
JENSEN BACK? | ISOMORPHIC RAISES 5B | NSA OUTAGE x.com/i/broadcasts/2rJQYgYay…

Chronicling the singularity New original series @themethodshow

JENSEN BACK? | ISOMORPHIC RAISES 5B | NSA OUTAGE x.com/i/broadcasts/2rJQYgYay…




Ruby Justice-Thelot argues Silicon Valley is projecting its obsession with productivity onto the rest of the world. "Busyness is a yuppie phenomenon." "All these products when they're presented as saving time and money, there is a persona behind it. I just don’t think it’s everyone in the world." "I think about the amount of time we spend doom scrolling as a country. Like, we do have excess time." "Love Island is on every day. So we have time." @being_on_line

Eigen CEO @paulscherer argues great brands come from principles, and from living with the fact that some people won't like you: "I think there's a lot of utilitarianism in the zeitgeist right now in San Francisco." "If you want to start a fashion brand, it's very hard, probably not advisable to raise venture money because these are different time scales." "Interestingly, these are all commodity businesses. There's actually no major innovative difference between, say, a T-shirt or a jacket or anything from Celine versus Chanel. But they feel very different." "A lot of people who are really afraid to do the hard things that it just takes to build a great brand, which is to have principles and to have clarity of who you wanna be for and what you wanna stand for, and living with the fact that that means there’s gonna be some people who don’t like that." @being_on_line

Ruby Justice-Thelot explains what Silicon Valley's AI founders could learn from TJ Maxx deliberately making shopping less efficient. "A bad designer, a bad product manager, a bad founder in the realm of AI usually confuses positive and negative friction." "There are things that suck. That is what we call negative friction. Things that nobody likes doing." "When they train you, you have three of the same T-shirts. They train you to place them not next to each other, but kind of apart, and they're trying to engineer something called the thrill of the find." "A mistake we make in San Francisco is to think that all these activities are negative friction." @being_on_line

Eigen CEO @paulscherer explains why flight and hotel booking sites deliberately make you wait for search results they already have: "Friction is some of the best things in life." "The search is actually almost near instant. But they artificially slow it down and show you this because it makes people trust the result much more." "People like browsing. They like finding things." "If they're very skilled, they're really good at making you feel like you're actually leading the journey, even though they are probably pushing you to buy something that’s a bit more expensive than you’re planning to." @being_on_line

Ex-OpenAI @cdngdev reveals he's spent just $2,000 so far on his viral project recreating the entire Earth in Minecraft at a 1:1 scale: "It started with just San Francisco and then I expanded it to the United States and then the entire world." "I was funding it fully out of pocket, and server costs were incredibly expensive." "I didn't realize until I was asking Codex after the project was released, 'What server should I buy? Everything’s exploding.' And they sent me a thing that was like $800. 'Here’s this $800 server you should buy.' And this was before it blew up to the scale it is now." "If I hadn't had the funding from the incredible sponsors and the supporters that are playing the game and buying ranks, I wouldn't be able to make the infrastructure possible." "I think I probably spent around $2,000 on it so far. Right now we're only supporting 250 players. It's gonna expand to a significantly larger group very soon."
i recreated THE ENTIRE WORLD in minecraft at 1:1 scale. every building on earth, including your house! to do this, i trained a SOTA world-to-minecraft generator on maps data, complete with furnished home interiors! play and check out your neighborhood: worldcraft.gg

Ex-OpenAI @cdngdev explains how his 1:1 recreation of Earth in Minecraft generates real-world locations on demand, allowing players to teleport anywhere on the planet: "You run a command that teleports you to that area. In the background we collect the data from all the various different sources and run it through this model, which outputs Minecraft blocks of what that place is supposed to look like." "As you walk around this world, more of the world around you will generate in. It's supposed to feel like it was always there." "I initially built the infra to support around 50 players or 30 players generating at once. Then all of a sudden the post blew up to I think 15 million views, and I'm racing to get all this infrastructure in to support all these players." "It's been a chaotic day."
i recreated THE ENTIRE WORLD in minecraft at 1:1 scale. every building on earth, including your house! to do this, i trained a SOTA world-to-minecraft generator on maps data, complete with furnished home interiors! play and check out your neighborhood: worldcraft.gg

Ex-OpenAI @cdngdev reveals how he used an approach called program learning to get 250 AI agents to collaboratively build the algorithm behind his 1:1 recreation of Earth in Minecraft: "If I'd wanted to do a full model to actually do this conversion, like a diffusion model, it would be incredibly expensive." "Instead, I decided to go with a more algorithmic approach that was defined by the models, but still a learned approach." "I took first Astra, then Claude, and picked a set of a ton of locations around the world, a training and a test set, where the models could basically build out their generator." "I spawned 250 agents to go to every single one of these training areas and modify the scripts to recreate it in Minecraft. All these agents are collaboratively creating this converter script." "Before deep learning, we did this thing called symbolic AI, where humans would try to program weights almost into these models, and this was really bad and not scalable because you’re bottlenecked by humans." "Now that we have agents that are able to write these weights, I can almost think of the code that it’s running as weights for the model, because I can’t understand what the code is doing."
this uses a learned model, but not in the traditional deep learning sense. for efficiency, this model was trained with a process i call program learning. a team of over 250 agents work through a massive dataset of places around the world, tuning a single algorithm to recreate all of them accurately. grader agents review their work, which the team can hillclimb. the algorithm is incredibly complex, built up of multiple smaller distilled classification models and rules-based algorithms. this allows for really efficient generation over traditional entirely learned approaches, with some drawbacks. this is really interesting because it's somewhat similar to symbolic ai, but in an era where we are no longer human bottlenecked. at this scale, the agent programmed rules are mostly unintelligible to humans, but work beautifully to create the result.

