Overview
Today’s feed had two strong currents: builders making systems faster, and big institutions putting serious money behind space and science. Between a Rust swap in vLLM, Bun’s base64url boost, and a revived computer vision paper, the engineering mood was practical and nerdy. At the same time, NASA’s Moon Base plan and a fresh Space Force contract for SpaceX kept the space timeline feeling closer than it should.
The big picture
Performance work is moving up the stack, from model serving frontends to everyday runtime utilities, because the bottlenecks have moved. Meanwhile, space is splitting into two lanes: public plans (NASA’s phased Moon Base) and defence infrastructure (Space Force data networks), both leaning on commercial launch and satellite operators. In the background, health and markets had their own story, with sleep architecture as a treatment target and profits concentrating around AI-heavy firms.
vLLM swaps Python for Rust at the frontend
vLLM merged a Rust frontend as a drop-in alternative to its Python API server, keeping the same engine and the same ZMQ boundary. The pitch is simple: as GPUs speed up, CPU-side work like request handling and pre-processing starts to dominate, so the boring bits need to get quicker too.
Early numbers shared look stark for preprocess-heavy workloads, and it’s opt-in via an environment flag, which should make testing painless for teams already deployed on vLLM.
Bun makes base64url decoding wildly quicker
Bun is promising a 41x speedup for Buffer.from(str, “base64url”) on large inputs in the next release. It’s a reminder that “small” primitives matter when they sit in hot paths, like auth tokens, web payloads, and anything touching binary data.
The note about using simdutf is also telling, this is the sort of optimisation that comes from treating language runtimes as performance products, not just developer ergonomics.
A 2015 head-pose idea gets a second life
Lucas Beyer resurfaced his early vision work, BiternionNets, on predicting continuous head orientation from discrete labels. The core theme is still relevant: pick output parameterisations and losses that match the geometry of the problem, rather than forcing generic regression to behave.
It’s also a nice time capsule of how people worked around limited data and low-resolution labels, and still got smooth outputs by being careful with circular quantities.
AI claims another crack at a famous geometry problem
Andrew Curran says Anthropic’s Mythos has solved the Erdős unit distance problem, with chatter about “overhang in discoveries”. Whether the broader community accepts the result is its own saga, but the pattern is hard to ignore: frontier models are now regularly producing candidate proofs that humans then attempt to formalise.
The interesting bit is not hype about one-shot genius, it’s that these systems keep finding plausible, tidy routes through existing ideas, which changes how quickly the field can explore the search space.
NASA puts a phased Moon Base plan on the table
The biggest attention magnet was NASA’s master plan for a permanent Moon Base at the lunar South Pole, framed as sustained presence rather than “flags and footprints”. The outline is phased: early robotic and commercial delivery, then infrastructure and power, then ongoing operations with routine cargo.
Commercial heavy lift is central to the whole thing, which is both pragmatic and a quiet admission that the logistics problem is the mission.
Space Force hands SpaceX a $2.29bn military data network contract
SpaceX picked up a $2.29 billion Space Force contract to develop a space-based military data network, pushing the relationship further from “we launch your satellites” into “we run the backbone”. This looks like part of the broader move to resilient orbital comms and data relay, designed to keep working under stress.
It also sharpens the Starlink vs Starshield distinction people keep asking about, civilian connectivity on one side, government-specific capability on the other.
Another 24 Starlink satellites, another routine milestone
SpaceX confirmed deployment of 24 Starlink satellites, the sort of post that feels mundane until you remember the cadence. At this scale, “routine” is the achievement, both in launch operations and in keeping the constellation coordinated.
The operational reality is constant traffic management, with collision avoidance and debris risk becoming as central as throughput.
Sleep architecture as medicine, not just more hours in bed
Crémieux highlighted Phase 2 data from Bright Minds Biosciences suggesting BMB-101 boosted REM sleep substantially in people with absence seizures, while total sleep time stayed roughly the same. The headline is the reported drop in seizure counts alongside the REM increase.
If these results hold up, it points to a future where sleep stages are treated as tunable physiology, not a vague wellness goal.
S&P 500 profits keep concentrating around AI-heavy firms
The Kobeissi Letter pointed to a widening gap in profit margins: the index looks strong in aggregate, but much of the expansion sits with the tech and Magnificent 7 cohort, while the rest lags closer to pandemic-era lows. It’s a concentration story dressed up as broad strength.
That doesn’t mean the numbers are wrong, it means the index is telling you where pricing power is, and where it is not.
Amjad Masad recognised by King Abdullah II
Replit CEO Amjad Masad shared that he received a medal for distinction from King Abdullah II on Jordan’s 80th Independence Day. It’s a personal milestone, but also a reminder that tech founders often become national symbols, especially when their story starts outside the usual hubs.
The replies show the full mix you’d expect: pride, congratulations, and the occasional political aside.
Episode #413: 27 May 2026
