9 min read · 14 sources

Google is about to find out whether AI hardware survives space. Its first Suncatcher test satellite, named MVP, launches October 1 on a SpaceX Falcon 9 rideshare - a refrigerator-sized box built by Planet Labs carrying four of Google’s custom TPU accelerators, powered by solar panels that supply about one kilowatt. If the thermal interface material and heat pipes keep those TPUs alive through radiation, vibration and vacuum, it’s a real step toward orbital data centers. If not, it’s a very expensive lesson in thermal engineering.

The bigger news is that RSA just got weaker. Researchers have found a way to forge signatures without factoring the key at all, using classical computing. It’s fully practical for 1024-bit keys - a few months on an academic CPU cluster - and it drops 2048- and 4096-bit keys below the 128-bit security threshold that’s been the floor for two decades. Widely used implementations are still safe, and the practical risk today is limited. But the assumption that breaking RSA requires factoring is now dead, and that changes the threat model for every system still leaning on RSA for long-lived secrets.

A new classical attack forges RSA signatures without factoring, cutting 2048-bit keys below the required 128-bit security threshold.

Google's Suncatcher satellite: TPUs in orbit, one kilowatt at a time

Source: arstechnica.com ↗

Project Suncatcher is Google’s moonshot to put AI data centers in orbit. The MVP satellite launching October 1 is the first real test of that vision. It’s roughly refrigerator-sized, built by Planet Labs, and carries four of Google’s custom TPU AI accelerators - the same ground-based hardware, not space-hardened variants. Solar panels supply about one kilowatt, which is enough to run the TPUs but not much else.

The mission runs a few months and tests how off-the-shelf AI hardware behaves under space conditions: radiation, vibration, heat. The cooling is the hard part - thermal interface material and heat pipes have to move heat away from chips that are used to liquid-cooled server racks. For engineers, this is the first real-world data point on whether you can run AI workloads outside a datacenter. If it works, the economics of AI compute change entirely: no power bills, no land, no cooling towers. If it doesn’t, the failure mode will be thermal, not computational.

RSA is breakable without factoring - and 4096-bit keys are now too weak

Source: arstechnica.com ↗

The new RSA attack forges signatures without factoring the modulus. It uses classical computing - no quantum hardware - and it’s faster than anything previously published. For 1024-bit keys, it’s fully practical: a few months on an academic CPU cluster. For 2048- and 4096-bit keys, the security level drops below the 128-bit threshold that NIST and everyone else treats as the minimum.

The practical risk is limited right now: widely used RSA implementations are safe, and the attack is computationally expensive. But the conceptual breakthrough is the story. The long-held assumption was that RSA’s security rests on the hardness of factoring. That’s no longer true. For engineers running systems with long-lived RSA keys - TLS certs, code signing, SSH - this is the moment to start planning a migration path to post-quantum or elliptic-curve alternatives, because the math just moved under you.

Waymo's Texas surge: 49% growth in three weeks, one-third now Chinese minivans

Source: techcrunch.com ↗

Waymo’s fleet data shows a company concentrating its bets. Since September 2024, the robotaxi service expanded from 3 to 15 U.S. cities, averaging 500,000 paid rides weekly. About 80% of its roughly 4,000 robotaxis are in California and Texas. The Texas fleet is the hotspot: it surged 49% in three weeks to 1,102 vehicles, driven by new “Ojai” minivans - modified Zeekr RTs - that now make up a third of the Texas fleet.

The reliance on Chinese-built vehicles is notable, especially with tariffs driving up costs. For anyone operating a fleet or a supply chain, this is a case study in scaling under constraint: Waymo is growing fast, but it’s leaning on a single vehicle type in a single state to do it. If the Ojai minivans hit a regulatory or supply snag, the Texas growth story stalls.

Blue Origin: $30 billion of Bezos's fortune, and a plan for $30 billion in revenue

Source: wsj.com ↗

Jeff Bezos has poured $30 billion into Blue Origin since founding it in 2000. That’s made it a large space company - around 15,000 employees - and now it’s raised capital from outside investors for the first time. The company expects $1.4 billion in revenue this year, and plans to grow that to more than $30 billion by 2030.

That’s a 20x revenue jump in four years, which is either ambition or fantasy. The outside investment is the tell: Bezos is diversifying his exposure, which suggests the burn rate is real. For engineers, Blue Origin is increasingly a serious player in launch and space infrastructure, not just a Bezos hobby. Watch whether the revenue numbers track the plan.

CPU shortages: spot pricing is dead, reservations take months

Source: blog.pragmaticengineer.com ↗

A new trend of CPU shortages is hitting cloud providers. Spot pricing - previously up to 90% off - has disappeared because there’s no idle CPU capacity anymore. Reservations for specific CPU types now need to be made months in advance, and providers will turn down requests if they don’t have the right hardware.

The driver is AI workloads, particularly reinforcement learning and agents, which are CPU-hungry, not just GPU-hungry. Even companies like turbopuffer are struggling to get CPUs across AWS, GCP, and Azure. For anyone running batch jobs or CI pipelines, this changes capacity planning: you can no longer assume elastic supply. Budget for lead times, or your jobs wait.