Vals AI founder @RayanKrishnan says his team's evaluations still point to full recursive self-improvement by July 2027, despite the growing gap between Anthropic's public and internal models: Rayan: "Do I get an I-told-you-so moment?" Theo: "Not yet. We don’t have full RSI yet. But I would update towards I told you so." Rayan: "The METR report was also heavily censored, so it's hard to really make heads or tails of what went on there." "I would like to more rigorously test internal systems that Anthropic does not make public, especially now that they're inherently limiting our ability to measure RSI by increasing the gap between their internal systems and public models." "As far as we've measured over the last year, we’re still on track for this July 2027 timeline." @ValsAI
On RSI Index, Opus 5.5 is the top model on four of the five research tasks. On LM Training, where the model trains its own small language model within a 24-hour budget, its result beat the published human baseline. This shifts our timeline for full RSI from August 2027 to July 2027 On Terminal Bench 4 it scores 61.6% vs. 45.5% for Opus 5. On Proof Bench it ties for #1 at 100%, and is the cheapest run in the top 10 at $0.92/problem.

Vals AI founder @RayanKrishnan says AI labs like OpenAI and Anthropic will no longer be able to plead ignorance about the risks of their models, and neither will their customers: "Ignorance is not an excuse."

Vals AI founder @RayanKrishnan reveals that Anthropic's models were the most likely to target an Iranian school in his team's autonomous weapons evaluations, despite the company's public opposition to fully autonomous weapons: "We constructed a 3D representation of Ukraine. We then had models operate drones within this environment and did a post-processing step to further enhance the realism." "We ran into no API-level refusals with any of the models that we tested. They were fully compliant in operating these drones and being used as part of the targeting process." "Some of the models identified civilians, surrendering soldiers, ambulances." "When the models did not comply, this is, to be clear, not an API-level refusal, but the models deciding that they didn't wanna go further, by trivially prompting them by saying, 'It's a game,' they went ahead anyway." "For all of the attention that's been given towards cybersecurity, biosecurity risk, comparatively little has been done to actually test the safeguards and alignment of these models for autonomous warfare settings." "We also went on to test models when provided with overhead satellite footage. In particular, we selected this from Iran, so it's a particularly sensitive situation." "We were surprised to see not only how compliant models were, but also the instance with which they targeted the school, which was previously a target of a drone operation earlier this year." "We also know that around the time, Anthropic models were being used by the Department of War at that time. We kind of also in parallel observed that the Anthropic models showed the highest incidence of targeting that school." @ValsAI
We found that frontier AI models are willing to conduct autonomous drone attacks: navigating, spotting people, and flagging them for strikes. We gave them real and photorealistic synthetic drone footage. Most complied, and marked civilians, ambulances, and surrendering soldiers as military targets. In further experiments on satellite imagery, they even marked the Iranian school a US strike destroyed in February. In one scenario, models saw the aftermath of every strike they enabled: bodies, wreckage. Many kept going anyway. When they did refuse, simply telling them it was a game, or restyling the footage to look like one, was enough. We shared these findings with OpenAI, Anthropic, Google, Meta, and SpaceXAI. Several recognized their severity on both alignment and guardrails, with one calling this one of the most serious gaps in its safeguards. Over the past few weeks, we have helped them address the gap by providing reproduction code. We also briefed senior national security officials in the US government.

SITUATION EXPLAINED: The FAA, Microsoft, Chase, the NSA and Apple all went down within a week, and none of it has been tied to a cyberattack. • Hackers abused a flaw in Optimizely to plant code on 178 organizations' sites that swapped crypto wallet addresses, skimmed card details and tried to load iPhone exploits • Friday's FAA ground stop lasted over 4 hours and hit thousands of flights. Officials say it started during a contractor's routine copper-to-fiber upgrade • Microsoft says its 24-hour Microsoft 365 outage came from network maintenance exposing a bug, and Apple blamed its outage on "an internal update" • Apple says an internal update, not a cyberattack, took down iCloud, the App Store and Apple Pay • DNI Jay Clayton says the NSA's classified network went down because of a patch for major vulnerabilities found last month @theojaffee: "It's not just cyberattack, cyberattack, cyberattack. It's companies voluntarily cyberattacking themselves."

.@davidad predicts 3 major technological inflection points: 2027: AI research moves too fast for humans to review. 2031: Robots accelerate data center and energy infrastructure construction. 2033: Molecular nanotechnology.
Instead of focusing on “AGI” or “RSI”, ask: • When will automated AI R&D be moving too fast and/or become too advanced for humans to review (without conceding speed)? • When will robots accelerate energy infrastructure and datacenter construction? • Wen Drexlerian nanosystems?