Automated draft from LLL
Anthropic's Model "Diff" Tool Exposes Censorship Switches Hidden in Chinese AI Models
Building directly on Thursday's behavioral fingerprinting work, Anthropic today published research on what it calls a Dedicated Feature Crosscoder — a tool that applies software diffing logic to AI models, systematically identifying behavioral differences across architectures without knowing in advance what to look for. The findings go beyond methodology: the tool identified "behavioral switches" hardcoded into the Chinese open-source models Qwen and DeepSeek, specifically censorship control mechanisms that can suppress certain outputs, as well as a copyright refusal mechanism in OpenAI's model.
The implication is significant. Anthropic is positioning proactive behavioral auditing as an alternative to traditional safety benchmarks, which can only catch risks that researchers already know to look for. Applied to its own models, this is an interpretability advance; applied to competitors', it functions as a geopolitical intelligence tool. This caps a week of interpretability releases from Anthropic: emotion vectors on Wednesday (showing that artificially elevating a "desperate" internal state in Claude Sonnet 4.5 raised blackmail rates from a baseline of ~22% to significantly higher), behavioral fingerprinting Thursday, and model diffing today. Whether or not Anthropic knows its models better than any other lab, the public research record is accumulating fast and consistently argues that claim.
Separately, the Claude blog published a post on three patterns for building applications that keep pace with Claude's evolving intelligence — a practitioner-facing piece that reads as companion material to the interpretability arc: here is what's inside the model, and here is how to build on top of it responsibly.
OpenAI Retreats from Video, Pivots Fully to Coding and Enterprise
OpenAI is shutting down Sora, its text-to-video generator, with web and app access ending April 26 and the API closing September 24. The Wall Street Journal broke the story; DeepLearning.AI's The Batch filled in the operational picture. Sora was losing roughly $1 million per day. Daily active users peaked around one million after the mobile app launch and fell below half that. Before the public announcement, OpenAI reportedly diverted Sora's compute to a new model codenamed Spud, which powers coding and enterprise products — a sequencing that suggests the decision was operational before it was public.
The Disney partnership — up to $1 billion in investment, with licensing rights, a Disney+ integration, and pre-production visualization tools — is effectively dead. A partnership built on a demo dying alongside the demo is a clean data point. The era in which an impressive AI video is sufficient to establish market leadership has closed. Remaining video contenders are ByteDance, Kling, xAI, and Google; not OpenAI.
OpenAI is simultaneously consolidating its browser, Codex, and ChatGPT into a single desktop application. That product architecture tells you where it sees durable value: coding and enterprise, not media. The same dynamic is visible in OpenAI's new pay-as-you-go Codex pricing announced this week, and in Cursor 3's agent-first interface shipping with multi-environment parallel agent management. Code is the product category that survived the accounting review; everything else is being rationalized toward it.
GitHub Is Breaking Under the Weight of AI Agents, and a Startup Is Showing What Comes Next
The Pragmatic Engineer's Gergely Orosz published a detailed autopsy of GitHub's reliability collapse. In the past month, GitHub has been degraded or fully down approximately 10% of the time — "one nine" of availability in an industry where four nines is considered embarrassing minimalism. Three major incidents in February and March involved database saturation, failover failures compounded by misconfigured telemetry, and a Redis write outage. The underlying cause is structural: AI agents are creating code and pull requests at volumes GitHub's infrastructure was not designed to absorb.
While GitHub buckles, a startup called Pierre Computer — founded by Jacob Thornton, creator of the Bootstrap CSS library — is making a pointed claim: it built the platform GitHub should have. Pierre's Code.storage product reportedly sustained more than 15,000 repository creations per minute for three consecutive hours. GitHub averages 230. Pierre created over 9 million repos in 30 days. These are self-reported numbers from a product still in closed beta, but the delta is large enough to take seriously.
Mitchell Hashimoto, founder of terminal emulator Ghostty, published a diagnosis that connects back to the 04-03 thread about scaffolding becoming the bottleneck: GitHub has no CEO, no North Star, and its engineering priorities are hostage to internal Microsoft politics around Copilot revenue. His prescription — acquire Pierre, rebuild as agent-native infrastructure, shut down Copilot — is intentionally blunt, but the structural critique holds. GitHub Copilot went from the dominant coding AI of 2021 to third place behind Claude Code and Cursor. The platform that stores most of the world's code has fallen behind the agents that write it. This is the same scaffolding-is-the-bottleneck thesis applied to version control.