Thinking in systems, shipping in loops

Source: tomtunguz.com ↗

The shift from writing code to designing AI loops is showing up in real numbers. Artemis engineers merged 16 PRs per day in August, up from 2 in January. SpaceXAI’s Grok team ships 2,000 PRs monthly. The work isn’t typing code anymore - it’s designing systems where AI writes the code correctly at scale.

The key design principles: resilience (AI checking its own work), self-organization (loops learning from failures), and hierarchy (composable components that get reused once verified). For engineers, this is a new career path: system architecture and steering AI, not hand-coding. If you’re still optimizing for typing speed, you’re optimizing the wrong variable.

The specter of neuralese: AI thinking in unreadable ways

Source: astralcodexten.com ↗

Neuralese recurrence - previously hypothetical in the novel “AI 2027” - appears to be real. OpenAI’s Astra model uses a reasoning technique called recurrent depth that operates outside sequential chain-of-thought, making intermediate results harder to monitor. That’s a problem for safety: if you can’t read what the model is thinking, you can’t catch harmful intentions.

There’s an emerging taboo on creating dangerous neuralese AIs, but it’s unclear whether OpenAI broke it - the line hasn’t been drawn yet. For anyone building safety monitoring on top of model outputs, this is a warning: chain-of-thought is not a reliable audit trail.

Zuckerberg's Tamagotchi pendant: Meta's iPod moment?

Source: bloomberg.com ↗

Meta’s Muse Charm is a Tamagotchi-like pendant for interacting with Meta’s Muse personal agent. It’s expected out in time for Christmas, features a customizable character, and can be worn, pocketed, or attached to a bag. The devices can interact with each other when placed nearby.

The Bloomberg opinion is that this could be Meta’s iPod moment - a consumer device that opens revenue streams that eclipse advertising. That’s a bold claim. For engineers, the interesting part is the device-to-device interaction: a network of cheap, always-on AI pendants is a different compute model than cloud-based agents.

Oracle's force majeure: Project Jupiter may slip

Source: cnbc.com ↗

Oracle sent a force majeure notice for its New Mexico data center project, “Project Jupiter,” to delay payment if it doesn’t come online as expected in 2028. Stock dropped 3%. The project is part of the Stargate AI infrastructure build-out, slated for over 2.4 gigawatts powered by Bloom Energy fuel cells. Oracle has $18 billion in debt tied to it, trading at stressed levels.

Setbacks include local opposition, environmental concerns, and delays in a natural gas pipeline. For anyone watching AI infrastructure, this is a reminder that gigawatt-scale projects are as much about permitting and pipelines as about chips.

NASA vs. China for the Moon's south pole

Source: arstechnica.com ↗

NASA Administrator Jared Isaacman is worried that China’s Chang’e 7 mission, delayed to early next year, could deny access to parts of the Moon’s south pole - particularly Shackleton Crater. The crater’s rims create “cold traps” with potential water ice, making it prime real estate for lunar settlement. Chang’e 7, with rovers and hopper drones, could establish exclusion zones.

Isaacman’s concern is that China is targeting the same location NASA wants, potentially hindering US plans for a long-term lunar presence. For engineers, this is about the infrastructure race: whoever controls the water ice controls the fuel for deep-space missions.

TLA+ can't verify everything

Source: hraness.com ↗

TLA+ has fundamental limits. It can verify invariants and eventualities, but it cannot express possibility and reachability properties - like “I can always shut down the computer” - or hyperproperties, statistical properties, or robustness against code changes. These limits stem from TLA+’s focus on logical formulas over individual behaviors.

Be cautious about claims that AI-assisted TLA+ will enable formal verification of all software. Many real-world properties are inexpressible in the formalism. TLA+ is a powerful tool, but it’s not a universal verifier.

Not everyone can code

Source: funcall.blogspot.com ↗

Effective debugging requires counterfactual reasoning - entertaining false premises to understand how a bug could occur. A significant portion of the population (40% or more) has difficulty with this cognitive skill. Many people are anchored to the concrete world and find debugging abstract software state agonizing or impossible.

This is a real barrier to entry. Not everyone can code or debug, and that’s a cognitive limitation, not a training gap. For hiring managers, it means screening for this skill matters more than ever.

The age of the soft skill

Source: rudyfaile.com ↗

The scarcity of technical capability is over. AI agents can reason about codebases and produce good code quickly. What’s now scarce is judgment - knowing the right questions to ask and understanding the constraints: who it’s for, what it can break. Output is table stakes; soft skills are the differentiator.

For engineers, this means the career path is shifting from writing code to asking the right questions. If you’re not developing judgment and communication, you’re commoditizing yourself.

US intercedes for X over European fine

Source: nytimes.com ↗

The European Commission fined X $140 million in December for failing to prevent deceptive behavior on its site. Now the US is interceding on Musk’s behalf. The diplomatic angle is notable: a US administration pushing back on a European regulatory fine for a private company. For platform operators, this is a signal that cross-border regulatory enforcement is becoming a geopolitical football.