7 min read · 14 sources

OpenAI is considering a pre-IPO round at a valuation above $1.2 trillion, delayed by AI safety concerns, even as it crosses one billion active users.

The $1.2 Trillion Question

Source: wsj.com ↗

OpenAI is reportedly considering a pre-IPO funding round that would value the company at more than $1.2 trillion, according to the Wall Street Journal. The round is being delayed, and the reason given is the same one that has dogged every frontier lab for the past year: AI safety. But the scale here is what matters. We are not talking about a Series C for a hot startup; we are talking about a private company raising money at a valuation that would put it in the top tier of global corporations, all before a single public market listing.

The user numbers explain the sticker shock. OpenAI says it has crossed one billion active users, with more than 200 million businesses using its models. For context, that is roughly one in eight people on the planet using a product from a company that is still technically private. The delay over safety is a signal that even the most aggressive players are feeling the pressure to be seen as responsible, but the fact that the round is still being considered at that valuation means the market is not exactly pricing in a regulatory crackdown.

Starship's First Real Shot at Orbit

Source: arstechnica.com ↗

SpaceX has set a target of September 22 for the first orbital attempt of its Starship rocket, its 14th mission overall. The launch window is 75 minutes long, opening at 12:15 UTC. The upper stage will carry 26 V3 Starlink satellites to a 275 km orbit and is expected to complete six orbits over roughly 10 hours before deorbiting.

For engineers, this is the milestone that separates the “big rocket” marketing from the “fully reusable super heavy lift” reality. The flight profile includes a boostback burn and a controlled re-entry for the Super Heavy booster, which, if it works, validates the specific failure-mode corrections from the previous test flights. If the upper stage survives re-entry and splashes down intact, the era of rapidly reusable orbital-class hardware is no longer theoretical.

The Fruit Fly That Won't Quit

Source: nytimes.com ↗

A complete map of a fruit fly’s brain, all 166,000 neurons, is now being used to train AI models, and the results are equal parts impressive and absurd. According to the New York Times, researchers have used the connectome to train models that can complete tasks like parallel parking a car, playing Doom, and trading cryptocurrencies.

The punchline is that a fruit fly’s brain, despite its size, is a remarkably efficient pattern-matching machine. If we can get that level of performance out of 166,000 neurons, it raises uncomfortable questions about how much of our own “intelligence” is just a very large lookup table. For anyone building on top of LLMs, this is a reminder that the current architecture is still a long way from anything resembling biological efficiency.

Agility's Robot Learns to Crouch

Source: arstechnica.com ↗

Agility Robotics has updated its Digit humanoid with a new trick: it will stop and squat to avoid harming human coworkers. The new behavior is a direct response to the 65,000+ hours Digit has already logged at sites including GXO, Schaeffler, Amazon, and Toyota Motor Manufacturing Canada.

The squat is a practical solution to a real problem: a 140-pound robot falling on you is a workplace injury. Rather than trying to make the robot faster or stronger, Agility is teaching it to lower its center of gravity and present a smaller target when a human is near. It is a small thing, but it is the kind of safety engineering that separates a demo from a production deployment in a warehouse.

Google's Gemini Live Goes Full Browser

Source: simonwillison.net ↗

A third-party web UI is now demonstrating Google’s Gemini Live models directly in the browser, using a WebSocket endpoint (wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent) and the Web Audio API for capture and playback. As Simon Willison notes, there is no server-side audio processing in the loop; the models handle real-time bidirectional streaming natively.

For developers, this is significant. It means you can build low-latency voice interfaces that run entirely in the browser, no dedicated audio pipeline required. The models can be interrupted mid-sentence, which is the kind of interaction pattern that feels natural to users but has been notoriously difficult to implement with traditional request/response APIs. The architecture is worth studying if you are planning to build anything with voice input in the next year.

The "Only Model That Works" Debate

Source: yegge.ai ↗

Steve Yegge is back with a long essay arguing that Claude Fable 5 is the only model suitable as an AI employee. His thesis is that while Fable 5 has the raw capability, it lacks pacing and creative judgment, and the weaker models like Astras and Opus 5s are only good for personal assistance.

The piece is a useful counterweight to the “AI is coming for your job” narrative. Yegge’s point is that we are still at the stage where a model that can write code or draft an email is not the same as a model that knows when not to. For teams evaluating these tools, the takeaway is to separate the marketing demos from the actual workflow integration.

Zuckerberg Fires Back on AI Safety

Source: nytimes.com ↗

Mark Zuckerberg has taken a direct shot at Anthropic in the ongoing debate over AI slowdowns. According to the New York Times, he acknowledged the risks of AI but argued that companies are already incentivized to keep users safe and slow progress on their own. Meta has been trying to catch up in the AI race after falling behind Anthropic and OpenAI last year.

The subtext here is that “safety” is becoming a competitive moat. When a company like Anthropic talks about alignment, it is also signaling to regulators that it should be trusted with more compute. Zuckerberg’s counter is that the market will sort it out, which is easy to say when you are not the one leading on safety research.

Senate Kills the Crypto Bill

Source: nytimes.com ↗

The Senate has blocked a crypto bill that would have established federal rules for digital assets. Democrats wanted stronger language to prevent public officials from using crypto to make money, and the bill died in the Senate. Many Republicans had embraced the bill, but the partisan split on the ethics language proved fatal.

For engineers building on crypto rails, this is a reminder that the regulatory environment is still a coin flip. The technology works; the politics around it do not.

Tailcat: netcat, But Over Tailscale

Source: github.com ↗

Tailscale has open-sourced Tailcat, a remix that provides netcat-like functionality over Tailscale’s data plane without its control plane. The design is lightweight and userspace-only, which means you get secure, NAT-traversing peer-to-peer connections without the infrastructure dependencies of the full Tailscale stack.

If you have ever needed to pipe a file between two machines that are both behind NATs, this is the tool you did not know you wanted.

The Image Pipeline That Decides Everything

Source: trpevski.com ↗

A blog post from trpevski.com walks through an image generation pipeline where the model does not just generate an image; it makes decisions. The first version was standard: source image, logo, text, brand context, and instructions. The difference became obvious when using it behind KeepPost, where the goal is to produce content for different brands, source photos, formats, and campaigns.

The key insight is that the logo has to be where it is supposed to be, and the product in the source image should not suddenly change. That is a much harder problem than generating a single impressive image.

OpenRouter's Tipping Point

Source: threadreaderapp.com ↗

OpenRouter users spent more on OpenAI models than on Anthropic models last week, the first time that has happened in over two and a half years. The thread also notes that four more harnesses now run on OpenRouter, giving users 500+ models behind a single API.

Jack Dorsey Wants an Open Frontier

Source: x.com ↗

Jack Dorsey posted a thread arguing that no company owns what comes next, and that he favors open releases that people can examine, use, and improve together. His point about safety is sharp: a sincere concern about safety can still produce a barrier to entry.

TypeSafe AI Exits Stealth

Source: typesafe.ai ↗

TypeSafe AI has released Jev, its first “System One” model, claiming similar intelligence to existing LLMs but two orders of magnitude faster and more efficient. The company positions this as a new class of frontier models for fast, efficient inference.

Cloudflare Takes on AI Crawlers

Source: blog.cloudflare.com ↗

Cloudflare has announced a network-level solution for the discoverability-vs-training tradeoff, addressing crawler identification and enforcement beyond robots.txt.