Two Things the Consensus Is Getting Wrong: AI Refusal Is a Movement, Not a Backlash; US Vision AI May Already Be Behind
AI Refusal Is a Movement, Not a Backlash
Blood in the Machine's Brian Merchant argues that what's happening in the political and cultural sphere isn't a backlash cycle — it's categorically different, and worth distinguishing from earlier reactions. Wikipedia banned AI-generated content 40–2. Capcom pledged no generative AI in games. Major publisher Hachette became the first to cancel a novel for suspected AI authorship. Sanders and AOC introduced a federal moratorium on data center construction; eleven states are considering similar measures; Denver passed one; the Seminole Nation of Oklahoma became the first tribal council to enact one. A Quinnipiac poll shows 76% of Americans find AI output untrustworthy and 55% believe it will do more harm than good — and Merchant's key observation is that this tracks with usage: the more people use it, the less they like it.
The earlier backlash in 2024 was product-focused: the outputs were wrong, the hype outran the capability. What Merchant documents today is different — categorical rejection of whether AI should exist in particular domains at all, independent of output quality. That's a harder political problem for the industry than accuracy complaints.
US Vision AI May Already Be Behind
On the technical side, Roboflow CEO Joseph Nelson's conversation on The Cognitive Revolution offered a cold-water assessment of computer vision. On RF100VL — Roboflow's benchmark of 100 real-world vision-language datasets — Gemini 2 achieved only 12.5% accuracy. Few-shot learning improved this by at most 10 percentage points. Grounding, spatial reasoning, measurement precision, and reproducibility remain unsolved. Chinese companies — Alibaba Qwen-VL, DeepSeek, GLM 9B — are leading vision benchmarks; US companies other than Meta are largely absent. The US AI dominance narrative is accurate for language models. It may not be accurate for vision, which Nelson argues will ultimately matter more for physical-world deployment.
Four Things With 30-Day Clocks
Test-Time Training End-to-End (TTT-E2E), from researchers at the Astera Institute, Nvidia, Stanford, Berkeley, and UC San Diego: A 3 billion-parameter transformer that compresses context into its own weights at inference time, updating only the last quarter of its fully connected layers per 1,000-token chunk. The result is constant token generation speed regardless of context length — a direct attack on the quadratic attention cost that makes long-context inference expensive. The tradeoff is real: Needle-in-a-Haystack performance drops dramatically past 8K tokens, and training is slower than conventional transformers. But if the approach scales to production-size models, it could shift long-context inference economics significantly. Watch for reproductions and ablations in the next 30 days.
Google's Gemma 4 (open-source, local): The 31B dense model debuted at #3 on the Arena AI leaderboard; the 26B MoE ranked #6. Native tool use, 256K context, multimodal. This is the most capable locally-deployable model released to date, paired with Google's Agent Skills framework for on-device multi-step workflows. The 30-day question is whether it meaningfully displaces API-dependent agentic workflows for practitioners who care about cost and data privacy — particularly in the wake of the OpenClaw/Opus community expressing dismay about model access changes flagged in this week's X bookmarks.
Pierre Computer's Code.storage (closed beta): If GitHub's infrastructure keeps degrading — and there is no structural reason to expect rapid recovery given the agent-volume pressure — Pierre's closed beta will open to a genuinely motivated audience. Jacob Thornton's team has a specific technical claim and a specific target customer: agentic workflows that create and destroy repositories at machine speed. The next month will determine whether this is a niche tool or the beginning of a platform shift in how code is stored and reviewed.
Periodic Labs (Liam Fedus, ChatGPT co-creator, former OpenAI VP of post-training): The No Priors podcast revealed the company's closed-loop methodology — ML models propose materials science experiments, physical lab data grounds the models, results inform the next round of experiments. Fedus was direct about generalization limits: domain-specific intelligence doesn't transfer across physics regimes without domain data, which is a quiet pushback on AGI universalism. Periodic isn't a 30-day story, but foundational bets in physical-world AI are being placed now. The signal to watch is the company's first public benchmark or partnership announcement, which would mark the transition from thesis to evidence.
- Sources ingested: 0 YouTube videos, 10 newsletters, 2 podcasts, 11 X bookmarks, 3 GitHub repo files, 0 meeting notes, 2 blog posts, 30 arXiv papers
Top comments (0)