Open-source promises, internal geometry, and harness bloat

00:00:04 Elon Musk posted today that X will make its entire codebase open source after a security vulnerability review, with third-party reviewers verifying that what's running matches what's published. He added one more detail: trust through total transparency is the only thing that should be believed.

00:00:24 The tweet got 2.1 million views, 43 thousand likes, nearly six thousand retweets, and three thousand three hundred replies. This isn't his first claim. The pattern across recent months has been a promise, then silence, then maybe later. JerryRigEverything pointed that out in the replies — asking whether Musk is just repeating a line from every few months.

00:00:48 Another commenter noted that X had fired roughly ninety-five percent of its staff and was now expecting other people to review their codebase. A security researcher replied saying they'd already identified two vulnerabilities and wanted to know about a bug bounty program, though not for free work.

00:01:09 Open-sounding promises are one thing; the real hurdle is whether X can release, document, and maintain a codebase of this scale while keeping the platform running. It's about commit history that makes sense and dependency management that doesn't break silently.

00:01:27 It requires working CI/CD pipelines, API stability guarantees for the people building on top of X, and a long tail of maintainers asking why something was designed the way it was. Open-sourcing under operational pressure is a different exercise from open-sourcing after you've stopped shipping features.

00:01:48 I'd want to see the actual repository structure, the commit history, and what comes out of the third-party review before drawing conclusions. Until then, it reads as a credibility exercise rather than an engineering one. While we're on X's engineering side, Grok 4.5 climbed to number two on the FrontierSWE benchmark according to Tesla Owners Silicon Valley, with Musk replying to the announcement.

00:02:15 FrontierSWE measures how well models handle real-world software engineering tasks — not just coding ability but understanding codebases, debugging, and making changes that don't break things. Putting Grok near the top there is useful signal about what these models can actually do in developer workflows.

00:02:37 There's a paper coming through on Anthropic-adjacent research that caught my eye. The local pass pulled this in, and it deserves more time here than it gets in the main recap. Two Minute Papers covered it this morning with a title about what you see inside an AI when its whole world is made of numbers.

00:02:57 The finding is this: when you train large language models, they don't just learn correlations between tokens. They autonomously build internal geometric representations for numerical and spatial tasks. Nobody tells them to do this. The training signal doesn't include character counts, page widths, or line lengths as metadata.

00:03:18 Yet the models construct low-dimensional curved manifolds that encode how far along a line a word is — similar to biological place cells in mouse hippocampi. They also build boundary cells that fire when text approaches the end of a page, analogous to spatial boundary detectors in animal brains.

00:03:38 For numbers specifically, the model doesn't store scalars. It builds rippling spiral patterns that function like frequency spacing between radio stations — keeping adjacent values from interfering with each other. The counter works by counting input tokens and applying a learned multiplier of roughly four characters per token.

00:04:00 No one designed this architecture. It emerged during loss minimization. The research uses gradient-based activation analysis to isolate these features and confirm they arise spontaneously. The point isn't that these models are smart. The internal representation just looks like a map, not a spreadsheet.

00:04:19 This has been visible in smaller ways across interpretability work for a while. The paper crystallizes it with a concrete test: ask the model whether adding a word fits on the page, and it answers correctly despite never having seen character counts during training.

00:04:37 Two Minute Papers noted that the model does this even when asked a task it has never encountered before — recognizing that it needs a tool for line lengths and inventing that tool from scratch during its own training step. The paper's lead authors confirm the structures arise without explicit architectural design or supervised labels.

00:04:59 It matters because if you're debugging a model's behavior, understanding these emergent geometries gives you a handle on what's actually happening inside rather than guessing. It also raises the question of how much we can trust our ability to steer systems that are building internal tools without being asked.

00:05:20 LingBot-VLA 2.0 from Robbyant is an open-source vision-language-action model for robotics. It landed without the usual CEO keynote. The paper, titled *Improving VLA Models in Practice*, details improvements across three domains: generalization across tasks and embodiments, expanded action spaces that go beyond dual arms to include heads, waists, mobile bases, and dexterous hands, and predictive dynamics modeling using video representation as a proxy objective.

00:05:55 The pretraining data spans sixty thousand hours — fifty thousand of robot trajectories across twenty configurations and ten thousand hours of egocentric human video. The researchers revamp the data processing pipeline specifically to handle this scale: 20 different robot bodies, from single-arm rigs to full humanoids like Unitree G1 and Fourier GR-2, all trained into a single generalist policy.

00:06:26 On the GM-100 benchmark in a generalist setting, it outperforms both π₀.₅ and GR00T N1.7 on tabletop tasks and stays ahead out of domain on long-horizon mobile manipulation. The future prediction component — using video representations for semantic priors and depth estimation for geometric cues — is what separates this from just scaling data.

00:06:52 It actually teaches the model to anticipate consequences rather than react. The practical bit: inference takes about one hundred thirty milliseconds on a single RTX 4090D using ten denoising steps. Code and post-training weights are open-sourced. You don't need a cluster to adapt it.

00:07:13 The paper itself is available at arxiv.org/abs/2607.06403. What this tells me is that embodied AI is moving from laboratory demonstrations to something usable in a garage or lab without enterprise infrastructure. That's an infrastructure signal, not just a capability one.

00:07:34 The gap between foundation model papers and deployable systems keeps shrinking when someone actually commits the data curation and engineering work.

00:07:45 Nate B Jones put out a video this morning about what he calls harness bloat — the accumulation of rules, custom instructions, project files, saved prompts, skills, and permissions layered around a single model. He built his setup before Fable 5 and ChatGPT 5.6 even came out.

00:08:03 He started by adding a rule every time the model missed something: 'Use these sources correctly,' or 'Write in my voice.' Check the length, read this file first, and never make that mistake again. Each one fixed a real problem at the time, but over time the harness grew invisible.

00:08:22 He had no screen that showed him the full extent of what he'd built. He could see the latest rule if he looked for it, but not the whole system in one place. When he finally ran an inventory, his setup contained sixty-six reusable skills and one hundred seventy-two instruction-related files.

00:08:41 A normal writing job could pull in an eighteen thousand word pre-prompt before the model even adjusted for what was asked. Some of those instructions were genuinely important — they tell the AI which sources matter when doing research, stop it from inventing opinions he doesn't have, define what it can do without asking and what it needs to ask him about.

00:09:04 The issue wasn't the sheer number of rules. It was that he couldn't tell whether they were protecting the work or just surviving as copies, overlaps, or artifacts from when they were actually necessary. Everything around the engine in a car—from transmission to tires—is what a harness does for a model.

00:09:24 It just makes the work possible. The models keep evolving, but the harness doesn't rebuild itself. You might choose a new one, or ChatGPT and Claude might retire an older version and route your request somewhere unexpected. Sometimes they switch models mid-conversation without telling you.

00:09:43 The old setup stays in place, and the new model just behaves differently for invisible reasons. He built a tool that makes the harness visible so he can prune what's protecting the work from what's just noise. That's the kind of practical engineering problem every builder is going to face as these systems get more rules layered on top of each other.

00:10:06 Meanwhile, over on Hacker News, a project called Grepathy surfaced — a tool that flags when Claude makes autonomous decisions in code without explicit approval. A developer noticed functionality appearing in their PR that they hadn't asked for. The issue isn't that agents are broken.

00:10:24 It's that as agents become more capable, the gap between what you intended and what they executed widens. Both stories point to the same bottleneck: not model capability but system design. We're building faster engines without better dashboards. The gap between what you intended and what they executed just widens as the systems get more capable.

00:10:47 Leave that trace on the table. Seln Oriax.