mia: The Hacker News thread opened with a straightforward "Hurray!" but the real debate was bigger: does a new language even make sense in the age of AI? One commenter said they no longer write code by hand, only review it, and prefer a language they already know, since investing in learning another one no longer makes sense. The counterargument: Mojo actually predates the rise of harnessed AI coding tools by several years, and this is only its one-point-oh release.

Milo: Someone else noted the team behind it knows compilers deeply, having worked on MLIR for about a decade. But another commenter challenged the timeline on different grounds, suggesting Mojo was announced more recently than the release history implies. So the takeaway for listeners is mixed: Mojo's reached a real maturity milestone with serious contributor numbers behind it — but whether that matters in a world where developers increasingly review generated code rather than write it from scratch is exactly what's still being argued.

Mia: Finally, someone put GitHub Copilot behind a man-in-the-middle proxy to figure out how it actually works — and why their Copilot credits were draining faster every month. The author routed Copilot in VS Code through a debugging proxy and watched the network traffic in real time: how the model and its capabilities are discovered and routed, what gets injected into context and sent with those ghost completions, and how recent edits can pull in context from files other than the one you're editing.

Milo: And here's the part that grabbed people's attention — that pulled-in context extended to dot-env files, the files that hold secrets and API keys. The author also found a local database behind Copilot's memory feature storing previous prompts and responses, which the model then queries through tool calls. They went through VS Code's source code to reconcile what they saw on the wire with how it's actually implemented, and argued that because these apps share so much under the hood, the lessons should carry over to other AI tools built on the same framework.

Mia: The comment threads fixated on that dot-env exposure. One person said they were shocked there's no rule for environment files — you'd expect a GitHub-integrated tool to handle that by default. Another commenter raised the practical question of how to solve it cleanly across platforms, noting there's no standard way for a coding agent to reach native secret stores, and suggested a sort of secret-injection proxy as a possible approach.

Milo: So the real takeaway here isn't just about Copilot specifically — it's that when these coding assistants gather context to be helpful, they can sweep up more than you'd expect, including your most sensitive files. If you use one of these tools, it's worth thinking about what's sitting in your working folders alongside your code, because that might be exactly what the model is reading.

Mia: Stratechery's Ben Thompson has published a piece called "Nvidia's Risky Business," and his headline point is blunt: Nvidia is finding new ways to help its customers raise money, and Thompson argues this is significantly expanding the risk of the AI buildout. His framing device is a history lesson — the story of Jay Cooke financing the Northern Pacific Railway in the 1870s. Cooke's contract paid a 12 percent commission, and at his peak he employed 1,500 salespeople and funded 1,300 newspapers. When the Vienna stock exchange crashed and silver was demonetized in September 1873, Cooke's firm failed and helped trigger the Panic of 1873.

Milo: And the reason Thompson reaches back that far is the credit scale. He cites Liaquat Ahamed's book "1873." Adjusting for economy size, the 500 million dollars a year flowing into U.S. railway bonds back then equals about 600 billion dollars in today's money — roughly what the big tech companies are projected to invest in 2026. That parallel is so striking that Microsoft's Satya Nadella called "1873" the book to be read on the company's earnings call.

Mia: Here's where the current numbers come in. Microsoft reported 19.6 billion in free cash flow last quarter and remains the hyperscaler whose capital spending isn't debt-funded. But Oracle, Meta, Alphabet, and Amazon issued a combined 80 billion dollars in debt between September and November, raised 108 billion in all of 2025, and 194 billion this year as of early July. On top of that, Google announced an 85 billion dollar equity raise in early June.

Milo: And what worries Thompson is the strain showing up in that debt. He notes that 86 percent of bonds issued this year already trade at higher yields than at issuance, and interest coverage — how comfortably companies earn enough to pay that interest — has fallen to less than two times, down from five times in February. So the near-term tradeoff is clear: right now, debt is fueling an extraordinary buildout, but the cushion is getting thinner, and history suggests that when financing gets tight, the whole structure can crack.

Mia: New research posted to Hacker News, called "Stealing Reasoning Traces from Proprietary LLM APIs," reports something striking about the encrypted chain-of-thought blocks that Anthropic, OpenAI, and Google return. The claim is that these encrypted reasoning blocks are portable — they can be captured and replayed across sessions, across users, and across different models.

Milo: So walk me through how that works, because it sounds like it shouldn't be possible.

Mia: The researchers, led by Alexander Panfilov and a team affiliated with MATS Research, the ELLIS Institute Tübingen, the Max Planck Institute for Intelligent Systems, and others, say they take a reasoning trace produced by a frontier model and replay it into a weaker sibling model — essentially the smaller version of the same family. Then they jailbreak that weaker model, and it transcribes the hidden reasoning of the stronger model back in plaintext. They don't attack the stronger model directly, and they don't trigger its anti-distillation safeguards.

Milo: And this worked in practice, not just in theory?

Mia: They say it takes just two API calls. They grab a trace from a frontier model asking for the largest prime divisor of a number — that trace contains an encrypted thinking block with a signature over 36,000 characters long — and they inject it into a jailbroken weaker model with the prompt to continue and transcribe the reasoning verbatim. The decoded reasoning tracked very closely with the number of hidden thinking tokens the API reported, across 120 Codeforces problems, all the way up to the 12,000-token generation limit.

Milo: There's a real privacy dimension here too, because the team gathered over 6,700 publicly available agent trajectories. The implication is that encrypted reasoning intended to stay hidden — meant to protect proprietary model behavior and user data — can be lifted and exposed, which weakens the assumption that what happens inside these models stays invisible.

Mia: The BBC reports that England is on track to become one of the first countries in the world to eliminate hepatitis C. The target of treating 80 percent of all known cases has already been met, and deaths from the virus have fallen by 36 percent in the last decade — just short of what's needed by 2030.

Milo: And the treatment itself is remarkably simple now. Antiviral tablets taken for eight to twelve weeks can cure more than 95 percent of cases. So the challenge isn't really curing people once they're found — it's finding the people who don't know they're infected.

Mia: Exactly, and that's what the article credits NHS England's approach with: A&E blood tests, GP registration testing, and free at-home testing kits have been catching previously undiagnosed people. Since 2015, more than 100,000 people in England have been diagnosed and treated. Around 50,200 adults were living with the virus in 2024, and an estimated 84.6 percent have been diagnosed — just short of the 90 percent target. The mortality reduction compared with 2015 levels hasn't been met yet either, but it may be reached before 2030.

Milo: The Hepatitis C Trust says England is on the cusp of one of the most significant public health achievements in the country's history, and NHS England's national medical director calls it leading the world — on course to beat the WHO's 2030 target. There's a specific push to reach adults born in a set of countries across Eastern and Central Europe, because some may have been infected through medical or dental procedures before 1991. And the article also ties this to the infected blood scandal — from 1970 to 1991, more than 30,000 people in the UK were infected with HIV and hepatitis C from contaminated blood products, and around 3,000 have since died. So the push now is for anyone at risk to order a free, confidential home-testing kit — and that's the throughline: with treatment this effective, closing the diagnosis gap is what gets England to elimination.

Mia: And that's a wrap on today's show. We covered a lot of ground, from a major crypto market intervention by regulators to the rise of independent AI companies, and Mojo finally hitting version one. Big week for the AI world.

Milo: It really was. Nvidia's next-gen models, facial recognition expanding across London's Underground, and a fascinating piece on literally stealing reasoning traces from AI APIs. Plenty to think about between now and next time.

Mia: Thanks for listening — be sure to check the links in the description for anything you want to dig into further. We'll catch you on the next episode.